Belitsoft > AI Agent Development: Chrome Extension as In-App Guidance Tool For ERP and CRM (2026)

AI Agent Development: Chrome Extension as In-App Guidance Tool For ERP and CRM (2026)

Client

Our client is a non-technical middle-level e-commerce business founder. They don't manufacture anything herself but buys from suppliers and sells white-label products via multiple channels. The business grew fast and scaled globally. Now they have branches in the US, Poland, and the UAE, selling on marketplaces and social media.

Running a retail network requires not only selling great products but also properly administering your tech stack.

Their tech stack includes multiple specialized, complex software systems to manage operations:

  • marketplace management software (each marketplace has its own system),
  • ERP and e-invoicing software (they use Dynamics 365 Business Central Cloud),
  • and Facebook and Google for ads management.

Each system has tons of operational nuances, a unique interface with its own rules and a million little quirks.

Challenge

With constant turnover, the staffing pain in e-commerce is onboarding

You need people who can instantly wrap their heads around ERPs, marketplace dashboards, and ad platforms.

Talents who can jump in and work right away are hard to find, and when you do find them, they are expensive. Most companies end up hiring less experienced people instead. However, training newbies takes forever and costs a fortune before they actually start delivering. It takes way too long to see any ROI.

These business applications are not just complex for novices but also constantly keep changing, breaking old instructions

Without a training system, new hires either guess clicking blindly, or wait to be spoon-fed with instructions, which drags down productivity. Written guides quickly become outdated, and fast-growing companies operating on thin margins have no budget for training new hires.

Many throw staff in the deep end and leave them to sink or swim, expecting them to figure everything out on their own. But that just drives people out the door even faster, and keeps churn high.

This is especially evident with ERPs - complex systems with extensive configurations and large manuals that are rarely read. Newcomers are forced to learn dozens of features they will never even touch, which kills their drive. Motivation dies somewhere around chapter three.

Founders trying to speed up onboarding in the end find themselves sticking in a never-ending loop of training people who do not stay. Every single time someone quits, they must manually repeat the same technical explanations teaching new hires how to navigate the company's tech stack.

Every new face means answering the same questions: "Where do I click?" and "What do I fill in here?". Some pick it up fast, some are totally lost but with such high churn, you're stuck in a loop.

Business owners don't want to be full-time software instructors anymore

They're tired of repeating themselves and exhausted from hand-holding everyone through these programs.

They say: "Give me an AI that overlays the software, shows them exactly where to click, points to the right button so they understand exactly what they must do, and guides them to the result, ensuring every task is completed successfully so I do not have to".

What they really want is to stop wasting their time training employees how to use software. They want to document the workflows and decision-making processes their experts know and turn individual expertise into organizational knowledge that can be preserved even after they leave, in the form of automated instructions for everyone else.

Consulting

There are three broad categories of in-app AI assistants based on the interaction models: doing things for you (classic agent), talking with you (chatbot), and guiding you (adoption platform).

AI chatbots

This is conversational help interfaces within the app (support chat, etc). They don't interact with the interface and are like a steering wheel and pedals that you must operate yourself. Such chatbots are limited because they just produce text answers based on the documentation on how to use software and do not interact with the interface.

Execution agents

AI that executes commands, auto-fill forms, and completes autonomously other tasks and entire workflows within the app. They operate almost like autonomous vehicles, turning, accelerating and braking on their own.

Digital adoption platforms

They provide onboarding assistance overlaid on the app interface. A digital adoption agent acts like a rally co-driver, offering timely instructions on when to turn and how sharp the corner is while leaving the driver in control. Our client asked us to build the guidance agent that belongs to this category. Its job is to overlay the interface, highlight the right elements and guide the user through each step while the user retains control.

Benefits of Digital Adoption Platforms (DAP)

The challenge our client faces is a typical digital adoption case. The global digital adoption platform market is growing fast. The compound annual growth rate was expected to be between 10 and 20 percent.

The goal of a digital adoption tool, or more specifically an in-app guidance tool and AI onboarding assistant, is to guide users step by step through a software interface, highlight the required buttons and fields on each page, and verify that the user followed the correct path to complete the task.

This makes staff more independent and helps companies avoid productivity loss caused by staff turnover.

AI agent development

The standard approach of off-the-shelf digital adoption platforms is to show popup overlays with task lists that users can follow (or ignore/close)

Limitations of Off-the-Shelf Digital Adoption Platforms

A typical ready to use out of the box digital adoption tool could theoretically help our client, but such platforms are not plug-and-play solutions that you can just set and forget. They won't automatically meet the business owner's expectations right after installation without ongoing management.

You have to map out and document the steps for every business process, and then keep those scenarios updated whenever the UI or workflows change. That is something you can live with. But the real issue is that most digital adoption platforms still require a lot of time just to build walkthroughs for all your processes.

Someone on the team must have to master the digital adoption tool itself to build and maintain these walkthroughs. Most such platforms are a pain to learn, and if you are not using them all the time, you forget how they work. When they break, you end up wasting time trying to fix them.

Also, there are so many of these tools on the market that choosing the right one can take forever. None of them are perfect - just look at the complaints on review sites like G2. Some of these can cost tens of thousands a year for a license, which is too expensive for most clients' budgets.

Lastly, standard digital adoption platforms don't actually force users to follow the right path. They just throw pop ups that users can close. But that doesn't fully meet the client's original request.

Comparison with Microsoft's Built-In AI Agents

Built-in AI Agents Developed by Vendors of Your Software

Your existing systems may already have the AI agents you need. Why build custom if it's already there? Microsoft pours serious investment into AI capabilities. Does Microsoft Dynamics 365 Business Central have something near what's required?

Let's look at what the major AI agents in Business Central can do as of early 2026.

Microsoft Copilot for Dynamics 365 Business Central

Microsoft Copilot for Dynamics 365 Business Central is an AI-powered assistant built directly into Business Central. It's designed for small and midsize businesses to automate tasks and provide insights across finance, supply chain, inventory, and sales modules. Copilot uses generative AI based on Azure OpenAI to help employees work faster.

Earlier, Copilot was just a chat assistant. It answered questions by searching official documentation (public docs on learn.microsoft.com or custom knowledge bases), and then created a response based on those sources. These responses looked like traditional chat. It didn't highlight interface elements or enforce a sequence of steps to follow. Copilot as the chat assistant still exists.

Copilot Business Central

Microsoft Copilot integrated into Dynamics 365 Business Central as a chat assistant that answers questions and provides information, but doesn't actively guide users through interface tasks

Later, Microsoft added several embedded AI features built into specific pages or forms where you're working. You click a button on that page and AI can do something right there, like Bank Reconciliation Assistance on the reconciliation page. But most ERP functions this embedded feature doesn't touch.

Finally, Microsoft added autonomous agents: Sales Order, Payables, Expense.

  • The Sales Order Agent processes emails and creates quotes automatically.
  • The Payables Agent extracts invoice data and drafts purchase invoices.
  • The Expense Agent processes receipts and creates expense reports.

These agents don't solve our client's problem at all. They don't need automation that does the work but guidance that teaches their staff where to click. All of these ready-to-use AI agents from Microsoft do the tasks themselves but none of them show a new hire how to navigate the interface.

Agent Playground

Microsoft also offers Agent Playground that lets you build custom agents that act for the user with autonomous execution in the background: open pages, navigate, and complete tasks automatically.

Agent Playground could theoretically automate more custom workflows beyond just orders, invoices, expenses. It could handle fulfillment, drop-shipment, some custom scenarios and reduce the number of tasks new hires need to learn manually.

But it still won't teach humans how to do those tasks. It won't preserve "how to do it" knowledge but only automates "doing it". You need detailed task definitions to set them up but none of them let you highlight interface elements or give you a step-by-step tour of Business Central.

So to actually onboard new staff and preserve institutional knowledge, you still need a custom guidance agent like the Chrome AI Agent extension. And you still need developers to define tasks, not just business users documenting workflows.

Solution

Custom AI Agent Development Approach

Companies facing challenges with ready-to-use solutions often have to make a choice. Either they hire engineers to build an agent from scratch, which may seem quick at first but incurs ongoing maintenance costs, or they purchase licenses for an existing platform, paying a high price to adapt another company's solution.

For businesses seeking to create an AI product in this space, especially startups, the obvious path is to develop a bespoke agent. Our client chose to invest in custom development rather than paying large licensing fees.

OPENAI-AI agent development

OpenAI's official toolkit for building AI agents. The client's solution uses OpenAI's models (ChatGPT API) for their custom guidance agent.

Concept and Stages

Our client really needed an AI software navigation agent. Think of it like Google Maps Navigation, but for software. The AI agent builds a route through the application and walks the user through exactly how to use it.

We broke the custom development down into stages. Since the client is a non-technical founder, we suggested a Time and Materials model.

The idea was to build and pay for a small milestone of the project first, something that would demonstrate our expertise and prove the concept on a real example. This is the so-called MVP, the minimal viable product.

For the MVP, we focused on one of the client's business apps, specifically Dynamics 365 Business Central Cloud.

AI agent development Chrome Extension

The AI agent overlays the interface, highlights the specific UI element the user needs to interact with (the Type dropdown), and provides step-by-step guidance in the sidebar while keeping the user in control of performing the actions.

AI Agent Google Chrome Extension Development

All of the client's business software is web based, which simplified the development effort, since the AI agent is easier to set up with browser access. Chrome is the most widely used engine, so we decided to build a Chrome extension as the interface for the AI agent.

AI agent development Chrome Extension

This image shows AI agents available as Chrome extensions - the exact same approach we used to build the client's guidance tool.

A Chrome extension is the perfect bridge for the agent since it can access the DOM and also grab screenshots to send over to the LLM for analysis.

Building the AI agent requires an LLM. Since fine tuning a custom model wasn't within the MVP budget of our client, we proposed using an existing commercial model. We chose OpenAI ChatGPT as the most well known option, although the choice was not critical. Most commercial models today offer very similar capabilities and comparable API pricing.

We built the extension to run as a panel on the right side of Business Central. The Google Chrome extension adds this sidebar without covering the ERP interface and blocking the main ERP screen.

This way, the user can see the full route for the task through that window. The agent's main job is to monitor the screen and keep the user following the steps defined as correct. We placed the relevant instructions in the customized system prompt for the LLM so the agent understands what is considered correct.

As the user moves through the task, the agent highlights the relevant parts of the interface (buttons and fields) in real time directly in the page's HTML to show where to click and what and where to fill in.

Process

Custom AI Agent Development Process

Just using the default OpenAI model wasn't enough. The model had no knowledge of the 2026 version of Business Central since it was released after the training cutoff, and lacked understanding of how the interface actually looks and works. We addressed this through prompt engineering customizing the commercial LLM for the client's specific requirements by providing detailed context in the system prompt.

Below is a breakdown of the customization process:

  1. We took the role of an expert user of Business Central and asked the model to build the correct workflow (the right path with the steps) for a particular business task.
  2. Every time the UI of Business Central changed (or not) after we did what the LLM suggested, we'd grab a screenshot and send it back to the model. This gave the model a way to see the results of its own advice from the point of view of a human user and track what was happening. We also sent the HTML along with it, so the model could see what was going on at the code level. That way, the model learned to double-check itself so its knowledge stayed relevant to how users of Business Central actually work within their specific workflows.
  3. By showing the LLM model the exact path, we trained it to verify user actions on the fly. In production, the LLM-powered agent will guide users without us based on this knowledge.
  4. We also trained the agent to inject JS into the current Business Central page to highlight what the user needs to do next - like clicking a button or filling out a field.
  5. We programmed the AI agent so that it won't mark the step as finished in the right sidebar - the step stays open in the task list - until the user actually finishes the action. That way, the user can track their progress and see what's already done and how many steps are still left to do.

The Ins and Outs of Custom AI Agent Development

You need more custom work hours to build an AI agent for Business Central than for other platforms. Business Central doesn't use modern frontend standards - it's packed with iframes, making it hard for AI agents to read the page correctly. You also need to design the agent to minimize requests to the OpenAI LLM.

Every request costs money, so the goal is to only hit the API when you absolutely have to, to cut down on unnecessary pings.

There is still no way to guarantee a reliable response 100 % of the time with any model on the market right now. Sometimes the LLM hangs or stops responding - LLM providers usually blame it on network problems. You must fix this under the hood by minimizing automatic background requests to the LLM, so the user never even notices a thing.

When the agent sees a request for the workflow from the user of the software for which the in-app guidance assistant is being built, it checks if the workflow for completing this task has already been built and saved in the system - either because a previous user has successfully used it, or because an expert pre-built it.

This way, the agent skips the LLM and grabs it straight from the database with storage of successful routes. If the interface updates and the old route breaks, the AI agent just goes back to the LLM to refresh it.

Finally, since each AI agent is built for specific guidance tasks, you cannot ask the guidance agent anything that is not part of the workflows it was designed for. It will not help you with financial questions - you would need a separate agent for something like that.

Challenges of Custom AI Agent Development

The challenge is not to build an AI assistant that helps users navigate complex web applications (enterprise software with tons of menus and options) but to make it work reliably and safely in a day-to-day business context.

Complex Workflows

Connecting the OpenAI or Claude API and writing a few prompts to help employees complete standard workflows feels almost magical for clients because they see immediate results. But soon they discover that the more complex the tasks they put behind the agent, the greater the number of surprises they face. Belitsoft's team fixes these issues with additional prompts, or fine-tuning a model but this shows that building an AI assistant may be an iterative, not one-time, investment.

HTML changes

It also quickly became obvious for clients that HTML changes to the software pages can break AI guidance agents. If vendors of this software make changes, you may need ongoing prompt updates and retraining as part of maintenance for your in app guidance agent. Belitsoft's AI team addresses this by designing a self-healing architecture that detects UI changes and adapts automatically using visual and structural analysis to identify elements by their function rather than by their position.

AI Context Remembering

If your employee is halfway through filling out a form and the page refreshes for some reason (timeouts, etc.), the assistant should not forget what they were doing, so they don't have to explain the whole task again to the agent. That was not an issue for Belitsoft's AI team because our AI engineers solved it upfront.

Employee Access Control

If your employee doesn't have permission to view certain data, the AI assistant must inform them they lack access instead of trying to bypass security controls or work around restrictions. Belitsoft's AI team built the AI agent to respect existing security policies from the ground up. We do things architecturally sound from the very beginning, so there were no problems here for us and our client.

Automated AI Execution Risk

The AI assistant must tell your employee what to click and which forms to fill out, but your employee should do the clicking and data entry. This type of AI agent should not execute actions on its own, so it cannot accidentally delete records, submit wrong forms, or trigger unintended workflows in your business systems. This is a known problem with execution agents. They can automatically click buttons and enter data without human oversight. Belitsoft's AI team chose the DAP approach from the start to avoid these risks entirely. DAP agents do not have this problem because the employee stays in control for every action.

Results

We built an AI-assisted Chrome extension MVP that solves the client's problem. The AI Agent suggests steps, highlights the UI, and walks the user through the whole process until their task is done.

Many companies invest in training their staff on the business software they use and cover the cost of the classes. But once trained personnel quit, you have to start the whole training process from scratch with someone new. This is where the AI assistant comes in. It supports newcomers without needing an experienced employee standing next to them constantly. Businesses usually need a way to spread the expertise of their best people across the whole team. By teaching their workflows to the agent, you are no longer hit so hard by staff turnover.

You can compare this kind of AI agent to Google Maps Navigationfor your software. It tells you where to turn and what to do next until the task is done. The only difference is that you are navigating a software interface instead of city streets.

To scale the MVP into a full-on product with the right investment, you could go the enterprise route, including fine-tuning specialized models that focus specifically on user experience.

AI guidance is quickly becoming the go-to approach. It's going from a specialized know-how to the norm for how apps are built these days.

Belitsoft is an experienced partner that has the expertise to work however you need. They can supply dedicated AI specialists via staff augmentation, or they can build the whole thing for you as a turnkey project, managing every step from the initial concept to full implementation.

1
2
3
4
5

Recommended posts

Belitsoft Blog for Entrepreneurs
Mastercard Agentic Payments: Belitsoft’s Report and Guide [October 2026]
Mastercard Agentic Payments: Belitsoft’s Guide [October 2026]
How an AI agent finds your products An AI agent can find out about your products in 4 ways. AI agents can find your products on your website  An agent can use your website like a customer: open pages on your platform, fill in your eCommerce store’s forms with card details (an agent certified in Mastercard Agent Pay, Visa Intelligent Commerce or American Express ACE uses a token instead of the real card details), and complete checkout to finalize the purchase. Registered and unregistered agents Some of the agents that visit your website are registered with payment networks such as Mastercard, which verify and certify these agents after a security review. For you as a seller, these agents are the best buyers because their purchases seem to cause the fewest problems. You just need to check whether the agent trying to visit your website and make a purchase is listed in Mastercard's trusted agent directory within the Mastercard Agent Pay program (of course, while maintaining your usual payment security and fraud checks). The other agents may be from AI platforms, such as ChatGPT in agent mode. One more type is consumer-owned agents built using open-source tools. Such agents, regardless of their registration in Mastercard's trusted agent registry, may still be legitimate shoppers, and Mastercard does not recommend blindly blocking them all. The strategy to block bots but allow shopping agents When you block bots on your website, you may forbid shopping agents from visiting your website as well without even knowing it, losing customers to competitors. Such an agent will never recommend your products. Naturally, as a seller, you don't want this, so Mastercard suggests a trust strategy for managing agents on your website. With this strategy, you keep your product pages, search and categories open to all agents, unless they are malicious bots. Then, you distinguish between certified agents and uncertified but still legitimate agents. For uncertified agents, you turn on additional protections, such as additional checks at checkout, like a sign-in or purchase limits.  Five levels of trust For this strategy, Mastercard splits the agents on your website into five levels of trust and calls this the trust spectrum. Level one is certified agents, and you give them full access and a simple checkout with fewer steps. However, your fraud protections stay active for these agents too. Level two is agents that AI platforms like ChatGPT send to your website, and Mastercard calls them platform-identified. You can identify them by the User-Agent header or by Web Bot Auth, and most websites use this header to spot automated traffic. This header is easy to fake, so a signature tells you more, says Mastercard: with Web Bot Auth, an agent signs each request to your website. Platform-identified agents have standard access, and you keep the normal checkout with payment authentication. Level three is consumer-owned agents that tell you who they are, for example with Web Bot Auth. For these agents, Mastercard recommends open discovery: they can look at your products, and you keep the checks that you have at checkout and monitor their behavior. Level four is agents that don’t tell you who they are, and Mastercard calls them anonymous. Such an agent can be a legitimate agent of a consumer or a malicious bot: an agent that a person uses to shop with you, or a bot that scrapes your website. Such an agent sends you no identity signals, so you can’t tell the two apart. These agents can also look at your products, but you add more checks at checkout, use rate limiting and analyze their behavior. Such a limit can be 100 requests per minute for agents, as in a Mastercard example. According to Mastercard, rate limiting stops abuse and lets legitimate agents shop. Then level five is bots whose bad behavior you have identified: Mastercard calls them malicious, and they don’t follow your rate limits or your terms of service. These are the ones you block and report. However, Mastercard doesn’t say who you report them to. Set up Web Bot Auth and track agents in your analytics Mastercard also recommends that you set up Web Bot Auth on your website. Web Bot Auth signatures show you which agents say who they are and which ones stay anonymous. Also, Mastercard says you should track agents as a separate channel in your analytics, so you see where in your checkout agents fail. Make your pages readable for agents Once agents can see your pages, put the product name and the price in the page as soon as the page loads, not added later by JavaScript. That’s what agents expect, and according to Mastercard they may not wait for a complex script to load. Also, the availability and the fields at checkout need to be in the page when the page loads. Then test two or three important paths on your website, like buying your most popular product or making a return. That way you can see where an agent could have problems. AI agents can find your products via ChatGPT For ChatGPT, you upload your product data to OpenAI, and ChatGPT will show your products to your customer. How visible your products are in ChatGPT depends on the quality of this data. You upload this data with the Agentic Commerce Protocol (ACP), an open standard that OpenAI and Stripe developed together. According to Mastercard, what you pay for in ChatGPT is traffic and discovery, and you don’t pay for each sale. Mastercard doesn’t say how much you will have to pay. If you want to take part in ChatGPT shopping, you can talk to Mastercard, and the company will tell you more. Your customer doesn’t pay inside ChatGPT anymore: in March 2026 OpenAI shut down Instant Checkout, the option to buy in the chat. However, OpenAI’s guide for developers, which has no date, describes what you implement as a merchant to support Instant Checkout in ChatGPT. Now your customer is sent to your website or your app to complete the checkout and the payment. As for ChatGPT, our report on Stripe covers this way to find your products. AI agents can find your products via Google According to Mastercard, you have to set up your products in Google Merchant Center with the native_commerce attribute. After that, people can find your products in Gemini Shopping and in Google Search AI Mode. In Gemini Shopping, how well your products are shown depends on your product data, not on paid advertising. Also, people can pay for your products in both places, but for that you have to implement Google’s Universal Commerce Protocol (UCP) on your website. UCP is an open-source standard for agentic commerce from Google. According to Mastercard, you can use this payment option if you are an eligible merchant in the US. Mastercard’s guide doesn’t tell you who is eligible. Also, Google’s announcement for retailers only says that shoppers can check out from eligible US retailers. However, Google said on May 20, 2026 that this checkout will roll out in Canada and Australia in the next few months, and later in the UK. As of October 14, 2026, Google has posted nothing newer about Canada, Australia and the UK. AI agents can find your products via Mastercard Agent Connect, a single integration that Mastercard promises Mastercard announced Agent Connect on September 9, 2026. Mastercard promises that you will only need a single integration with AI agents and payment service providers (PSPs), and the agents that participate in Agent Connect will be able to find your products. In Agent Connect, you choose which AI experiences you connect to. If you want to join, you can speak with the Mastercard team, says the Mastercard page for merchants that includes Agent Connect. Mastercard doesn’t say when Agent Connect will be available or whether you will have to pay for this service. According to Mastercard, technology companies and platforms such as Getnet, Global Payments, Nexi, Samsung and Trip.com expect to work with Agent Connect, beginning in the United States. Mastercard also promises onboarding options without code for eligible merchants, but doesn’t tell you which merchants are eligible. How you accept the payment How Agent Pay works at your checkout An agent certified by Mastercard pays you with Mastercard Agent Pay. Agent Pay is Mastercard’s agentic payments program, which lets certified agents make secure purchases for your customers. According to Mastercard, the agents in Agent Connect will also pay you with Agent Pay. In Agent Pay, the agent pays you with a token instead of a card number. The agent platform asks the payment network for this token, and the agent enters the token in the usual fields on your checkout form. After that, your payment gateway receives the token. Your gateway then exchanges this token with the payment network for the card data. So you don’t have to change anything in your payment processing. Can your customer buy with an agent today? So can your customer buy from you with an agent today? Yes, if your customer is “present through the agent,” says Mastercard’s Brice van de Walle. Your customer tells the agent what to buy at the time of the purchase. A purchase where “the consumer is no longer present” is what van de Walle calls “the future.” However, with a customer who is present, a “single merchant, single product” purchase is possible now: “Right now, we can already cater for that,” claims van de Walle. He also says that a purchase with several merchants and a customer who is present is “something which we believe can be managed already today.” Check the Web Bot Auth signature With Agent Pay, only registered agents can transact on the network, says Mastercard. So you need to know whether the agent at your checkout is certified. You know this from the Web Bot Auth signature that the agent sends with each request. The tag in this signature shows you which payment network certified the agent. When you check the signature, you check the agent against a trusted agent registry: a payment network like Mastercard adds each certified agent to this registry. You can check this signature in your backend or with your edge provider, a company like Cloudflare, Akamai or Fastly. Mastercard recommends the edge provider: ask this company how to set this up. Read the tag in the signature If the signature is valid, you look at the tag. The tag is Agent-pay-auth if Mastercard certified the agent. If Visa certified the agent, you see the tag agent-browser-auth for browsing and agent-payer-auth at checkout. Other agents that support Web Bot Auth have a generic tag, web-bot-auth: ChatGPT in agent mode is one of these agents. According to Mastercard, you should check for the tags of the payment networks and for the generic tag. Agents without a signature However, Mastercard says that not all legitimate agents have this signature yet. Without the signature, you can identify some legitimate agents by their User-Agent header or by their behavior before they adopt Web Bot Auth. Mastercard calls these signs known agent patterns: you give them less trust than a verified signature. However, a User-Agent header alone doesn’t prove who the agent is: Mastercard warns that this header is easily spoofed. Mastercard tells you to relax your checks step by step as these agents show more signs of trust and to watch how they behave. If a request has no valid signature and shows no known agent pattern, you apply the rules that you have for bots. In this case you have three choices: block, challenge, or allow and keep a log. However, Mastercard also tells you not to turn agents away just because they lack scheme registration. Make your checkout form readable for agents An agent that shops on your website like a human has to fill in your checkout form. If the form is hard to read, the agent can fill in the wrong fields or give up on your checkout. According to Mastercard, an agent reads the names of the fields, the labels, the input types and the autocomplete attributes. So you check that every field has a name that tells agents which data to enter, the correct input type and an autocomplete attribute. Also, you need to link the label to the field in the code. These fields should use the native input, select and textarea elements instead of your custom components. That’s because agents can’t work with non-standard elements. These points are in a Mastercard checklist for forms: use this checklist to verify that your forms are ready for agents, says Mastercard. Check the CAPTCHAs on your website Mastercard says that CAPTCHAs are bot checks made to stop scrapers, so you should avoid aggressive CAPTCHAs that stop legitimate agents and use rate limiting instead of blocking all the bots. 3-D Secure and payments by agents At checkout, an agent can also be stopped by 3-D Secure. 3-D Secure is an extra step in an online card payment: the bank asks your customer to approve the payment. This step starts on your website, and an agent usually can’t approve the payment for your customer, so the step fails. However, a certified agent verifies your customer before the purchase: Mastercard says that this check is the job of agent providers. So your systems don’t need additional authentication logic to check your customer. However, Mastercard says that you have to adapt your 3-D Secure to the limited support for payments by agents today. So you check the Web Bot Auth signature to see if a certified agent sent the request, says Mastercard. If you see that a certified agent sent the request, you can skip 3-D Secure. You can also leave 3-D Secure on. If the step fails, you don’t want to count the failed step as fraud or block the payment of a certified agent just because 3-D Secure failed. According to Mastercard, the payment networks are working on 3-D Secure support for payments by agents, so you should watch for the announcements of Mastercard, Visa and American Express. Checks for agents that are not certified For agents that are not certified, you use the security checks of any online card payment: 3-D Secure and checks on the address and the card security code (CVV). Also, match extra checks to the risk of each purchase, says Mastercard: rate limiting, a required sign-in to an account, a confirmation by email or by text message (SMS) and purchase limits. Mastercard’s guide for merchants on agentic commerce has a table of 3-D Secure cases for payments by agents, and this table doesn’t mention certification. If you start 3-D Secure and the payment then proceeds without authentication, the table tells you not to count this as a sign of fraud. Set up merchant validation Mastercard also describes merchant validation: an agent can check that you are the seller and not a copy of your website. For now, this check is in development, and you can sell to agents without this validation. However, Mastercard recommends that you set up merchant validation. Mastercard’s guide describes the benefit in two ways: in one table, certified agents are expected to recommend validated sellers first, and in another table certified agents buy only from validated sellers. Ask your payment service provider (PSP) or acquirer to register you Merchant validation of your business starts with a registration, and according to the page The Rails for Agentic Commerce you should be registered to take part in Mastercard Agent Pay as a merchant. So who registers you? According to Mastercard, your acquirer or your payment service provider (PSP) usually does that for you. So you should contact your PSP or your acquirer: ask if they are ready for agents and ask them to register your business. Your PSP will need your merchant identifiers, your website addresses and your business details. Also, ask who manages the signing keys, the keys that prove your identity to agents: your PSP or you. Who takes the loss when an agent makes a mistake The question Mastercard is working on, and an answer from IXOPAY After the agent pays you, who takes the loss if the agent bought the wrong product? That’s a question Mastercard is working on. In January 2026, Mastercard said that the company was answering this question by building standards for agentic commerce with partners, including the AI platforms. Fortune asked Ed McLaughlin, Mastercard’s President and Chief Technology Officer, in an interview published in September 2026: “The liability falls on whom?” McLaughlin did not name anyone. He had just made a promise to your customers and to you: “you have zero liability if you have a valid transaction on our network, you have full recourse, and likewise to a merchant that we will stand behind that payment commitment.” He did not tell you whether a purchase with a mistake by the agent counts as a “valid transaction.” However, Jill Willard, the chief technology officer of IXOPAY (a payment platform that connects online merchants with payment providers), claimed on September 21, 2026 that today, in practice, Mastercard, Visa and other networks put the loss on the merchant and on the merchant acquirer (the bank that takes card payments) A protection program from American Express However, American Express has announced a program for these mistakes, Amex Agent Purchase Protection. American Express tells you that this program is an intention for the future. The company intends to protect eligible Card Members and merchants from charges after a mistake by an agent that is registered with American Express. So far the terms that American Express has published tell your customer to start a return with you first if a return is possible. These terms cover credit cards from the United States. However, Mastercard’s guide describes this program as cover for cardmembers against losses from mistakes by registered AI agents. The dispute rules that apply today So what rules apply to you today? According to Mastercard the standard chargeback rights and dispute processes apply when an agent buys from you: your customer keeps the right to dispute the payment, and you can’t ask your customer to give up this right for what an agent bought. However, Mastercard says that you should expect new rules: the payment networks are developing dispute rules for purchases by agents. Evidence for each kind of dispute When you receive a dispute, you first find out if an agent made the purchase. For agent purchases, Mastercard says you should prepare your dispute processes for new kinds of disputes and names the evidence for each kind. If the agent misunderstood your customer and bought the wrong product, your evidence is what the agent searched, viewed and selected. If the agent made a mistake such as the wrong quantity, you show the cart at each step. If your customer says that a malicious agent made the purchase, you prove that you checked the Web Bot Auth signature and the certification of the agent. If your customer doesn’t remember the purchase, you need the logs of the agent’s session and the time when your customer confirmed the purchase. Each time, you have to prove that the payment was authorized and that the agent acted within the limits that your customer set. That’s why Mastercard tells you to store the context of each purchase by an agent at the time of the sale, and calls this context your primary evidence. This context holds the evidence for each kind of dispute, and also which agent bought from you and who your customer was. What your support team needs to see In a dispute, Mastercard wants your support team to see five points. These are the order, the agent session and the identity of your customer. The other two are the intent exchange and the signature data as proof of the authorization by the agent and by your customer. The intent exchange is what your customer asked for and confirmed, and you can find this point in the context that you store today. Also, Mastercard has a code example for developers for this storage. Intent data from the agent and the networks Also, the intent, what your customer allowed the agent to do, travels with the payment, according to a speaker in the interview with IXOPAY. However, today the networks “don't actually let you see the intent object when they pass it through,” says a speaker in this interview. Mastercard says that the payment networks are building systems to share this intent data with you and that you will receive this data as part of the transaction later. Intent data that you receive straight from the agent is also part of the future, says Mastercard. Verifiable Intent Verifiable Intent, a protocol from Mastercard and Google, adds a tamper-resistant record to the intent data of your customer. This record links your customer’s verified identity, the agent’s instructions and the result of the purchase with cryptographic signatures. According to Mastercard’s guide, this protocol is in use or in a late stage of standardization at version 0.1. In September 2026, Mastercard said that Agent Pay uses Verifiable Intent to help confirm that purchases have the authorization of your customer. When an agent sends you a checkout with Verifiable Intent signatures, you can store this intent as evidence for disputes, says the guide. According to the guide, such a record is much stronger evidence than ordinary transaction logs in disputes about a wrong product, and the guide tells you to consider adding Verifiable Intent records to the context that you store when more companies adopt the protocol.
Stripe Agentic Payments: Belitsoft's Report [October 2026]
Stripe Agentic Payments: Belitsoft's Report [October 2026]
Your E-Commerce Platform Blocks AI Shopping Agents Agents are shopping, and Stripe is on the back end powering the checkout behind agentic payments. Agent commerce "is real. It's happening today," claims Stripe's Jillian Chi. "It's all early days right now," cautiously claims AWS, yet it is already hurrying to integrate with Stripe. There is an unexpected problem. Even if the purchase has been authorized through Stripe, the agent often still can't buy anything, because it got blocked by your retailer's server as bot traffic. Merchants should allow the good agents, the ones with a paying shopper in them, to come through the front door to complete the purchase. "That's the challenge that we're facing right now," says Stripe's Danny Smith in the September 24 podcast. Of the retailers who block, half do it on purpose, the other half don't realize they're doing it. "Walmart and Amazon both block third-party agents because they look like a threat. They look like fraud," says Simon Taylor at Tempo on September 28. So, how's this thing supposed to pay if it can never get through to the checkout in the first place? Smith tells how he sent the Muse and Instinct agents to buy a leather jacket: both got blocked by every single retailer. The agents found the jacket, the host had authorized the purchase, he was a high intent shopper who knew exactly what he wanted. "But I got rejected at the front door. And these were some major retailers."Smith Since September 15, Cloudflare blocks AI agents by default on pages with ads, by Cloudflare's announcement. On September 21 Amazon blocked Meta's Muse from Amazon.com, saying the agent does not identify as an agent while storing customer credentials.(Meta says that Muse has no visibility into passwords or payment methods). Stripe, says Smith, is "working on some potential protocols that can help identify good bot versus bad bot." However, what can you do for now? You must know whether your door is shut. According to the Link documentation, the tags an agent uses to classify what happened at a store are captcha, anti_bot_script, cdn_block, waf_block, login_required. So check the settings on your side, in your CDN, WAF or checkout, for the agents you want to let in. The shopper's agent pays with Stripe's Link or a one-time card 72% of consumers say they won't share their card details with AI, according to a September 11 report by Fortune. The agent must never see the shopper's card details - that's the main rule of agentic payments. Stripe addresses this with Link, its own consumer wallet. Stripe has published instructions at link.com/agents that an agent can read to learn how to use link-cli, a tool that lets agents communicate with Link. Shoppers who want their agent, for example in Claude, to pay using their Link wallet need to tell it to "learn this skill." If your store accepts Stripe's Link The shopper's agent will first try to use the payment details already saved in Link, without entering card details. Link generates a token, says Smith, and the shopper approves the payment in the Link app. If your store does not accept Stripe's Link and requires card details to be entered in a form What does the payment flow look like for a shopper if your site only has a standard card payment form? Stripe has a solution for this too. Link issues a single-use virtual card to link-cli, and the agent can safely fill out your form using that card's details. Done. Stripe also works with some well-known agents, such as Meta's Muse, Grok Bot and Instinct, which have integrated Link's wallet for agents. For these agents, the link-cli integration is already in place. If you want to be found inside ChatGPT and Gemini, hand Stripe your catalog Today, agents can't discover products from many merchants. Your pages need to be indexed because agents search the web. Otherwise, you don't exist. But indexing isn't enough. An agent is a software that needs a structured feed to search your catalog. A standard commerce interface can be difficult for agents to interpret, so they leave the site. Most merchants are focused on how to show up better in search. If agents can't discover your products, Stripe promises to bring them "directly to the agents" in ChatGPT and Gemini. With Stripe's Agentic Commerce Suite, merchants submit their product catalog data to Stripe, which then passes it to the platforms behind these agents. Stripe passes the catalog on as provided, so merchants need to format the data according to the agents' requirements. According to Jillian Chi, you push "your product, your inventory, your pricing, your product information onto the ACS platform on Stripe," and Stripe pushes it to ChatGPT, Copilot, and Gemini. According to Stripe's documentation (the seller guide and the guide to managing the integration), the steps are as follows: Open your Stripe account and go to the Agentic commerce page in Dashboard Add your terms of service, privacy policy, return policy, and a tax code for each taxable product Upload your product catalog data as a CSV file, with one row per item Make sure the feed is current: titles, descriptions and images are updated once a day, stock and prices every 15 minutes, or agents will report that a product isn't available Select the AI app you want to sell through, read its terms, and request a connection. The agent has to accept your request. And that's it, as long as you're in the US, Canada, or one of 32 European countries. "If OpenAI allows you as a merchant to onboard, sure, but if they don't, then you're stuck there." Jillian Chi "You can refine product catalog data. You do a lot of refinement, AEO, GEO, things that we don't do."Danny Smith Interestingly, OpenAI at first did not want the user to leave the chat to complete the purchase, and had instant checkout. Stripe was powering those experiences, and Etsy was their first merchant. Then OpenAI shut that down and pivoted to discovery. Today the checkout in ChatGPT is a punchout that goes back to your website. Etsy now has an app inside ChatGPT that sends the buyer to Etsy. "Traffic from agentic experiences is still less than 1% of our overall traffic, [but] they're higher intent, higher average order value traffic." On its August earnings call, Etsy's CEO Kruti Patel Goyal Walmart also tried integrating directly with OpenAI and found that conversion was worse than on their website. That's why instant checkout in the their-party website's chat did not work out for them. Even so, the idea of instant checkout inside the their-party chat has not disappeared. For example, Gemini together with Stripe has a couple of merchants in a pilot, and an official announcement from Google on instant checkout in the chat is expected late October. Conversational Commerce: Chat App with in-chat Stripe checkout on your website Merchants build an agent to improve the customer experience on their own websites. In this case, not just the conversation but the purchase itself happens inside the chat. Today, it's called conversational commerce. For a shopping flow facilitated by AI chat on the ecommerce website, the average conversion rate is "four times higher than that without an AI agent," says Stripe's Jillian Chi. At AuthX on September 28, she showed what a shopping flow looks like when you embed Stripe's Payment Element in the chat on your ecommerce website or app. A customer who's low on a product goes to the agent and tells it what they need. The agent talks to the LLM, gets all the product information that is required, and gives it back to the customer as a recommendation. The customer then chooses a product and clicks on check out. Through the MCP server, it then brings up a Stripe payment element in the chat window, so the buyer does not need to leave the agent interface. The client can just use their card information, give their OTP and one complete the check out quickly. By Stripe's documentation, a PaymentIntent is created on your server for that form, and Link is on by default in the Payment Element You can build something like this on your website today. If you sell an API, Stripe helps you charge agents per call An API-first software business needs to say in a machine readable way through the machine payments protocol: this is what I sell, how much it costs, and how you pay. Stripe's machine payments protocol (MPP) was co-authored by Stripe and a company called Tempo. The way that it works is pretty simple: the seller's service tells the agent the price, the agent authorizes the payment and the service hands over whatever was being paid for. AWS, which works with Stripe, showed at the Stripe Developer Meetup at Stripe's office in South San Francisco on September 22 how an agent pays (AWS offers a platform for building agents, AgentCore). The speakers were Chethan Shriyan, Principal Product Manager, Amazon Bedrock AgentCore, and Hasan Tariq, Principal Solutions Architect, AWS. Carrying out the task "go and get the contents from this URL," the agent may face the answer "HTTP 402 code," meaning "payment required," for example, at a cost of $0.10. To pay, the agent must have access to a wallet from which the agent can spend those funds, to pay for the content and get the contents. The agent also needs the user's consent: the user gives consent to the Privy app once, to use the wallet for payments, and the agent spends within the payment limit the developer has set up. Stripe on the buyer's side provides this technology together with AWS: the AWS product is AgentCore Payments, the Stripe wallet is Privy, in stablecoin. The buyer needs to give funds to the agent, into that Privy wallet, through either traditional payment methods like credit card or through stablecoin. For the agent to be able to buy the content from you with that 402 answer, you need, by Stripe's documentation, a profile in your Stripe Dashboard and the mppx library on your server. Your end point returns a 402 with the price, then verifies the payment, and the money settles into your Stripe balance. According to Stripe's documentation, the minimum charge is 0.50 USD by card and 0.01 USDC by stablecoin. What are agents really paying for in these early days? What can you, as an API seller, sell them today, and what is emerging? Chethan Shriyan at AWS sees two use cases from agent developers. So-called pay-per-use. APIs or MCPs already sell premium data and charge per API call. So-called pay-per-inference. This is when, instead of one model, you set up your agent to switch between LLM models dynamically while doing a task. Here the agent pays the model provider per call instead of a subscription. Stripe's MPP is still small. By September, six months after the March launch, transactions "in the last 30 days" had exceeded 20 million: "not big numbers," says Chi. 37 cents a payment on average. How to accept agentic payments with Stripe If you are on Stripe's hosted checkout, you are already connected Your checkout process must be structured clearly enough for an AI-driven bot agent to go shopping and successfully complete the checkout. The Pay button stays locked until the payment form is complete. Previously, to pay, the agent had to use browser automation and web scraping to figure out how to pay through a checkout form built for humans. Stripe Checkout for agents is a way of making a human-readable checkout machine-readable, powered by WebMCP. According to Stripe's Danny Smith in the September 24 podcast, if you're using one of its hosted checkout services, you don't have to do anything. As long as you're using Stripe hosted invoices or Stripe hosted checkout, "the agent readiness is already built in," and "you don't have to do anything if you're a Stripe merchant," says Danny Smith. Thus, the checkout flow for agentic payments will be automatically enabled. Smith quotes "some initial statistics": "42% less tokens burned by an agent when it's trying to check out using Stripe Checkout for agents" and "38% less tool calls," compared with attempts to pay through a form built for humans. If you built your own checkout on Stripe's API If your developer built a custom, bespoke payment page that calls the Stripe API, you're not ready to sell agents. Danny Smith put it this way: "we will have to work with you to migrate you over", whatever that means. Stripe writes in its blog post that it "decided to incorporate agent-friendly WebMCP across Stripe Checkout UIs," so that merchants get "access to future agentic shoppers without them needing to change their integration." For a custom payment form, Stripe provides no steps as of 4 October. Conclusion Agentic commerce hasn't yet had its breakout moment. Nevertheless, retail advisors say companies should be getting ahead of it, even though not all individuals will fully turn the purchasing process over to AI.
EU AI Act Enforcement: What Changes on 2 August 2026
EU AI Act Enforcement: What Changes on 2 August 2026
What Happens After a Complaint Is Lodged Against Your AI System The complaint is submitted to the national market surveillance authority. If this authority has sufficient reason to consider that your AI system presents a risk, it starts a formal evaluation (Article 79). And during this evaluation, the burden of proof lies on you. EU AI ACT Article 5 For the prohibited practices of Article 5, you must show the authority how your system actually works, what data it uses and where this data comes from, and that it behaves in the way you declare. The form of this proof is a written assessment for each AI system (checked against all 8 prohibitions from Article 5, dated and signed), a description of how its engagement and dialogue mechanics work, an inventory of the data it uses, and test and monitoring records. The most important part of this proof is the records. According to the guidelines of the Commission, when your system causes harm that you never intended, the same incident can be qualified as a prohibited practice or not, depending on whether you took safeguards and can demonstrate them. This means that if your chatbot caused harm to a user, but you can show the guardrails, the tests, and the monitoring you used to prevent it, the prohibition does not apply. If you cannot show them, it does. EU AI ACT Article 50 For the content-marking and labelling obligations of Article 50, the authority first checks whether you follow the Code of Practice on Transparency of AI-Generated Content. If you have not signed this Code, you must show that your own marking and disclosure measures achieve the same result, and the Commission expects a gap analysis comparing your measures with the measures of the Code. In this scenario, the authority will check you more closely: it can request additional information and additional access to test whether your marking actually works. The chatbot disclosure and the emotion recognition notice rules from Article 50 are not covered by this Code: in this case, the authority checks your implementation directly. Large fines for incomplete information You must be very careful with the information you provide during the investigation: not only misleading or incorrect, but even incomplete information is a separate violation with a fine of up to €7.5 million or 1% of turnover. If the authority finds your AI system non-compliant, you receive a maximum of 15 working days to correct the violation, withdraw the system, or recall it. In practice, this means that a transparency violation must be corrected, and a prohibited practice must be stopped. If you do not do this, the authority blocks the system on its national market and notifies all other EU member states. And the fine for the violation itself is added to this: up to €35 million or 7% of turnover for a prohibited practice, and up to €15 million or 3% for a transparency violation. Whether you will actually be fined, and the amount of this fine, depends on the circumstances: quick correction of the violation and cooperation with the authority are taken into account in your favour (Article 99). EU AI Act Compliance Checklist: What to Do Before 2 August 2026 To state that the EU AI Act does not apply to your business, you must answer several questions for each of your AI features: whether this feature is an AI system within the meaning of the official definition, whether it is available to users in the EU, whether its output is used in the EU, and whether it falls under Article 5 or Article 50. This check is performed system by system. This means that to reach the conclusion that the law does not apply to you, you in fact perform the same inventory, and the only question is whether you record its results. You must record them, because a conclusion without documentation does not protect you: under Article 85, a complaint can be lodged even against a company that considers itself outside the scope of the law. If this happens, a documented negative scope assessment for each system is the only way to respond to the authority quickly. The scope of the law can also change for your product over time. Your first client from the EU can appear, your feature can be repurposed (in this case Article 25(1)(c) makes you the provider of this feature), and your marketing department can describe the purpose of the product on the website in a new way. In each of these scenarios, your previous conclusion that the law does not apply to this product becomes outdated. If you do not update this list, nobody in your company will notice this change. EU AI Act Inventory The rules that apply from 2 August 2026 do not contain the obligation to keep an inventory of your AI systems: neither Article 5, nor Article 50, nor the guidelines of the Commission require such a list. But you need this list, for several reasons. You cannot check your systems against the eight prohibitions of Article 5 and implement the four transparency duties of Article 50 if you do not know which AI systems your company actually uses. When a complaint is lodged against your AI system, the burden of proof lies on you, and you receive a maximum of 15 working days to correct the violation. Within this period, you have no time to find out how many AI features your company runs and which model works behind each of them. This information must be collected in your inventory before the investigation starts. AI systems appear in a company not only through your own development plans. Your team can subscribe to an HR tool with emotion recognition, and your manager can connect a bot to customer calls, and in such scenarios your company starts violating the law without knowing about it. The inventory allows you to find such systems before the authority finds them. From 2 December 2027, high-risk AI systems must be registered in the public EU database, and the registration form requires information about each system. If you build this list now, you prepare this registration in advance. The obligations of the EU AI Act apply to each AI system separately and depend on its intended purpose. This means that the unit of your inventory is one AI system with its intended purpose, and not the model behind this system. Products that work on the same model can receive completely different legal outcomes under the EU AI Act. For example, three of your products can run on the same GPT-class API.  The first product analyses webcam footage and concludes which of your employees are unhappy. This product is a system for emotion recognition in the workplace, and Article 5 prohibits such systems. This prohibition applies in any case: informing the employees does not make this system legal.  The second product answers your customers in a chat. This product is legal, but under Article 50 it is obliged to disclose to the users that they are speaking with AI.  The third product classifies incoming emails as spam. This product is also an AI system within the meaning of the law, but no obligation of the law applies to it, because it does not fall under the prohibitions of Article 5, it does not fall under any high-risk category, and it does not speak with people and does not generate content for them, so it has no transparency duties under Article 50 either.  How to Make an Inventory of Your AI Systems The first step is to collect the candidates for this inventory. To collect them, you must request three lists: the list of all products and features with generation, chat, recommendations, scoring or recognition from your product managers, the list of every place where your code calls the API of OpenAI, Anthropic, Google or your own models from your developers, and the list of every purchased SaaS product with AI features (support bots, HR screening, call analytics) from your procurement department, because for these purchased products you are the deployer. The second step is to check every candidate against the official definition of an AI system. Whether your software legally qualifies as an AI system is determined according to the Commission's Guidelines on the definition of an artificial intelligence system (February 2025). According to these Guidelines, software that works on ordinary if-else rules and static formulas is not an AI system, and software that involves training or inference is an AI system. The third step is to run every system through the official EU AI Act Compliance Checker on the Commission's AI Act Service Desk. You answer the questions about one system and receive the result: whether this system falls within the scope of the law, whether you are its provider or its deployer, and which articles and obligations apply to it. The Compliance Checker is free, and its result must be added to your records. The fourth step is to put all these systems into a table with fixed columns. There is no official template for an internal inventory, but you do not need to invent the columns: you can take the structure of Annex VIII (the official list of fields that providers of high-risk systems submit to the EU database at registration) and reduce it to your needs. In this table, you must record the name of the system, its intended purpose, the model behind it, your role (provider or deployer), the availability of this system to users in the EU, the result of the screening against Article 5 with the date of this screening, the applicable duties of Article 50, the classification under Annex III, and the owner of the record with the date of the next review. The fifth step is to assign an owner of this inventory and to establish one rule: a new AI feature must receive a new row in the table before its launch. If you do not establish this rule, your inventory stops corresponding to the actual state of your AI systems within a month. EU AI Act Enforcement August 2026 News: What Applies and What Was Delayed After 2 August 2026, the key provisions of the EU AI Act were expected to become fully applicable, meaning fines and penalties from €7.5 million to €35 million (or from 1% to 7% of global turnover). And indeed, for AI systems the rules of Article 5 (the prohibited practices) and Article 50 (the transparency requirements) are in force as of 2 August 2026. This is the Act's enforcement start date. From that day, anyone who has grounds to believe that the AI you use breaks the EU AI Act rules can officially lodge a complaint against your business, and a national market surveillance authority can investigate your AI-based systems. And it doesn't matter where you're based: in the EU or, say, in the US. The Act applies to providers and deployers worldwide (lawyers call this its extraterritorial reach). But the implementation timeline for high-risk AI systems is on pause: the deadline for Annex III systems (hiring, credit scoring, education and so on) was moved from 2 August 2026 to 2 December 2027. The European Commission proposed the delay as part of the Digital Omnibus on AI, and the Parliament and Council agreed. The Omnibus also allowed a small concession for generative AI systems that were already on the market before 2 August 2026: the machine-readable marking duty of Article 50(2) doesn't have to be implemented until 2 December 2026. Belitsoft offers AI consulting services and provides senior AI developers with expertise and experience in architecting, engineering, and maintaining AI systems fully compliant with the EU AI Act. Let's talk Who Does the EU AI Act Apply To? The EU AI Act applies not just to the companies that train the models, as many would like to believe. According to Article 2(1) of the AI Act, with the definitions in Article 3, the affected include providers, deployers, importers, distributors and product manufacturers. Providers: anyone who develops an AI system (or has it developed) and offers it in the EU under their own name or brand. And if you build a product on top of a third-party model, even by simply calling the OpenAI API, the letter of the law considers you a provider too (a "downstream provider"). Deployers: companies and public authorities that use AI systems in the course of their business from an HR department running a screening tool to a bank using a chatbot. Importers and distributors of AI systems. Manufacturers that embed AI into their products: cars, medical devices, machinery. And location doesn't matter: a US or Chinese company falls under the Act as soon as its AI product is available to users in the EU, or as soon as the output of its AI system is used in the EU. AI Recruitment Tools under the EU AI Act: High-Risk Systems and Prohibited Practices AI recruitment tools are applications that analyze and filter job applications, rank candidates against a job description, and score interview answers. Founders and HR directors buy them to handle large volumes of applications, and developers usually build them on top of a ranking model trained on the client's own hiring history. The EU AI Act (Regulation 2024/1689) treats many of such tools as high-risk (read as "they cannot be sold in the EU without bias testing, technical documentation, audit logs and registration in a public database"), and some of them are prohibited.   AI emotion recognition in video interviews Owners of video interview products need to know which of their features can be sold in the EU at all.  The problem is not the use of AI in a video interview as such.  Most of the AI Act applies from December 2027, but the ban on emotion recognition in the workplace has been in force since 2 February 2025, with exceptions for medical and safety purposes only. There is no compliance path for a prohibited practice, and the penalty is up to €35 million or 7% of worldwide turnover. Simple example: a retailer runs a first round of interviews for shop floor staff. The candidate answers five questions on camera, and the product returns two scores. The first one rates the content of the answers against the job description. The second one rates confidence and enthusiasm from the candidate's voice and face. The first score is high-risk and can be brought into compliance. The second one has to be taken out before a client in the EU can use the product. The distinction matters because not every AI feature used in recruitment performs the same function. Scoring the content of an answer is one thing, but inferring confidence, enthusiasm or personality from a candidate's face and voice is another.   A feature must be removed if:  the model scores facial expression, micro-expression or gaze; the model scores voice tone, pitch, pace or hesitation as a proxy for confidence, enthusiasm, honesty and so on; the product builds a "cultural fit", "engagement" or "personality" score on top of any of the above. When a customer asks us to prepare a video interview product for the European market, we usually start by separating these scores from the rest of the pipeline. They are cheap to remove while the code is still in development. Differences between decision support and automated decision making With a decision support tool, the recruiter evaluates the data and makes the final decision. In an automated decision making system, the outcome is determined by the system without meaningful human involvement. A recruiter may formally approve the outcome, but does not independently assess each case. Almost every AI screening product on the market is sold as the first one: adoption of AI resume screening among HR teams went from 26% in 2024 to 43% in 2025.  Speaking to the BBC about the screening tools, Kelly Trindel, Chief Responsible AI Officer at Workday, describes how the company sees the division of responsibility: "Workday AI does not make hiring decisions and is not designed to automatically reject candidates. Customers retain full control and human oversight throughout their hiring processes."  Laura Holden, an AI lawyer and founder of the Bonsai AI and legal consultancy who ran responsible AI programmes at Unilever and Diageo, describes what happens on a recruiter's desk. Providers market these products as decision support tools, but they're designed and sold in a way that encourages employers to reject large numbers of applicants on the basis of AI-generated scores. So in practice they work as automated decision making systems. The GDPR defines profiling as the processing of personal data by automated means to evaluate certain personal aspects relating to a natural person. This includes aspects relating to work performance. This definition includes grading resumes, giving them grades from A to D, or ranking candidates by "suitability." Profiling deals with the assessment of personal characteristics and not merely with the use of personal data. Profiling is performed by an AI recruitment system automatically evaluating candidates' personal data and ranking them according to their suitability. Even if the final decision is officially approved by a recruiter, the system is therefore in the high-risk category. The EU AI Act includes, under Annex III, systems that analyze and filter applications and assess candidates. They are classified as high-risk systems. Article 6(3) provides for an exemption from the 'high-risk' category for some systems listed in Annex III where they do not pose a significant risk of harm to fundamental rights and perform one of four limited tasks. However, this exception does not apply if the system engages in profiling. Deployer Obligations (Article 26): When Using Someone Else's AI Falls Under the Act Article 26 is the checklist for deployers or companies that use someone else's high-risk AI system in the course of their business. If the system you use isn't high-risk, this list isn't about you: your duties as a deployer are in Article 50, and they apply from 2 August 2026. The Article 26 checklist, thanks to the Digital Omnibus, starts to apply only on 2 December 2027 (Annex III systems) or 2 August 2028 (AI embedded in regulated products). As a deployer of a high-risk AI system, you must: Use the system according to the provider's instructions for use, backed by technical and organisational measures. Assign human oversight to people who have the competence, training and authority to actually intervene. Make sure the input data you feed into the system is relevant and representative for the system's purpose (to the extent you control that data). Monitor how the system operates. If you suspect the system presents a risk even when used by the book, inform the provider and the market surveillance authority and suspend the use. A serious incident must be reported immediately. Keep the logs the system generates for at least six months (when the logs are under your control). If you're an employer: inform the affected workers and their representatives before switching the system on at the workplace. If your AI makes or helps make decisions about people (the Annex III cases: hiring, credit, education and so on): tell those people that AI is being used on them. Use the provider's technical information for your own GDPR impact assessment, and cooperate with the authorities. If you're a public body: check that the system is registered in the EU database (if it isn't - don't use it). And if you're a public body, a bank, an insurer or a private provider of public services: run a fundamental rights impact assessment before the first use and notify the results to the authority (Article 27). A deployer's obligations are light (Article 26). But Article 25(1) lists 3 actions that makes you a provider, with a provider's obligations: You offer someone else's high-risk system under your own name or brand. You buy a third-party credit scoring engine and sell it to your customers as "YourBank Score" (white-labelling), and now you're the provider of that system (unless your contract with the developer allocates the obligations differently). You significantly modify someone else's high-risk system, for example re-train it on your own data, and the result is still a high-risk system. The original provider's certification covered their version, not yours, so the modified system needs a provider of its own: you. You take a system that is not high-risk at all (including general-purpose AI like ChatGPT) and repurpose it so that it becomes high-risk. You integrate a general-purpose API into your candidate selection process. The system was just a chatbot, you turned it into an HR tool, and now you're the provider of a high-risk AI system. According to Article 25(2), once this happens, the original provider is no longer considered the provider of that system. OpenAI won't be liable for your CV-screening tool built on their model. As for banks and insurance companies, they already have internal control, monitoring and record-keeping systems in place under financial legislation (CRD, Solvency II and others), so they don't need to build a new control system for the AI Act duties to monitor the operation of the AI system and to retain logs: complying with their existing financial governance rules is deemed to fulfil these two duties (Articles 26(5) and 26(6)). Provider Obligations Under the EU AI Act (Article 16) Article 16 is the checklist for providers of high-risk AI systems. If your system is not high-risk, your duties are in Article 50. And thanks to the Digital Omnibus, high-risk providers now have time: these obligations start to apply on 2 December 2027 (Annex III systems) or 2 August 2028 (AI embedded in regulated products). A provider of a high-risk AI system must: Make the system meet the technical requirements for high-risk AI: risk management, quality of training data, technical documentation, automatic logging, instructions for deployers, human oversight, accuracy, robustness and cybersecurity (Articles 8–15). Put its name (or trademark) and contact address on the system, its packaging or its documentation. Run a quality management system covering the whole lifecycle (Article 17). Keep the technical documentation for 10 years after the system reaches the market (Article 18). Keep the logs the system generates — for at least six months, when the logs are under its control (Article 19). Pass a conformity assessment before the system reaches the market, draw up an EU declaration of conformity and affix the CE marking (Articles 43, 47, 48). Register the system in the public EU database (Article 49). If the system misbehaves: take corrective action, withdraw or recall it, inform the authorities and the distribution chain (Article 20). Demonstrate the system's conformity when a national authority asks for it. Meet the EU accessibility requirements for people with disabilities. EU AI Act Concessions for SMEs and Start-ups: What Applies in 2026 and What Comes Later The EU AI Act's impact on startups and small businesses is softer: the law comes with lower fines, free regulatory sandboxes and simplified paperwork. Here is what an SME gets and when. Who Counts as an SME Under the EU AI Act An SME here means fewer than 250 employees and up to €50 million turnover (Recommendation 2003/361/EC). You must count the employees and the turnover of your parent and affiliated companies too: if your company belongs to a large group, you're not an SME. EU AI Act Obligations and Timeline for SMEs and Startups: Key Dates The compliance dates for SMEs are the same as for everyone else: the Act doesn't give small companies extra time. Here is the timeline of obligations for providers, deployers and SMEs. The prohibited practices of Article 5 (in force since 2 February 2025). The transparency duties of Article 50 for providers and deployers (from 2 August 2026). The enforcement: anyone with grounds to believe you breach the Act can ask a national authority to investigate you (from 2 August 2026). Generative systems already on the market before 2 August 2026 have until 2 December 2026 to implement the machine-readable marking. High-risk AI systems: from 2 December 2027 for Annex III systems (hiring, credit scoring, education, etc.) and from 2 August 2028 for AI embedded in regulated products (medical devices, machinery, etc.). Lower Fines for SMEs and Start-ups For a breach of the prohibited practices of Article 5, a large company pays €35 million or 7% of global turnover, whichever is higher. SMEs and start-ups pay whichever is lower. The same reversal applies to the €15 million or 3% for the transparency duties of Article 50, and to the €7.5 million or 1% for giving incorrect, incomplete or misleading information to the authority. EU AI Act Regulatory Sandboxes: Free Priority Access for SMEs and Start-ups From 2 August 2026, every member state must run at least one regulatory sandbox, and SMEs and start-ups get priority access to it, free of charge. A sandbox means you agree a plan with the national authority and, for a limited time, develop and test your AI system under their supervision before it goes to market. What you get out of it: The compliance problems in your system are pointed out to you, with guidance on how to fix them before the system is on the market. No administrative fines are imposed for breaches of the Act, as long as you follow the sandbox plan and the authority's guidance in good faith. You get a written participation record. The authorities and certification bodies must take these documents into account positively. Simplified Documentation and Reduced Fees If you're an SME or a start-up, you get two concessions that start to apply on 2 December 2027 (Annex III systems) or 2 August 2028 (AI embedded in regulated products). You may draw up the technical documentation for your high-risk system (Article 11, Annex IV) in a simplified form. There is no official form yet. And the law says the conformity assessment fee has to be lower for an SME than for a large provider (though, as of July 2026, by how much is unknown). By the way, the Digital Omnibus extends the simplified technical documentation and the lighter quality management system (Article 17) to small mid-cap companies below 750 employees and €150 million in turnover. High-Risk Classification: How to Tell If Your AI System Is High-Risk (Annex III) There are two ways your AI system can turn out to be high-risk (Article 6). The first way is through product regulation. Your AI is a safety component of a product (or is itself a product) that already requires third-party certification under the EU product laws listed in Annex I: machinery, medical devices, toys, lifts, cars, aviation. If that's your case, the AI is high-risk — with the obligations starting on 2 August 2028. The second way is Annex III: a closed list of eight areas where the use case itself makes the system high-risk, with the obligations starting on 2 December 2027. Check your product against the list: Biometrics: remote identification of people, categorisation by sensitive traits, emotion recognition (a simple "is this person who they claim to be" verification is excluded). Critical infrastructure: safety components for digital infrastructure, road traffic and the supply of water, gas, heating and electricity. Education: deciding admission, evaluating learning outcomes, assessing what level of education a person can access, catching cheaters during exams. Employment: recruitment and CV screening, targeted job ads, evaluating candidates, decisions on promotion and firing, allocating tasks, monitoring employees' performance and behaviour. Essential services: eligibility for public benefits and healthcare, credit scoring (fraud detection is excluded), pricing in life and health insurance, classifying and dispatching emergency calls. Law enforcement: assessing the risk of someone becoming a victim, polygraphs, evaluating the reliability of evidence, assessing the risk of offending, profiling in criminal proceedings. Migration and border control: risk assessments of travellers, examining asylum and visa applications, identifying people (verifying travel documents is excluded). Justice and democracy: helping judges research facts and apply the law, influencing elections and how people vote. The risk class of your AI product does not depend on the technology inside it, but on the purpose you declare for it. If you offer your GPT-based tool as a writing assistant, it's low-risk. But if you offer the same tool for screening candidates, this is an employment use case from Annex III (recruitment and CV screening) and considers you the provider of a high-risk AI system, with the obligations starting on 2 December 2027. There is an exemption (Article 6(3)): if your system only does auxiliary work and doesn't materially influence the decision (converts CVs into a table, polishes a letter a human has written, flags patterns for a human to re-check, prepares materials for a human decision), it doesn't count as high-risk. But the exemption doesn't apply if the system profiles people, and any scoring of a person counts as profiling. If you decide under Article 6(3) that your system isn't high-risk, you must document that assessment before the system reaches the market and register the system in the public EU database (Article 49). The List of Transparency Duties (Article 50) Article 50 of the EU AI Act contains a list of 4 transparency duties that companies and public authorities must comply with when their AI-based software and devices talk to people, generate content or watch people. If caught violating them, they face fines of up to €15 million or 3% of their total worldwide annual turnover for the preceding financial year, whichever is higher (or whichever is lower for SMEs and start-ups). In practice, market surveillance authorities will interpret Article 50 in line with the European Commission's Guidelines on transparency obligations. Article 50 requires that all four disclosures must be made clearly at the first interaction or exposure: - An AI system that talks to people must tell them they are talking to AI (unless this is obvious). This is the duty of the provider. Examples of AI systems affected by these requirements according to the Commission guidelines on the interpretation of Article 50: AI voice assistant or chatbot speaks with users in customer support, complaints management, e-commerce, finance, healthcare, education or public services. AI hotline takes incident or fraud reports from people. Humanoid robot, cobot or AI companion interacts with people. AI avatar interacts with people in a virtual reality environment. AI bot speaks with people on social networks and media. Coding agent or other AI agent interacts with people directly, including in multi-agent architectures. Robotic companion pet looks and behaves so much like a real animal that people cannot tell whether they are interacting with AI. Realistic human-like avatar or voice speaks with users in VR or AR, where children, the elderly or people with disabilities cannot easily tell human from AI. Helpdesk chatbot on an online platform gives replies that users may take for human-written. Police chatbot on an official website lets people report criminal offences (crime-reporting systems are not covered by the law-enforcement exception). AI-powered telephone hotline of the police answers public inquiries. Fraud-reporting hotline of a bank or public authority collects reports of suspected financial crimes. Virtual assistant collects witness statements. The guidelines also explain how exactly AI systems must disclose themselves. Examples of disclosures that are enough according to the Commission guidelines: Chatbot starts the conversation by mentioning it's based on AI technology Provider places a label or banner "You are interacting with an AI system" in the interface Voice assistant says at the beginning of a session that it's powered by AI AI agent puts an AI label at the top of the email it generates Examples of disclosures that are not enough: Company mentions AI only in its documentation, terms and conditions Provider embeds a machine-readable watermark or metadata that users cannot see during the interaction Company calls the bot simply an "assistant" or gives it a human name and photo Website shows a generic "Services on this website use AI" notice Company states only the underlying technology ("this system uses LLMs") without explaining that the user is speaking with AI The AI system must also admit it's AI if a user asks about it or if the conversation suggests that the person believes they're speaking with a real person. The Digital Omnibus on AI also allowed a delay until 2 December 2026 for the machine-readable marking duty (Article 50(2)), and only for generative systems already placed on the market before 2 August 2026. The other three transparency duties and all new systems get no delay: they must comply from 2 August 2026. - An AI system that generates audio, images, video or text must mark its outputs in a machine-readable format, so the content is detectable as AI-generated. This is the duty of the provider. Examples of AI systems affected by this requirement according to the Commission guidelines: Chatbot or general-purpose AI system writes texts: articles, emails, reports, marketing copy. Image generator creates pictures from a text prompt. Video generator creates clips, 3D content or virtual reality environments. Voice generator clones voices or turns text into speech. Music generator composes songs. Photo or video editor removes, replaces or inserts objects and people, or swaps faces. AI agent sends people content it has generated: emails, documents, presentations. App creates digital twins: virtual replicas of real people or objects. Tool mixes AI-generated fragments with human-created material. This guideline also explains which changes to content require marking and which don't because they're simple edits. Examples of changes that require marking: AI tool summarises text. AI tool paraphrases or rewrites text, changing its style, structure or meaning. AI tool removes, replaces or inserts objects or people in images and video, or replaces faces. AI tool synthesises realistic speech in a specific person's voice. AI tool generates realistic video of events that never happened. AI tool alters a person's body shape or skin colour in a picture. AI tool makes extreme colour and contrast changes that alter the meaning of the content. AI tool builds composite images or clips that change how people, objects or events are represented. The marking requirement doesn't apply to source code, to outputs that only machines exchange with each other, or internal technical materials (such as engineering designs, technical specifications and documentation) to which only a pre-defined circle of professionals in your company and your client's company has access, and which are never shared outside. If you build your AI product on a model from a third-party developer and that model already has a watermark, you may use that watermark instead of your own. However, if that watermark fails, the fine is yours. So you must check that it actually meets the four legal requirements (effective, interoperable, robust and reliable) and be able to show this if the authority asks. If you wish to ensure that the regulatory authority accepts your marking, you may sign and follow the voluntary Code of Practice on Transparency of AI-Generated Content. - People exposed to an emotion recognition or biometric categorisation system must be informed that the system is running on them. This is the duty of the deployer. This duty covers any biometric categorisation, even one that isn't high-risk. If your camera simply estimates the age or gender of visitors, you must inform them. It doesn't matter whether the system works in real time or analyses recordings afterwards. And you must inform everyone exposed to the system, including children. Examples of informing people according to the Commission guidelines: Computer game shows a pop-up before launch saying that the player's face is recorded to capture their emotions. Exhibition hall places a visible notice at each entrance saying that visitors' faces are captured to assign them to an age group. The notice only needs to state the fact: the system is running on people. The AI Act doesn't ask you to explain why you process their data (the purposes, the legal basis and the rest are what your GDPR privacy notice must already cover), because GDPR applies in parallel with the AI Act. Note that informing people doesn't make the system legal by itself. Emotion recognition at work and in education stays prohibited under Article 5 no matter what notices you put up. - Deep fakes must be disclosed as artificially generated or manipulated (for evidently artistic and satirical works in a way that does not spoil the experience); the same applies to AI-generated text published to inform the public on matters of public interest (unless it has passed human editorial review and someone holds editorial responsibility). This is the duty of the deployer. Examples where you must disclose the AI origin:   AI generates a video of a celebrity influencer promoting a product. AI generates a realistic synthetic avatar of a company CEO congratulating employees on the corporate results. AI generates a product image for an ad or packaging that makes the product look better or different than it really is. AI simulates humans advertising a product in a teleshopping-style video. AI generates an image of celebrities implying their involvement in activities that never happened. AI-generated synthetic influencer tests a sponsored real product. AI generates realistic holocaust scenes that are shared on social media. Movie features de-aged actors or digital replicas of dead actors. AI generates music resembling the individual style of existing artists. AI-manipulated image places a politician in a scene that clearly mocks his policy decisions. Game imagery includes deep fake simulations of real people. AI generates a summary of an article about a town council decision, and the newspaper publishes it without editorial review. AI manipulates parts of a lifestyle article comparing the effects of diets on a disease. AI manipulates corporate reports with investor information on a listed company's website. AI generates a storm warning on a meteorological institute's social media profile. Website posts AI-generated articles on EU policy without any human review. Another AI reviews AI-generated articles, and a human editor only checks the grammar. Self-published AI-generated book on climate change goes to an e-commerce platform without any review. The duty applies even if you had no intention to deceive anyone, and the disclosure must be visible or audible to people. For evidently artistic, satirical or fictional works (a movie with digital replicas of actors, AI music in the style of a real artist, political satire, a game) you may disclose the AI origin in a way that doesn't spoil the work. The List of Prohibited Practices (Article 5) Article 5 of the EU AI Act contains a list of activities that companies and public authorities are banned from performing with AI-based software and devices. If caught, they face fines of up to €35 million or 7% of their total worldwide annual turnover for the preceding financial year, whichever is higher (for SMEs and start-ups — whichever is lower). In practice, market surveillance authorities will interpret Article 5 in line with the European Commission's Guidelines on prohibited AI practices. This is a 134-page official interpretation that helps anyone understand what exactly the wording of Article 5 means in practice. Article 5 lists 8 prohibited AI practices. Each point describes, in general terms, one category of prohibited practices without giving examples. In turn, the Commission's Guidelines give dozens of concrete examples of what is prohibited (we counted around 80). Article 5 prohibits placing on the market, putting into service or using: - an AI system that applies covert, manipulative or deceptive influence techniques to materially distort a person's behaviour so that they take decisions they would not otherwise have taken, causing (or having the potential to cause) significant harm to them or to others. Examples: A game with a brain-computer interface covertly trains players' brains to reveal sensitive information from their neural data (bank details, intimate information). AI system plays background audio or shows images that imperceptibly alter a person's mood, increasing anxiety and mental distress. AI system creates highly persuasive messages tailored to a person's data and vulnerabilities. AI system learns manipulative techniques on its own (from manipulative patterns in its training data or by gaming reinforcement learning from human feedback) even if the provider never intended it. AI chatbot impersonates a person's friend or relative with a cloned voice to run scams. AI system recognises when it is being evaluated and temporarily hides its undesired behaviour, resuming it once the evaluation is over. Advertising chatbot flashes brief visual cues and embeds inaudible audio signals to push people into harmful purchasing decisions without their conscious awareness. Wellness chatbot gives users dangerous health advice. AI chatbot nudges users towards self-harm or violence against others. AI companion app uses anthropomorphic features and emotional cues to make users emotionally dependent on the service. Chatbot offers fraudulent products. AI system uses addictive design that fosters compulsive use, anxiety and depression. AI systems facilitate harassment, gender-based violence and sextortion: deepfake porn and blackmail with AI-generated images. AI-generated deepfakes impersonate real people to deceive them. - an AI system that exploits people's vulnerabilities related to their age, disability or a specific social or economic situation to materially distort their behaviour, causing (or having the potential to cause) significant harm to them or to others. Examples: AI-powered toy encourages children to complete increasingly risky challenges (climbing furniture, handling sharp objects) in exchange for digital rewards and virtual praise. Game analyses a child's behaviour to create personalised rewards through addictive reinforcement schedules and dopamine-like loops, driving compulsive play. AI system targets older people with deceptive personalised offers or scams, exploiting their reduced cognitive capacity. Assistive robot forces older persons to do certain activities against their free choice. Therapeutic chatbot for people with mental disabilities exploits their limited intellectual capacities to push them into buying expensive medical products. AI system identifies women and young girls with disabilities online and targets them with more effective grooming practices. Predictive algorithm targets people in low-income postcodes with advertisements for predatory financial products, exploiting their financial desperation. Provider or deployer knows its system unlawfully discriminates against people in a vulnerable socio-economic situation and likely causes them significant harm and takes no corrective measures. Personalised chatbot identifies the discontent of migrants and pushes them towards extremist views, including violence. AI system generates child sexual abuse material or develops strategies for grooming and sexually extorting children. AI system uses addictive reinforcement schedules to keep young users dependent on the service, causing anxiety, depression, eating disorders, self-harm and suicidal behaviour. Anthropomorphic AI simulates human-like emotional responses with children, fostering unhealthy emotional attachment and distorting their understanding of real human relationships. AI system targets older people with expensive medical treatments, unnecessary insurance policies or deceptive investment schemes. AI-enabled differential pricing exploits people's economic situation, charging lower-income consumers higher prices for the same insurance coverage. Emotion recognition system meant to support mentally disabled individuals manipulates them into buying ineffective and expensive products with "unrealistic mental health benefits". AI chatbot targets socially or economically disadvantaged groups with fear-based narratives, inciting them to violence against others. AI system targets older persons with insurance offers by exploiting their reduced cognitive capacity. AI system uses imperceptible images to steer choices; emotion recognition covertly detects when a consumer is most susceptible and offers products at higher prices in that moment. - an AI system that scores or classifies people over time based on their social behaviour or personal characteristics (known, inferred or predicted), leading to worse treatment of them in contexts unrelated to where the data came from, or to treatment that is unjustified or disproportionate to their behaviour. Examples: Credit scoring system produces a "probability value" on whether a person will repay a loan by predicting that they will behave the way other people with similar characteristics have behaved (banned when the score uses unrelated data or leads to disproportionate treatment). Partly automated surveillance system at refugee camps (cameras, motion sensors) evaluates migrants over time to assess whether they are at risk of absconding. Public authority assesses people's trustworthiness combining an AI score with human assessment, when the AI score plays a sufficiently important role in the final decision. Public authority obtains a person's creditworthiness score produced by a separate specialised company (outsourcing the score does not escape the prohibition). National tax authority selects tax returns for closer inspection using not only income and assets, but also taxpayers' social habits and internet connections. Social welfare agency estimates fraud risk of benefit recipients based on having a spouse of a certain nationality or ethnic origin, having an internet connection, behaviour on social platforms and performance at work. Public labour agency scores unemployed people for state employment support using marital status, health data on chronic diseases and addictions. Public agency profiles families to detect "children at risk", and children are taken away over minor transgressions like occasionally missed doctors' appointments or traffic fines. Municipality scores residents' trustworthiness and people are blacklisted and lose public benefits for not returning library books on time, leaving rubbish out on the wrong day or paying local taxes late. Tax authority profiles child-benefit recipients into "deliberate intent / gross negligence" categories using low income and dual nationality, with benefits cancelled and families driven into heavy debt. Public authority scores fraud risk in student housing grants using applicants' internet connections, family status and level of education. Government runs a comprehensive citizen rating across all areas of life (social interactions, online activity, purchasing habits, bill punctuality) and low scores restrict access to public services, raise loan rates and block travel, renting and jobs. Insurance company takes spending data from a bank (unrelated to life insurance eligibility) and uses it to refuse contracts or set higher premiums. Private credit agency decides whether a person gets a housing loan based on unrelated personal characteristics. - an AI system that predicts the risk of a person committing a crime based solely on profiling them or assessing their personality traits. Examples: Law enforcement authority predicts criminal behaviour (such as terrorism) solely from a person's age, nationality, address, type of car and marital status. People are deemed likely to commit crimes they have not committed based on personal characteristics alone. National tax authority reviews all tax returns to predict criminal tax offences solely on an AI-built profile using personality traits such as double nationality, place of birth, number of children and opaque inferred variables that are hard to verify. Police risk tool scores children and adolescents for "future violent and property offending" based on their relationships. A child is deemed higher-risk simply for having a sibling or friend with a high risk score and children end up registered, monitored and referred to youth "care" services. Private company, asked by a law enforcement authority, analyses masses of data from national registers, banking transactions, communications and geo-spatial data to predict which individuals are potential human-trafficking offenders. AI system links location-based crime data to a specific person and scores their criminal risk solely on profiling, including the fact that they live in a high-crime area. - an AI system that creates or expands a facial recognition database through untargeted scraping of facial images from the internet or CCTV footage. Examples: Scraping tool avoids collecting "everyone" at once and instead harvests faces group by group (all residents of one city today, another tomorrow) until the database covers everyone anyway. System legally searches for specific individuals but along the way harvests and stores the faces of everyone else it encounters. Scraping facial images from CCTV footage captured by surveillance cameras in airports, streets and parks. Facial recognition software company scrapes photos of faces from social media (Facebook, YouTube, Twitter, Venmo) with an automated image scraper, stores them with the source URL, geolocation and sometimes names, converts facial features into hashed mathematical representations and lets a user upload anyone's photo to find a match (business model of Clearview AI). - an AI system that infers people's emotions in the workplace or in education institutions (except for medical or safety reasons). Examples: AI system infers that an employee is unhappy, sad or angry towards customers from body gestures, a frown or the lack of a smile. AI system infers from voice and body gestures that a student is furious and about to become violent. AI system infers emotions from the way a person types, their facial expressions, body postures or movements. Call centre uses an AI system with webcams and voice recognition to track its employees' emotions, such as anger. AI system monitors the emotional tone of hybrid work teams from voice and imagery in video calls, even when the stated purpose is conflict prevention. AI emotion recognition system evaluates candidates during recruitment. AI emotion recognition system monitors employees during their probationary period. Supermarket uses AI-powered cameras to track its employees' emotions, such as happiness. Education institution requires students to use a language-learning app with AI emotion recognition. AI eye-tracking software at online exams detects not only cheating but also students' emotional arousal and anxiousness. Education institution uses an AI system to infer students' interest and attention. AI emotion recognition system assesses applicants during admission tests. AI system monitoring online lectures detects students' emotional arousal, anxiousness and interest. Education institution runs an AI emotion recognition system on both teachers and students. AI system detects burnout or depression at work or in education (the medical exception does not cover this). AI emotion recognition system assesses employees' or students' wellbeing, motivation and job or learning satisfaction (wellbeing does not count as a medical reason). Employer's AI-enabled wearables measure employees' anxiety or boredom (allowed only where stress poses a specific danger, such as operating dangerous machines or handling chemicals). - an AI system that categorises individual people based on their biometric data (face, voice, etc.) to deduce or infer their race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation. Examples: AI system categorises social media users by their assumed political orientation, analysing biometric data from the photos they have uploaded, to send them targeted political messages. AI system categorises social media users by their assumed sexual orientation from photos shared on the platform and serves them advertisements on that basis. AI system estimates people's attributes such as age, gender and ethnicity from bodily features (face, height, skin, eye and hair colour) or singles out individuals by a specific feature such as a scar under the right eye or a tattoo on the right hand (banned when this is used to infer the sensitive traits from the list). AI biometric system claims to be capable of deducing a person's race from their voice. AI biometric system claims to be capable of deducing a person's religious orientation from their tattoos or face. The 8th point prohibits using (but not developing or selling) real-time remote biometric identification systems in publicly accessible spaces for law enforcement purposes (with narrow exceptions). Examples: Police use a real-time facial recognition system to identify a shoplifter by comparing their face against criminal databases (theft is not among the three permitted objectives). Police at a busy festival screen everyone passing the entrances against a watchlist of wanted individuals with outstanding warrants, hoping to spot criminals in the crowd (untargeted "fishing" deployment is banned). Police run real-time facial recognition in the streets for general security, crime prevention and overcrowding concerns: constant, time-unlimited surveillance of everyone.
Google Launched Antigravity 2.0 with Gemini 3.5 Flash. Will Agents Be 4x Faster?
Google Launched Antigravity 2.0 with Gemini 3.5 Flash
What Google announced at the Google I/O 2026 According to Google, 3.5 Flash "outperforms Gemini 3.1 Pro" on most benchmarks and generates tokens "4 times faster than other frontier models." Faster token generation reduces the part of an agent's time spent generating text, and for agentic coding that text portion is often large (long plans, lots of code output, multi-turn reasoning). So agents get faster, but not 4x. Gemini 3.5 Flash beats both Opus 4.7 and GPT-5.5 on MCP Atlas, Finance Agent v2, and MMMU-Pro, ties GPT on CharXiv, and tops GPT on Toolathlon: agentic tool-use, financial analysis, and multimodal. But Gemini 3.1 Pro is still winning long-context and hard reasoning. Google's pitch is that Antigravity 2.0 with Gemini 3.5 Flash can now operate many AI coding agents at once, faster and lower-cost than other frontier models. Building an app, cleaning up a codebase, preparing financial documents can now be done in a fraction of the time and at lower cost. The top AI Ultra plan dropped from $250 to $200 with the same 20x limits, so that plan got cheaper. The subscription got cheaper, but per-token API pricing for 3.5 Flash went up. According to early developer reactions, Gemini 3.5 Flash costs about three times what the previous Flash model charged for each output token: $3 to $9 per million. One developer ran the Artificial Analysis suite and found 3.5 Flash not just cost more than the older Gemini 3.1 Pro but also ranked lower, though another questioned that calculation. The only feature that users agree is better is speed. Gemini 3.5 Flash is fast, which matters if your application needs rapid responses. But "fast" is the only benefit most developers agree on. Whether it's cheaper or smarter is disputed. Several called it more expensive for less. A few users of Antigravity 2.0 were unimpressed. The old version that worked like VS Code was better and the new version is essentially a conversation with an AI agent. One developer called it a "cheap knockoff" of Codex that lost its identity. Beyond the Interface, Several Things in Antigravity Changed Specifically for Developers Antigravity CLI replaces Gemini CLI A CLI is a tool you type commands into in a terminal. Google had one called Gemini CLI. They've now made a new one, Antigravity CLI, and they're shutting the old one down for regular paying users (Pro, Ultra, free Code Assist) on June 18, 2026. If you use Gemini CLI, you have to switch. The new one is built in a faster programming language (Go) and can have several AI agents working in the background at once. Antigravity SDK An SDK is a set of building blocks (code libraries) you use to build your own software. This SDK lets a developer build their own custom AI agents using the same agent engine Google uses internally, and host them on their own servers instead of only inside Google's app. Why you care: before, you used Google's agent inside Google's product. Now you can build your own agent on top of Google's engine and host it yourself. Managed Agents in the Gemini API Normally an API call sends text and gets text back. A Managed Agent means one API call starts an AI agent that can think, use tools, and execute code by itself, inside its own isolated Linux computer in the cloud. You can come back later and continue the same session. You don't have to build the agent infrastructure (the sandbox, the tool-calling loop, state) yourself. One call, and Google operates the whole agent for you. The single thread connects all three: Google built one shared agent engine ("harness") and is now exposing it three ways: terminal (CLI), code library (SDK), and API (Managed Agents). Why Google announced this at the Google I/O 2026 Google is behind in the only AI category making real money, and is using its scale to undercut on price and expand the customer base from developers to everyone. The coding-assistant market is the most profitable AI category right now. Anthropic is on track for $45 billion annualized revenue, up 5x in a year, almost entirely from its coding model and Cowork product for business. Anthropic holds about 40% of the market. Google holds 10-15%. Despite Google and DeepMind leading AI research for years, OpenAI and Anthropic got to the money first, and Pichai has been criticized for letting it happen. So this launch is Google trying to take that revenue back. 3.5 Flash is pitched as matching Claude Code and OpenAI Codex but operating faster and at as low as half the cost. The Ultra plan dropped to $200 to line up with what OpenAI and Anthropic charge. Google owns the chips, data centers, models, and products, and is spending up to $190 billion in capex this year. That is how it can cut prices.
Google Makes Gemma 4 Up to 3x Faster Locally
Google Makes Gemma 4 Up to 3x Faster Locally
Gemma 4 is a popular tool among software engineers for tasks ranging from code generation and debugging to building complex agentic systems. According to Google, the Gemma 4 models were downloaded more than 60 million times since April 2, 2026. Recently, Google has added Multi-Token Prediction to its Gemma 4 family of open models to make text generation faster, achieving up to 3 times the speed without reducing quality.  Previously, generating text involved waiting for the AI to produce each token sequentially. With multi-token prediction, a lightweight drafter model proposes several tokens at once, which the main model verifies in parallel. This leads to faster response times. Since the drafters are open source under Apache 2.0, companies can run them locally. A Different Approach Than Gemini Gemma 4 and Gemini are built for different audiences.  Gemini operates in Google's cloud infrastructure with access to large pools of server-class TPUs. When users ask Gemini questions, the system allocates resources to run the calculations across multiple TPUs.  With Gemma 4, Google has refocused on the needs of developers who want AI that run on their own hardware. Gemma 4 (in its smaller variants) works for consumer grade GPUs SMBs have on their desk: dedicated gaming graphics cards or purpose-built AI accelerators. MultiToken Prediction speeds things up for locally hosted AI. Google also released Gemma 4 under the Apache 2.0 license.  Previously the company used more restrictive licenses that limited how developers could use the models. Now it is easier for businesses to build their own systems: fact-checkers, document summarizers, chatbots, code generators. This opens the door for a new wave of custom applications that business is built on-premise or in private clouds. The Bottleneck MTP Targets The main bottleneck for local large language models was moving weights throughout memory during generation of each word. Since every word requires running the model again to predict the next token, most time was spent waiting on data transfer.  When the model runs at home on a gaming PC or a server with limited video memory, it is often waiting for weights to get fetched from VRAM. You end up with high costs: either you wait long periods for AI to generate text or you need expensive workstation hardware with video cards that have thousands of dollars worth of VRAM to get any real work done. Google's MultiToken Prediction extends the inference speed of Gemma 4. Previously slow models can now handle real-time applications with multi-phrase conversations or text generation. How the Drafters Work The new drafting system has two text generation models: a main model (the target model) and a smaller model (the drafter). Quick token generation: When you ask it to write something, the drafter will quickly come up with a few tokens. Shared memory cache: Meanwhile the main model stores its previous attention computations in another 'bank' of memory called the key value cache. The drafter shares this cache, so it can generate draft text very quickly, without rebuilding from the ground up.  Intelligent mimicking: The drafter is less correct than the main model, so it guesses intelligently based on its training to mimic the main model. Simultaneous testing: The main model compares the drafter guesses against its own predictions and approves or not. It tests all candidates simultaneously. Parallel verification: These guesses are verified in parallel by the larger target model. Keeping the best guesses: When a draft token matches what the target would have generated, the system keeps the text. If every guess in a batch is right, all of them get accepted at once.  Dropping and proposing again: Otherwise it drops that token and everything after it, then the drafter proposes again from that point. The system performs better on speed while preserving output quality. Hardware Performance Gains On Google's Pixel phones, a 2.8 times speed increase on E2B models with the new drafting system. The same Pixel phones perform 3.1 times faster in the E4B configuration with the latest draft system. The new drafting approach is also available on Apple's M4 Silicon chips, where the latest 31B Gemma models run 2.5 times faster with the new speedup. Availability and Integration Open source licensed (Apache 2.0) new drafting Multi-Token Prediction (MTP) Gemma implementations are now available.  This means anyone can use them and change them for free. These reference implementations are already integrated with popular local inference frameworks: MLX, vLLM, SGLang and Ollama.  Today, developers can take the latest release of Gemma and plug drafting in without rewriting existing code or infrastructure. If you just want to evaluate the model, then log in at Google's AI Studio at aistudio.google.com and select Gemma 4 from the list of available models. If you want to run Gemma 4 on your laptop without any server setup, run it locally using Ollama. Gemma 4 can be fine-tuned to specialize for your domain. The recommended way is to use QLoRA with the TRL library. Businesses are already using Gemma 4  Businesses are already using Gemma 4  This is the first time a major AI lab has put frontier-capable AI in the hands of businesses without the need for cloud subscriptions or for data to leave the premises, which is changing AI affordability for SMEs. Running the model locally keeps the client data under the company’s control - an important requirement for GDPR and other regulations.  Gemma 4' function calls and JSON output allow firms to build AI agents that extract data from invoices, route support tickets or generate code without prompt engineering.  Different model variants suit different workloads:  E4B for low‑latency customer service and chat bots,  26B MoE for contract review and financial report analysis,  31B Dense for code generation and complex decision support,  E2B for factory‑floor monitoring and retail assistants.  For those finance leaders who cannot send general ledgers to public cloud APIs, Gemma 4's local deployment and native JSON functions offer automating invoice processing, revenue recognition and accounts payable while keeping payroll and vendor data air‑gapped.  Companies like WaveFlow use Gemma 4 to perform OCR on invoices, match them with local purchase orders and automate accounts payable entirely on local hardware.
Top-7 AI Voice Agents for Clinical Trial Screening
Top-7 AI Voice Agents for Clinical Trial Screening
1. Belitsoft's Custom AI Voice Agents for Clinical Trial Screening Belitsoft is an experienced custom AI agent development company that builds AI voice agents for clinical trial screening for each client individually. They don't force you to buy ready-to-use solutions that you must adapt to, but create tailored AI voice agents based on how your screening process works exactly. They don't sell all-in-one software with a lot of features where AI agents are just a part, but create separate AI voice agents on demand for each client's request and work directly with clinical recruitment platform vendors to create AI voice agents for their specific needs and integrate them into their CRM,  their customers' EHRs and other systems. If your business related to clinical trials relies largely on making a high volume of calls a week to check patient eligibility for trials, this is exactly what Belitsoft designs AI voice agents for. Their AI voice agents for clinical trial screening are HIPAA-compliant, and you have exclusive rights on it. There are no subscription fees or monthly payments for third-party products you never fully control. Let's, for example, take a look at' one of the recent case studies of Belitsoft from March 2026. Belitsoft engineered a custom AI voice agent for a clinical trial company. The AI voice agent can call a patient within 30 seconds after receiving information about them, confirm their identity, and speak with them. Throughout the call, the AI voice agent checks patient qualifications and writes the resulting data into the client database. Belitsoft engineers built a memory tool to track the conversation history. The client receives uniform, accurate data from every call. The AI voice agent evaluates patients in hours instead of days. It can dial hundreds of numbers simultaneously, which drastically reduces the need for human operators. Belitsoft developers programmed the AI voice agent with strict limits to protect patient information and comply with privacy laws. For this solution, their developers selected an independent HIPAA-based LLM vendor that specifically supplies compliant language models. The final AI voice agent software operates on Amazon Web Services and makes calls using Twilio. 2. Bond Health Bond Health provides an AI-driven, cloud-based patient recruitment and workflow automation platform that helps hospitals find and enroll patients for clinical trials. Bond Health plugs its software into the hospital’s medical database, like Epic and Cerner, using APIs, HL7, and HL7 FHIR communication channels. Hospitals can usually set up these systems in up to six weeks. Once connected, the software uses an LLM to scan the electronic health records. The AI reads the text and compares the information it reads against the rules of the clinical trial to find the individuals whose medical histories match the trial’s requirements. Automated voice programs call these individuals, ask questions from a prepared script, and arrange clinic visits. If an individual needs human assistance, the program transfers the call to a staff member. Bond Health stores its data on internet servers, securing access to it with passwords, user roles, and AES-256 code. The vendor states they follow medical privacy laws, such as HIPAA.   For payment, Bond Health charges a base fee to install the software. Then, they charge an additional fee for every patient who officially starts a trial. 3. Grove's Grace Grove sells a digital tool, Grace, to help hospitals run clinical trials. The company pairs Grace with a specialized database to manage patient contacts.  Grace is an artificial staff member that communicates with people over the phone, text messages, and email. The software answers common questions, evaluates if a patient qualifies for a trial, sets up dates for clinic visits, and connects the patient to a human when necessary. The software relies on approved medical scripts and displays data directly inside the main database system. Grove says the software speaks 14 languages and supports various accents. Grove relies on Google Cloud to store information. They keep each client separate and encrypt all data. The platform connects to Calendly, Outlook, Salesforce, HubSpot, and tools like Advarra Clinical Conductor, RealTime, and CRIO. Grove states they follow HIPAA, SOC 2 Type 2, 21 CFR Part 11, GxP, and ISO Artificial Intelligence standards. Grove does not share how much they charge for this product. Recently, Hippocratic AI purchased Grove. 4. TrialWire TrialWire operates a cloud-based conversational AI platform that works as a specialized CRM extension on Salesforce to help find patients for clinical trials.  The company recently introduced a software product they call AI Voice Screen. Their software uses artificial language models to converse with people over the phone and record their medical details. They run their platform on the Salesforce Health Cloud to follow medical privacy laws, such as HIPAA for the United States and GDPR for Europe. 5. SOGLIA's ELLA Soglia sells a software program called Ella to help hospitals talk to potential patients.  Their conversational AI and patient recruitment platform uses voice agents to hold conversations over the phone. It explains research studies to candidates, checks if they qualify, and directs them to doctors.  The company states their system holds realistic conversations and allows hospital staff to customize exactly what the system says to callers. 6. Power Power promotes a conversational AI and patient recruitment and  protocol-specific pre-screening tool to replace human call centers. Their automated voice programs call medical patients and ask questions to verify their health history.  The system records the conversation, organizes the patient's answers, and connects them to electronic medical records.  A human staff member makes the final decision on whether a patient qualifies for a trial. The company states their voice program finishes 98% of calls successfully, and the average call lasts 8 minutes. 7. Ushur Instead of a separate product just for voice calls, Ushur provides an AI agent-powered customer experience platform. They have a clinical trial automation module to help biological and pharmaceutical research staff automate clinical trials: recruit patients, evaluate candidates, and collect electronic consents. Their goal is to recruit patients faster, give research staff less work, and help hospital staff eliminate unnecessary phone calls to patients. The company asserts its software comes with built-in security and supports HIPAA, GDPR, HITRUST, SOC, and ISO 27001. To show their system works, Ushur promotes statistics like lowering patient dropouts by 15-20% and eliminating 42% of outbound staff phone calls.
Claude vs ChatGPT
Claude vs ChatGPT
The Pentagon Situation Anthropic doesn't allow use of its models for fully autonomous weapons or mass domestic surveillance. When the US military used Claude during a January raid to capture Venezuelan President Nicolás Maduro, Anthropic objected. The Trump administration then labeled Anthropic a supply chain risk and told federal agencies to stop using it. However, reports showed the US military continued using Claude for intelligence work, target selection, and battlefield simulations during the US-Israel strikes on Iran, despite the ban. Defense Secretary Pete Hegseth said the military would keep access to Anthropic services for up to six months while switching providers. OpenAI reached a new agreement to run its models on classified Defense Department networks. CEO Sam Altman said the deal still follows OpenAI's own rules, which ban mass domestic surveillance and require human oversight over use of force. Shift in the Market Many users disagreed with the OpenAI agreement. Around 700,000 users reportedly cancelled ChatGPT. Claude's iOS app hit the number one free app spot on the App Store. ChatGPT dropped to second, Gemini to fourth.  Anthropic says free users are up over 60% since January, daily signups tripled since November, and paid subscriptions more than doubled this year. Claude Outage Following the surge in users, today, on March 2, Anthropic's Claude went down. Users faced complete outages. Reports peaked at nearly 2,000 on Downdetector within a short period. In total, around 10,000 reports were submitted. Anthropic said the Claude API was working fine, and that the failures were limited to the claude.ai website. Switching from ChatGPT to Claude Most users avoid switching AI tools because they don't want to lose their conversation history. Anthropic built a few ways to fix this. The fastest option is the Import Memory feature. You paste a prompt from Anthropic into ChatGPT, copy the output, and paste it into Claude's memory settings. Done in under a minute. For a full data transfer, you can export your entire ChatGPT history through settings. It takes 24-48 hours to receive the zip file. You then upload it into a Claude Project, giving Claude access to your full history. You can also download individual ChatGPT conversations as markdown files and upload them directly into Claude. Claude vs ChatGPT Comparision Both ChatGPT Plus and Claude Pro cost $20 per month via web version. In terms of input tokens, GPT-5 is cheaper than Claude Sonnet 4.6. Claude Sonnet 4.6 is priced at $3 per million input tokens and $15 per million output tokens. GPT-5 starts at $1.25 per million input tokens and $10.00 per million output tokens, while GPT-5.2 is priced at $1.75 per million input tokens and $14 per million output tokens. However, Claude is better at coding, generates more cautious and structured answers, and works well with large documents and code. Claude Pro is a better choice for coding quality, and its output is often more natural and context-aware. Claude Pro is also well suited for large documents, long-form summarization, and prototypes, while ChatGPT Plus is better for image generation, video with Sora, and web research. In one comparison, Claude 4.6 Sonnet outperformed ChatGPT 5.2 on most tests. Claude did better at step-by-step reasoning, financial analysis, and tone control. ChatGPT does better at explaining complex topics in simple language and storytelling. For day-to-day professional work, reviewers currently rate Claude as the better tool.
Microsoft Copilot Tasks
Microsoft Copilot Tasks
Unlike traditional LLM-based chatbots, where you ask a question and receive an answer, Copilot Tasks agent completes tasks autonomously: browse the web, coordinate across applications, create documents, and managу schedules.  You describe what you want in plain language. It can be a one-time job, scheduled, or recurring, and the Copilot Tasks agent delivers a completion report when finished. As CEO Mustafa Suleyman puts it: "AI that talks less and does more, no complicated setup or coding skills required". It's designed for everyday users, not just developers or enterprises. There is no need to manually configure AI tools or MCPs (Model Context Protocols) that currently makes agentic AI accessible primarily to technical users. How Copilot Tasks Agent Works Copilot Tasks does not need full access to your device like OpenClaw, for example, because it runs on Microsoft's cloud servers using a remote computer and browser. Since the processing happens on Microsoft's servers, there is no significant battery drain while tasks are running in the background on your laptop, and your hardware is less affected. If an attacker gains access to Copilot Tasks, it can only affect Microsoft 365 and any connected accounts, not everything on your device. Copilot Tasks can work only with what is available in your Microsoft account or other services you have connected to it. What Copilot Tasks Agent Can Do According to Microsoft, Copilot Tasks can: prioritize urgent emails and draft replies each evening compile a weekly summary of how your time was spent recap meetings schedule recurring syncs turn a syllabus into a study plan with practice exams build slide decks with talking points from emails and attachments tailor resumes and cover letters to specific job postings track apartment and car listings, watch hotel prices and rebook when they drop, and gather quotes from local contractors or plumbers book rides timed to your flights and adjust for delays help reserve party venues find local tradespeople unsubscribe from promotional emails you never open and cancel subscriptions you're no longer using How To Control Copilot Tasks Agent We need to know the AI agent will not email the wrong person, charge our card without permission, or schedule meetings in the middle of the night. Microsoft built in a control step before the Copilot Tasks agent takes action.  Before it spends money, books a meeting, or sends an email, it stops and asks for your approval.  You can review what the Copilot Tasks agent plans to do and approve, pause, or cancel it.  That is how Microsoft aims to build a credible solution. Copilot Tasks Agent Alternatives Microsoft Copilot Tasks puts Microsoft into the autonomous agent market alongside OpenAI's ChatGPT Agent, Anthropic's Claude Cowork, Perplexity Computer, launched in February 2026 as a digital worker, and Google's Gemini, which is adding the ability to carry out supported actions in Android applications. AI agents do more than answer questions. They complete tasks, and some use more than one model. Competition in the AI agent market is shifting to the orchestration of large language models. For example, Perplexity Computer coordinates Claude, Gemini, Grok, and ChatGPT, selecting the model based on the task. Microsoft launched Copilot Tasks on February 26 as a research preview to a small group of users, with a waitlist open for others to join ahead of a broader launch. Microsoft has not provided a timeline for general release.
Agentic AI Coding: What Still Remains Expensive Amid a 90% Drop in Costs
Agentic AI Coding: What Still Remains Expensive Amid a 90% Drop in Costs
Benefits of Agentic Coding Engineering Cost Reduction Agentic coding has changed the software development market by cutting the labor cost of implementation for simple and internal tools. The cost of writing has dropped by up to 90% compared with similar work a decade ago.  Agentic coding excels at CRUD applications, simple web forms, standard workflows, small internal tools, simple test suites, and basic API glue code. For these types of projects, development costs have dropped approximately 90%. AI tools let companies delay hiring managers and larger engineering teams, because a small senior team can do more. Tools like Cursor plus Claude allow a single experienced engineer to generate output that used to require a small team. With these tools plus smaller teams, a handful of people can now achieve roughly an order of magnitude more than before. Engineering Time Savings For a typical internal tool where data modeling is already complete, work that previously required a small team can now be finished in a few hours with an agentic coding command line interface. AI coding agents like Claude Code can generate a full unit and integration test suite for a fairly complex internal tool in a few hours. The AI-generated test suite, which contained more than 300 tests, would have taken several software engineers several days to write by hand. A project that previously took roughly a month from start to release can now be completed within about a week when using agentic coding tools. New Market Opportunities For niche platforms, AI makes it economically viable to build a product where the total market size is no more than 10 million USD  in annual revenue and hiring a full team would not have been worthwhile. Faster Product Idea Exploration Agentic coding tools are extremely good at turning business logic specifications into well-written application programming interfaces and services. AI agents allow faster exploration of product ideas, which shortens the loop between product and engineering.  Pairing a business domain expert with a motivated software engineer and these tools produces an extremely powerful combination. Instead of a larger squad that pairs a business specialist with a group of software engineers, we will see much tighter two-person pairings. Such pairings allow extremely rapid iteration on software products. If the chosen direction is poor, the team can discard the current software and quickly start again, using what they have learned. In this new mode, the hard work is the conceptual thinking rather than the typing. Faster B2B SaaS Customization There is a lot of customization per client in B2B SaaS, so the fast conversion of unclear requirements into working prototypes that reveal misunderstandings is the biggest gain.  AI agents help most when software requirements are fuzzy, polish and long-term support are less critical right now, and you only need to iterate quickly by creating simple prototypes. Excel Spreadsheet to Web App Conversion Every organization has hundreds, and possibly thousands, of Excel sheets that track important business processes. Those Excel-based processes would be much better expressed as applications. In some cases, a professional development agency can turn these spreadsheets into an application for around 5000 dollars by combining a competent software engineer with AI tools. Cutting Recurring SaaS Costs with Self-Hosted Internal Solutions, Built Cost-Effectively with AI Agents As AI lowers the cost of custom development, SaaS tools that mostly wrap simple workflows are no longer viable as monthly subscription products. High SaaS subscription prices make it easier for companies to justify replacing them with AI-coded internal solutions. Some multi-billion dollar corporations are already replacing SaaS tools with custom internal solutions built with AI assistance. They:  replicate Salesforce-like features and embed them directly into internal systems to reduce costs. replace tools like Fivetran with internal ETL solutions based on open-source platforms and custom code, saving 40,000 dollars per month, reducing maintenance costs, and making customization inexpensive. rebuild key features of expensive back office SaaS in several weeks. For many internal workloads or custom mobile apps, it now makes sense to build rather than buy. Low-code platforms combined with AI agents enable companies to build business applications quickly, replacing subscription-based alternatives. Optimal Use Cases for Agentic Coding AI is especially powerful in small companies or teams that already have an engineering culture and testing practices.  Production of Small Applications Many developers now produce far more small internal or personal applications, even if those applications never become public products. Even if AI may produce boring, ugly code, it is still good enough for personal tools such as IDE plugins that saves developer time. Component Development Large language models are very good at writing small components from the bottom up and stitching them together.  A software engineer can feed a REST API specification to an AI coding tool and receive a module that largely works. They can also write documentation and explanations for protocols or complex code paths. Composing Small Parts With guidance, AI is very good at composing small parts, such as several API calls plus error processing, into a coherent routine.  Senior developers are comfortable with letting AI generate entire small components such as simple dialogs or small modules. Predictable Tasks, Known Patterns and Libraries AI tools work best when they can compose known patterns and libraries rather than invent new designs.  AI coding tools work excellent when the underlying task is predictable, such as generating wrappers or predictable user interface patterns. They are strong at identifying libraries that solve a problem when given freedom to choose the approach.  The Main Use Case: Understanding Legacy Code The main value of AI agents is not code volume, but having a second brain that thinks faster than a human developer. Senior developers use AI agents to understand large legacy codebases, not necessarily to write big changes autonomously.  Legacy projects are dangerous for agentic coding, because AI agents may generate large diffs that are hard to review, and lack of tests makes correctness very hard to assert. However, even if they are not good at generating new code in that environment, AI tools are excellent at parsing existing legacy code and using it to explore scenarios and hypotheticals.  AI can extract complex business logic even from obscure implementations such as templates and plugins. AI is also good at explaining low-level code, including assembly for retro platforms. What Makes Code Legacy When models are fed well-structured code even from legacy systems, they can help deliver changes faster without a large team. AI agents can easily read a specific function, understand what it does, and propose changes with high accuracy because there are boundaries in the well-structured code. However, most legacy projects are those that are not maintained and have little or no test coverage, and where engineers fear changing anything. Legacy is a business decision: a system becomes legacy when the business declares it obsolete and stops investing in it.  Poor engineering and mismanagement can create legacy code even on a brand new project, so age is not the key factor.  Most vibe-coded apps become legacy almost immediately, because no one wants to invest in cleaning them up. 10x Productivity Gains A 10x personal productivity increase may occur when working on a large legacy codebase if your code agent can: scan and understand big old codebases, answer questions about them, propose targeted changes, and assist with testing and debugging. AI is excellent at understanding existing code, summarising it, and answering questions, and this is the main productivity gain for work on legacy systems. AI is a crew of excavators compared to a single shovel when exploring large garbage heap codebases. Developers have spent a lot of time trying to understand codebases that are several years old. Agents make understanding these older codebases easier: explain what the code is doing, and locate bugs in that code.  New Automation Opportunities Many use cases senior developers previously would not have bothered to script or automate are now easy, and they have cranked out small scripts and small web services in several hours using AI. AI-Assisted Code vs. Contractor Code Developers would rather inherit a repository written with the help of an agent and a good engineer in the loop than one written by a cheap contractor of questionable quality from India who left several years earlier. The kind of repository left by such a questionable contractor typically has no tests and code is often a mess of classes and methods. What Remains Expensive in Custom Software Development Even with AI Agent Coding AI hype around code is comparable to the hype around self-driving cars: solvable in theory, but much harder than many assumed. In fact, coding time is often only a small fraction of the total time spent on a complex enterprise project. Hidden Costs The main costs are not in the initial building but in future maintenance, feature additions, operations, and organizational coordination, which AI reduces far less. Debugging and supporting software in production is still difficult and has not been automated by AI. Production software also has many hidden costs that include: security, upgrades and patches, hosting and uptime at scale, customer interactions, regulatory and compliance aspects, product management and design, and data migration and integration. The cost of having software today also includes the cost of dealing with cloud platform complexity, such as Kubernetes, distributed databases, queues, and multiple user interfaces. Distribution Still Costs More Than Building  AI agents can reduce building cost, but large companies still have advantages in brand, distribution, and customer trust.  Marketplaces and discovery are controlled by algorithms that favour big players and established brands. Many excellent small products will remain invisible until their features are copied by large companies.  You May Still Require a Team Decision makers in larger organisations are unlikely to trust a one-person shop for core systems, regardless of how cheaply that person can code. Coding time is often only a small fraction of the total time spent on a complex enterprise project where you may still require a team of human engineers. Yes, smaller than even a year ago, but still a team. Before building the app from the ground up, a team would set up continuous integration and continuous delivery, define and implement data access patterns and the core services. After that, the team would usually build a backend, dashboards and graphs for users. Near the end, the team would ideally add automated unit tests, automated integration tests and automated end to end tests to make the product fairly solid. The release of such a product, depending on the complexity, may happen even about a month after the work started. This description above only covers direct labour. Every additional person on the project adds coordination overhead that includes daily or regular standups, ticket management, code reviews, handoffs between frontend and backend contributors and waiting on other team members to unblock your work. Senior Engineering Effort is Still Required to Solve Complex Production Issues There is currently enormous value in having a human supervise the agent and check its work. AI makes it very easy to create huge volumes of code, but this can be dangerous because code for critical systems is considered a liability, and less code is usually better. When AI writes code, it is tempting to accept it quickly, but this can damage long-term quality and maintainability. With a senior engineer in the loop, AI agents can create very high-quality software very quickly. AI is best used as a partner, not as an autonomous agent. Senior engineers still need to design, review, and direct the work. Using AI for code generation requires the same effort in design, coding, and review, except the code under review is not their own. Many engineers use AI primarily for exploring codebases and libraries, generating first drafts of code or tests, refactoring proposals, explaining errors, and drafting documentation. They may ask AI to implement something and use the AI output only as motivation. Experienced software engineers never just copy AI-generated code as is, and always review and adjust it. They may ask the model to write some code, then ask it to list the top ten problems, and then ask it to fix the most important ones. AI agents are very good at checking code and then critiquing or fixing it, so engineers actively use this workflow. To get good results from AI, you need to learn how to ask good prompts, plan the work, and supervise what the model does. Experienced software engineers often start with a planning step before any code is written. They ask the agent to propose a high level plan so that the person and the model are aligned. A simple CLAUDE.md or instruction file can teach Claude Code how a specific project works and help it stop repeating the same mistakes over time.  Agents are claimed to be strong at writing large volumes of unit or integration tests quickly. However, there may be challenges with the quality of these tests. They can create a false sense of security. While a human writes a test based on requirements, the AI may write a test based only on the code it sees. If the original code is wrong, the test will also use the wrong logic to verify it - the test passes, but the bug remains. That's why AI agents work best when you already have good automated tests, because those tests keep the model in check and make its output more trustworthy. AI can replicate the behavior of people who put together StackOverflow snippets without fully understanding them. With AI, average software engineers lose the educational value of spending hours reading documentation.   The educational loss is a reason why AI is more useful for senior software engineers than for junior engineers.  Will AI Replace Software Engineers? Current models may be pretty bad at programming anything non-trivial and  "fix" issues by removing functionality. AI coding assistants may hallucinate functions or APIs that do not exist, use deprecated interfaces, reintroduce bugs during refactors, and remove tests silently. They may "fix" failing tests by changing the test instead of the code. However, many human programmers also do not manage complexity well, and for senior engineers there is no fundamental difference between improving large language models and training junior developers. Moreover, AI agents and models are still improving at a rapid pace. Existing benchmarks do not really capture how fast these agents and models are improving. Newer models will soon make the current ones look completely obsolete. Model providers are now actively filtering code training data for quality, so the new generation will use higher quality code than older models.  Even without AI, developers were always at risk of field disruption from new technologies. There will still be strong demand for engineers who can supervise AI, manage complexity, and align software systems with business goals. Knowing the business and industry specifics gives senior programmers an advantage over others in the era of AI-assisted coding. When you add business domain understanding on top of technical and agent skills (which architectural decisions fit a project, which framework to use, and which libraries perform best), it feels as if the 10X engineer has arrived. Such engineers move toward full stack and product engineering or roles with more business responsibility, propose solutions rather than just implement specifications, and engage with customers, sales, and marketing. Modern developers use AI to learn faster about business processes, regulations, and industry history (verifying information). AI is a partner that amplifies us. We plus AI are much more valuable than we or AI alone.
Composer LLM from Vibe Coding Platform Cursor 2.0: Cool or Overhyped?
Composer LLM from Vibe Coding Platform Cursor 2.0
What Is Composer LLM There are two common bottlenecks in agent coding: reviewing code and testing changes. Cursor 2.0 targets both. Its new agent-centered interface includes an option to switch back to the classic integrated development environment view. Cursor 2.0 adds a native browser tool. Agents can open a browser, click, analyze the user interface, take snapshots, detect issues, and iterate until the output is correct. The browser is used to fix and validate user interface changes that would otherwise require manual feedback. The developers promise that with Composer, programming tasks in Cursor now finish faster (often in under 30 seconds) and are completed more accurately in production.  The model can run unit tests, fix linter errors, and refactor code.  Cursor now allows up to eight coding agents to run at the same time, working either together or independently to plan, write, test, and review code. Each agent operates in an isolated workspace using git worktrees or remote machines. Developers can compare multiple results from concurrent agent runs and select the best output. The base model used for Composer remains undisclosed. It may be built on Qwen3-Coder or GLM 4.6 according to some observers.  It is known that Cursor did not train Composer from scratch but focused on reinforcement learning (RL) post-training on coding examples  rather than pre-training. How to Use Composer LLM Effectively Here are the timestamps for the practical workflows demonstrated in the video filmed by one of Cursor’s investors. Upgrading a project from SDK v4 to v5 (4:45 - 6:58) The presenter had an old software project. He told Cursor to update the entire project to use the new SDK and gave Composer a link to the migration guide. The AI agent read the instructions, created a to-do list for all the steps, then edited files across the repository, making changes in different places in the project. The Composer LLM in Cursor 2.0 understood how all the files in the repository (the project folder) were connected and made changes wherever needed. Running three models in parallel (progress UI) (6:06 - 7:41) When you work with Cursor, it just prints lines of text to a black screen (the "CLI" or Command Line Interface). This makes it hard to see what is actually happening. The presenter wanted to redesign this text output. Instead of a long scroll of text, he asked the AI to build a table that updates live. The table would show which tests are running, how many are completed, how many failed, and the average time taken. He asked three different AIs (Composer, Haiku, and GPT-5) to do the exact same task at the exact same time. This is a core new feature in Cursor 2.0. Some were faster (like Haiku), some were better at creative UI (like GPT-5). This parallel test lets the developer see which AI gives the best answer. The presenter pointed out that, for this task, Composer LLM was making mistakes. This is why you would want to run three models at once, in case one fails. Building a complete, working web application from an empty folder (14:19 - 15:23) He gave Cursor a one-sentence instruction: "Build me an app that is an image generator". It wrote all the code using a modern website framework Next.js  to build the app. This included all the backend setup files needed to manage the website's content, files, layout code, and settings. It then ran the installation commands, linted the code (ran a spell-checker for code to find any errors), and started the development server (ran the final command to turn on the website). At the end, a fully functional image generator website was shown in the browser. The presenter emphasizes how fast all this happens. A human developer would take significantly longer to set all of this up manually. Iterating UI styling (Composer, Haiku, GPT-5) (16:06 - 19:18) The image generator app that Cursor built in the previous step worked, but the presenter thought it was "so ugly". He gave the AI a command: "Rethink it. It should be minimal, nice-looking, and in dark mode". He gave that exact same prompt at the same time to Composer (Cursor's new, fast model), Haiku (another fast model), and GPT-5 (a model known for being very good at UI design). He got three completely different website designs. He was able to click through each one and compare them. He decided the GPT-5 version looked the best and, with one click, he applied those changes. Using the browser tool (19:20 - 20:48) The final result still was not perfect. The image was pushed to the side, and there was a header bar the presenter did not like. He told the AI to fix the layout: "Make sure the image output is centered on the page. You can remove the top bar if necessary". The presenter used the "@browser" command, which is like saying, "Do not just guess, open a web browser and look at the app yourself". Cursor debugged its own visual work. It opened the app in a new Chrome window and took a screenshot to analyze. After it understood exactly what was wrong, it wrote new CSS and HTML code to fix the layout.  Then Cursor checked the browser again to see if the fix worked. Cursor and Composer LLM Reviews Some users believe Composer 1 is like GPT-5, but completes tasks much quicker. They see the parallel models and multitasking feature as a major improvement. Long-time users say that Cursor keeps them from leaving by releasing an update right after they start thinking about canceling. Some have already canceled their subscriptions but now want to come back because of Cursor 2.0. However, according to some unhappy developers, Composer fails to update configuration files correctly, generates duplicate or broken code objects, includes Chinese characters in outputs, and produces poor pull request code rated 2/10 or 3/10 in security checks. The update broke some extensions, especially for C and C++ coding. They believe Composer is an early alpha: unstable and not worth upgrading yet. They think Cursor is still miles away from how their marketing describes it, because it is "just a VSCode fork with AI" that "does three things at once but none correctly". The integrated browser is appreciated for debugging. However, it does not behave exactly like a full, standalone Google Chrome browser. Users are frustrated about being forced to share data and telemetry by default. For a tool that reads a company's entire codebase, this seems like a privacy violation. Users are also worried that the tool could be using their private, proprietary code to train its models. Tab completion is an example where users agree it is good, but it is also too aggressive and annoying and interferes with typing. Users are looking at other tools (like Claude Code) and wishing Cursor had their features, such as slash commands and hooks. Naming the core feature "Composer", when a major, widely used PHP tool already has that name, is seen as confusing branding. Who Owns Cursor AI The San Francisco-based company Anysphere Inc., maker of the AI coding assistant Cursor, has surpassed $500 million in annualized revenue, mostly from monthly subscription plans. Most revenue comes from subscriptions. Pricing for individual users starts with the free plan (Hobby tier) and maxes out at the $200 per month Ultra plan. Business plans start at $40 per user per month for Teams. Enterprise clients get custom contracts with compliance, so prices vary. The cost of using the API is the same as for GPT-5 and Gemini 2.5 Pro: $1.25 per million tokens when you send the prompt to ChatGPT, and $10 per million tokens when it generates a response. The company’s valuation is $9.9 billion. It is considered the fastest-growing software startup ever, reaching $100 million in annual recurring revenue in 14 months. The previous record was 18 months, held by Wiz. More than half of Fortune 500 companies now use Cursor in some form, and over a million people use it daily, including coders at companies such as OpenAI, Spotify Technology SA, and Nvidia Corp. Cursor AI Alternatives The market is moving toward agentic coding. The AI plans and writes code across multiple files instead of just suggesting one line at a time. Cursor competes with GitHub Copilot, Claude Code, Windsurf, and Cline. Composer is Cursor’s feature for these agentic tasks.
Amazon Cut Tens of Thousands corporate Jobs to Invest more in AI Automation
Amazon Cut Tens of Thousands Jobs to Invest in AI Automation
Amazon staff reduction due to AI automation Andy Jassy, Chief Executive Officer, says the reason is to reduce the "excess of bureaucracy". The vision is to operate like the world’s biggest startup and to make the company leaner, with fewer layers and more ownership, so Amazon can move more quickly. However, the key reason may be the increased use of AI that cut jobs by automating repetitive and routine tasks. It seems that AI-driven productivity gains within corporate teams were sufficient for a substantial reduction in force. Amazon has long-term investments in building out its AI infrastructure and in the short term must offset costs. The company is expected to spend $118 billion in capital expenditures for the year, with much of it going towards building AI and cloud infrastructure. Beth Galetti, Senior Vice President of People Experience and Technology, says the reduction is necessary because this generation of artificial intelligence is the most transformative technology since the Internet, and it accelerates the pace of innovation across existing and new market segments.   Amazon had more than 1,000 generative artificial intelligence services and applications in progress or built, but that figure was a "small fraction" of what it plans to build. Amazon shares rose 1.2 percent to $226.97 on Monday following the report. The company appears to be expecting another big holiday selling season and plans to offer 250,000 seasonal jobs to help staff warehouses, among other needs, which is the same seasonal hiring level as in the prior two years. This contrasts large seasonal warehouse hiring with corporate reductions. Amazon Web Services Amazon Web Services, the company’s cloud computing unit, is affected among others. AWS reported second-quarter sales of $30.9 billion (17.5 percent increase year over year), and that growth was well below gains recorded for Microsoft Azure (39 percent and for Alphabet’s Google Cloud at 32 percent in the same period). This competitive pressure may be driving Amazon to restructure AWS. The division has been making headlines recently for a fifteen-hour internet outage last week that disrupted many widely used online services. Amazon Robots Amazon executives believe the company is on the cusp of a major workplace shift that will replace more than 500,000 jobs with robots. Robotic automation could allow the company to avoid hiring more than 600,000 people by 2033, even while selling twice as many products. Amazon’s automation team expects the company can avoid hiring more than 160,000 people in the United States by 2027. To mitigate fallout in communities that may lose jobs, Amazon policy avoids using words such as "automation" and "artificial intelligence" when discussing robotics and instead substitutes phrases such as "advanced technology" or the word "cobot" to imply collaboration with humans. Daron Acemoglu, a professor at the Massachusetts Institute of Technology who studies automation and who won the Nobel Prize in Economic Sciences last year, said that once companies work out how to automate profitably, the practice will spread to other firms.
ChatGPT Agent Builder: Main Module of the AgentKit Platform
ChatGPT Agent Builder: Main Module of the AgentKit Platform
Why LLM-Based Agents Matter Many try to do different things in ChatGPT, Claude, and Gemini but do not get the results they want. Not because these systems cannot do it, but because they lack correct instructions, do not call the required tools, or do not have access to the right internal files. LLM-based agents are here to solve this. What Can You Use OpenAI AgentKit Agents For? The main use case now is conversational AI chat agents: customer support in e-commerce, internal agents for knowledge base search, or data analysis. E-commerce Applications Why use an agent in e-commerce? How is it different from a regular text chatbot? An LLM agent converts a vague customer query into a precise, machine-readable command (JSON). The LLM agent understands that "I want to return this product" means the "returns" category. This allows you to launch traditional scripts with precision. The LLM agent can extract data from your personal database (if you grant access) and create a personalized response directly in the chat interface, rather than just providing traditional search results. The LLM agent outputs interactive elements (order forms, "buy" buttons) directly into the chat. The user can make a purchase without leaving the chat window. Other Scenarios Turn your ChatGPT projects into AI agents and workflows that your team, friends, family, or customers can use or buy If you have built ChatGPT projects that allow you to do things in seconds that previously took hours, thanks to how you trained them, you can now sell or share these as autonomous agents hosted on your website or behind a paywall. Create your own AI Brain Go to templates, find "Internal Knowledge Assistant". This can classify and answer employee questions. Or maybe you have a general audience asking questions or just want to remember everything from everywhere. Click on the Internal Knowledge Assistant template. It rewrites the user's question to make it better, so users do not have to be experts. It classifies the question: "Q&A," "Fact-finding," or "Other." Then it processes with three types of agents: Internal Q&A, External Fact-finding, or a standard agent. You can train each one. You can upload a knowledge base, or go to "Tools" and connect to MCP Server and add various tools - Gmail, Drive, Dropbox, third-party servers - so your team, clients, or audience can ask questions here instead of asking you directly. Example of Building a Classification Agent with ChatGPT Agent Builder The key advantage of agents is the level of control over responses they give you. Unlike classic GPTs, which rely on a single prompt and file uploads - leaving the model to interpret everything on its own - agents minimize errors. The Agent Builder can help you create a classification agent. When a user sends a request, the agent determines which of your company's information categories it belongs to. Use simple prompts like: "Classify a user's request into: manufacturing, inventory, or logistics". The model will assign it to one of those three categories. Choose your model and set how much reasoning it should do. Usually, leave "Low" reasoning turned on. Classification is not hard for GPT-5. More reasoning uses up more tokens and costs more money. Switch the output from "Text" to JSON. This makes sure the response fits your prompt exactly. The model will only return one of the three categories, nothing more. Next, add If/Else logic. That turns the JSON output from the classifier into a set of instructions. If the response says "manufacturing", your next agent knows to do one thing. If it says "inventory", it will do something else. Each output can trigger other agents or send the final answer to the user. Why Is an Agent Built with OpenAI AgentKit Genuinely New in Terms of Business Value? An agent created with Agent Builder offers something new for business: it combines the power of LLMs with control over logic and security, making it possible to deploy LLMs in corporate processes with a high degree of reliability. Security Guardrails Before Agent Builder, using LLMs in corporate systems always carried risks of unpredictable behavior, confidential data leaks (PII), or toxic content. Built-in Guardrail nodes (PII, Moderation, Jailbreak) are ready-made architectural modules that automatically filter both incoming and outgoing data. This makes LLMs safe to use in sensitive sectors (finance, healthcare, legal). Programmable Logic Agent Builder lets you program multi-step decision logic (via If/Else) based on the LLM's semantic interpretation. Businesses can create complex predictable workflows. The agent does not just "reply" - it acts on a business scenario like "classify the client" (LLM/JSON) -> "check status in CRM" (MCP integration) -> "initiate a return" (MCP integration) -> "send a form to complete" (Widget). The agent can execute transactional actions (book, buy, change status in CRM), becoming an active participant, not just an advisor, and changing states in external databases. Data Control Agent Builder solves the key corporate problem: it lets you use the full power of AI models while minimizing unpredictability. You can deploy an agent via SDK/API, integrating the Agent Builder core into your mobile apps, chat systems, or backend processes. Execution logic and confidential data remain under your control. Deployment via SDK/API means the LLM core processes and classifies data, but confidential data remains inside your secure server or database/CRM. Critical data (names, addresses, phone numbers, credit card info, client activity, order details, CRM interactions) never go to the LLM. They are filtered at your system boundary. The LLM receives only anonymized context to generate a response. How Do You Start Using Agents with OpenAI AgentKit? After creating an agent with Agent Builder, you either get code to embed the chat widget on your site, or use the Agent SDK (or API) to integrate the agent's features into your mobile app or backend system, including vibe-coded ones. What Exactly Does Agent Builder Allow You to Do to Create a Custom Agent on OpenAI LLM? Creating a custom agent in Agent Builder means building a workflow using a drag-and-drop interface. As before with ChatGPT customization, you create a System Prompt (role, persona, response style), and select the LLM. Once you complete basic customization (prompt, LLM, RAG), you begin to configure parameters that control the LLM externally. You connect the database to the agent, but now you do not pass personal data from this database to the LLM - they can be filtered by middleware on your server. This shift from prompt-only management to management via nodes is the key business value of Agent Builder. Workflow Configuration You force the LLM (Agent Node) to interpret a vague customer request ("I want to buy," "how do I return") and output a structured variable (for example, "pathway": "purchasing"). With If/Else nodes, you use this JSON variable to guide the conversation down a custom path: if pathway = "purchasing", activate the "Purchase Agent"; if "returning", activate the "Return Agent". You create or upload unique interactive elements (widget files), such as a checkout widget or purchase-complete widget. This allows the agent to do more than reply with text. The LLM-based agent (Agent Node) is trained to fill these widgets with data from the conversation or your knowledge base. For example, it inserts the product name and price into the order widget, making the widget unique to the current dialogue. Modular Agent Design Instead of a single "all-knowing" GPT, Agent Builder lets you create multiple Agent nodes, each a narrow specialist: Greeting Agent (focuses on tone and classification) Purchase Agent (focuses on product knowledge and sales) Return Agent (focuses on policies and procedures) This modularity allows precise control over instructions and behavior at every stage of the customer journey. Agent Optimization Platform Agent performance optimization is part of a broader set of tools called Agent Kit. Optimization is done on the agent optimization platform, which includes evaluation, trace grading, and datasets. The main issue with complex agents is that they work well 80% of the time but "break" or hallucinate the other 20%. Evaluation tool automatically tests the agent on large datasets to check if it always returns the correct answer and uses correct logic. Trace Grading helps you see step-by-step how the agent arrived at an answer. If it made a mistake, tracing shows at which node (If/Else, Agent, Tool Call) the error happened. Reliability also means cost savings. If the agent gets it right the first time, no extra logic loops or retries - lower token use. By analyzing traces, you may find that the agent takes a long, expensive path for a simple task. You can fix the logic to make the path shorter and cheaper. Agent Builder VS n8n Agent Builder is designed for chat-oriented workflows. n8n is designed for automation. Agent Kit is for consumers and teams in the OpenAI ecosystem needing quick, simple chat workflows. n8n is for developers needing complex, custom, autonomous systems. Background Automation The most powerful automations work in the background, invisibly. In OpenAI's Agent Builder, you are limited in triggers and cannot schedule or run things in the background - n8n clearly wins here. You only have one trigger for Agent Builder - Start. It is for chat agents. The only way to launch the agent is to send a message (via chat or API). No scheduled triggers, no app events, no web hooks for CRM events. If you need an agent to process invoices or track leads automatically, you need real triggers. Agent Builder does not have them (at least, not now). n8n wins for background autonomy. Chat Interface Excellence Chatkit is where Agent Builder truly shines: easy creation of stylish, branded chat interfaces, Widget Studio for interactive visual elements (forms, calendars, product recommendations) that the agent fills with data. Professional, deployable chat interface in minutes without coding. Ease of Use vs Control For a complete beginner making their first agent, Agent Builder is unbeatable in simplicity. It is faster, less intimidating. But this simplicity comes with trade-offs. In Agent Kit, it is difficult to track how variables move between nodes in debug/evaluation mode - easy to see in n8n. n8n gives you full control of infrastructure (self-hosting), model choice, and cost control. Agent Kit is locked to OpenAI cloud only, but allows connection to OpenRouter for hundreds of models via one API. Local models supported in self-hosted n8n. Technical Knowledge Required To work effectively with Agent Builder, you need to know or understand core programming and development concepts Code Logic Basics (System Concepts) How to build conditional logic for workflow routing What "true" and "false" are, how agents use them for If/Else decisions What repetition is and the risk of infinite loops or credit consumption Data Structures and Variables How to save and reference data at workflow steps What JSON is - structured data for formatting agent output for next node use Why data must be manipulated for readability and usability at each step Architecture and Deployment Understanding that LLM is only one part, integrated with external tools and controlled by logic and security nodes Deployment (Chat Kit, SDK) requires coding knowledge (e.g., avoiding UI widget breakage with bad output) To work effectively, you must think in terms of software, data flows, and logic, not just writing a prompt.
OpenAI’s ChatGPT Agent Outperforms the o3 Model Alone: What This Means for Developers
OpenAI’s ChatGPT Agent Outperforms the o3 Model Alone
ChatGPT Agent Meaning ChatGPT Agent operates on its own virtual computer accessible from within the ChatGPT.com interface and alternates between reasoning and direct action to complete complex, multi-step tasks.  Unlike a standard chatbot, it can click and scroll through pages, fill out forms, download and manipulate files – all in response to natural language instructions.  At the core of this capability is a model that combines the web-interaction skills of OpenAI’s earlier "Operator" prototype with the deep analytical and synthesis skills of the "Deep Research" mode.  ChatGPT Agent News and Availability Click the "+" icon to the left of the prompt field, open the Tools menu, and then choose "Agent mode" (it’s the last item in the drop-down list). If you don’t see "Agent mode" there, the feature hasn’t been rolled out to your account yet, and there’s currently no other way to enable it. Not everyone can try it yet. On launch day OpenAI began rolling the feature out to paying Pro, Plus, and Team subscribers. Enterprise and Education customers are queued for "the coming weeks", while free-tier users must wait. European Economic Area and Swiss residents are still in the waiting room while regulatory work proceeds, so if you don’t see "Agent mode" in your Tools menu, the feature simply hasn’t reached your region or plan. OpenAI ChatGPT Agent EU availability Users based in the EU, who have a Pro subscription, on the evening of 18 June still can’t see the new feature. Pro users can try a stop-gap at https://operator.chatgpt.com, but this isn’t the announced ChatGPT Agent — just one component: a simple browser agent that loads pages and executes tasks. It’s quite buggy, for example, CAPTCHAs on websites can’t be solved within the agent’s interface, either automatically or manually. We’ll see how the full ChatGPT Agent handles jobs like these once it finally rolls out. Deep Research mode is available from the Tools menu, but it runs more slowly and cannot click through sites as mode "Operator" does. OpenAI has not committed to a timeline. Tech press summaries simply repeat that official EU access is "coming" but "indefinite for now". However, if you need unofficial access and has PRO account (not PLUS), just use an American VPN. Even if you have a European card, it still works. However, this method is unofficial and access could be blocked at any time. OpenAI ChatGPT Agent availability UK Many UK Pro users received access the day the feature launched. UK users get the same virtual computer with visual/text browsers, terminal, and connectors described in the launch post. OpenAI. If "Agent mode" still doesn’t appear, give it a little time - OpenAI is enabling accounts in batches. Capabilites the new ChatGPT agent brings to web users of ChatGPT.com Real-World Utility Examples ChatGPT Agent’s unified skills significantly broaden its usefulness in both professional and personal contexts.  Professional tasks: It can automate burdensome office chores, such as updating spreadsheets with new financial data while preserving all formulas and formatting. Instead of you manually doing these repetitive tasks, you can delegate them in natural language and the agent will handle the execution end-to-end. Personal tasks: In everyday life, it can serve as a smart concierge. For example, it can find medical specialists and book appointments for you based on your schedule. It can orchestrate complex personal projects or errands that involve multiple steps and information sources, saving you a lot of time and effort. User Control, Oversight & Privacy As ChatGPT Agent works, it provides a live narration of its actions and explicitly asks your permission before taking any step with real-world consequences (such as making a purchase or sending an email).  You can intervene at any time – pausing or stopping the agent, taking over the browser manually, or steering it in a different direction.  Certain critical actions run in a supervised "Watch Mode" that requires you to approve each step. Likewise, the agent is trained to proactively refuse obviously high-risk requests – for example, it will decline to carry out a bank transfer or other sensitive financial operation on its own. If a task is long-running, you’ll get a mobile push notification when it finishes so you can review the results. Strong privacy controls are also built in. With a single click, you can wipe all browsing data the agent has accumulated and log it out of any websites it accessed on your behalf. (Otherwise, site cookies persist according to each site’s policy, which helps speed up repeated visits by avoiding extra logins.)  In secure browser takeover mode – when you temporarily take control to enter credentials or navigate – anything you type (like passwords) is kept private and not stored or sent to the model. This ensures the AI never actually sees your sensitive inputs, since it doesn’t need them, and thereby prevents those details from being retained in the AI’s context.  You can also disable any connected apps or accounts (connectors) at any time to limit what data the agent can access.  Finally, ChatGPT Agent supports scheduling recurring tasks: once the agent completes a task, you can set it to run automatically on a schedule (for example, "generate my weekly metrics report every Monday morning") to further automate your workflow. Built-in Tools, Connectors & Environment To accomplish its tasks, ChatGPT Agent has access to a suite of built-in tools and can choose the best tool for the job.  These include a GUI visual web browser (to interact with websites like a human would), a text-based browser (for quick retrieval and reasoning over large text documents), a full terminal for code execution, and the ability to make direct API calls.  The agent can also use ChatGPT connectors – integrations to third-party services such as Gmail or GitHub – which allow it to securely fetch information from your email, calendars, or other apps and incorporate that data into its responses. If needed, you can hand the agent an authenticated session on a website via a secure takeover mode (logging it into a private portal yourself) -  this lets it access gated or user-specific content, and you can resume manual control at any moment. All these tools run within one coherent virtual computer environment that preserves state across tool switches. This means the agent can move from one modality to another without losing context.  For example, it might retrieve information through an API (say, querying your calendar events), then use the text browser to reason over a lengthy document, and finally switch to the visual browser to interact with a site’s interface – all within one continuous session.  The virtual machine allows it to download files from the web, manipulate them via terminal commands or code, then open the results back in the browser for you to view.  This adaptive approach enables the agent to choose the most efficient path to complete tasks: it might process structured data via API, handle large text analysis in the text browser for speed, and use the visual browser only when necessary for human-oriented pages. By preserving memory and state, the agent can carry out multi-step workflows with speed and accuracy, without redundant back-and-forth or losing track of intermediate results.  You can interact with it in a highly iterative and collaborative way – you can interrupt at any time to give new instructions or clarifications, and the agent will incorporate your feedback and continue the task without starting from scratch. Similarly, the agent may ask you for additional details if needed to ensure it’s on the right track. Risk Mitigation & Safety Framework Because this is the first time ChatGPT can directly take actions on the live web, OpenAI has implemented extensive safety measures.  The controls from the Operator preview have been expanded to address the new risks of a general-purpose agent with web access. This includes safeguards around handling sensitive user data, limiting what it can do via the terminal (network access is restricted), and especially defending against prompt-injection attacks that could be hidden in webpages or metadata.  Prompt injection is when malicious instructions are embedded in content the agent might read (for example, invisible text on a webpage) to trick the AI into doing something harmful or revealing private information.  To counter this, ChatGPT Agent is trained to recognize and resist malicious or hidden prompts, and OpenAI runs a real-time monitor on all tool outputs to catch suspicious patterns.  If the agent encounters anything that looks like a hidden instruction or an attempt to hijack its behavior, it will refuse or seek confirmation.  Moreover, the system requires explicit user reconfirmation before any high-impact action (like those involving sensitive data or transactions), which adds a human check that further reduces the chance of a hidden prompt causing damage. Users are also advised to be mindful of what data they expose the agent to and to disable connectors when not needed, as additional precautions. Given the agent’s expanded capabilities, OpenAI has classified ChatGPT Agent as having "High Biological and Chemical Capability" potential, so they have nonetheless activated the highest tier of safety measures for this category. That safety stack includes threat modeling for misuse scenarios, special training to refuse or flag any dual-use dangerous content, always-on classifier systems and reasoning monitors watching the agent’s behavior, and strict policy enforcement pipelines to intercept potentially harmful actions.  OpenAI has also been coordinating with external experts in biosecurity, government and academia – even hosting a biodefense workshop with national labs and NGOs – to get outside input on shoring up these defenses.  As part of the safety push, OpenAI has published a detailed system card describing the agent’s risk analysis and mitigations, and launched a public bug bounty program to encourage outside researchers to find any vulnerabilities or unintended behaviors so they can be fixed quickly. Availability, Roll-out & Quotas ChatGPT Agent became available starting July 18, 2025 for subscribers on the Pro, Plus, and Team plans. Pro users (who pay for the highest tier) got access immediately on launch day, while Plus and Team users were slated to get it over the following few days as the rollout progressed.  Enterprise and Education tier customers are expected to gain access in the coming weeks after launch.  Usage of the agent is metered: Pro subscribers receive a generous quota of 400 agent messages per month, whereas Plus and Team users get 40 agent messages per month included. If users need more, OpenAI allows purchasing additional agent message capacity via a flexible credit system. (These quotas are separate from regular ChatGPT usage and exist because agent tasks consume more resources.) Currently, the feature is not enabled in certain regions – pending regulatory clearance or additional compliance measures. OpenAI has acknowledged this regional hold and will presumably turn it on in those locations once they address local data and privacy requirements. The introduction of ChatGPT Agent also means some earlier beta features are being retired or folded in. The standalone Operator preview site (operator.chatgpt.com), which previously demonstrated the web-browsing agent capabilities, will remain live only for a few more weeks and then be shut down. Its functionality is effectively merged into the new agent mode. Additionally, the original "deep research" mode (which provided very lengthy, in-depth research answers with extensive citations) has been integrated as part of the agent’s capabilities. For users who prefer the old deep-research style – which can take longer but gives more exhaustive reports by default – that option is still available in the ChatGPT UI by selecting "Deep Research" from the mode dropdown in the message composer.  Benchmark Performance Highlights OpenAI tested ChatGPT Agent on a variety of challenging benchmarks and the results show substantial performance gains over previous models: Humanity’s Last Exam (HLE) On this extremely difficult expert-level QA test across many subjects, the agent achieved 41.6% accuracy (pass@1), which is a new state-of-the-art and a huge jump from prior models. This beats the previous best from the dedicated Deep Research mode (26.6%) on the same benchmark. In other words, ChatGPT Agent answers correctly on roughly 4 out of 10 graduate-level questions, compared to only about 1 out of 5 for its predecessor – a significant improvement on a test explicitly designed to be harder than what any AI had seen. FrontierMath (Tier 1–3) On the hardest open-ended math problems (novel, unpublished challenges that even human math experts struggle with), the agent scored 27.4% accuracy. This vastly outperforms earlier models – for context, the previous o3 model was around 10.3% on these problems. Having an agent that can solve over a quarter of these elite math problems is a major leap forward (prior systems were in the single-digits or low teens percentage range). Complex Knowledge-Work Tasks In an internal benchmark mimicking real-world knowledge work (conducting research and analysis tasks that might take a human hours), ChatGPT Agent’s outputs were rated comparable to or better than expert human results in roughly half the cases. It also significantly outperformed the older models (o3 a) on these tasks across all tested time budgets. In other words, given difficult professional assignments (like writing an analytical report or creating a detailed project plan), the agent’s work was on par with human experts about 50% of the time. This demonstrates that the agent isn’t just good at quizzes – it can handle practical, economically valuable tasks and often do as well as top human professionals in those domains. DSBench (Data Science Benchmark) This benchmark tests realistic data analysis and data modeling workflows. ChatGPT Agent not only surpassed other AI models here, it even exceeded human performance on some metrics. For the data analysis portion, the agent scored 89.9%, versus around 64% for human analysts performing the same tasks. On data modeling tasks, the agent reached 85.5%, well above the human baseline (around 65%). These are remarkable results – the AI is actually outperforming experienced humans in interpreting and modeling data in this test, highlighting how effective the agent can be when it can use tools (like Python) to assist its analysis. SpreadsheetBench This is a suite of 912 spreadsheet editing questions (requiring formula adjustments, data manipulations, etc). ChatGPT Agent achieved 35.27% accuracy overall on this benchmark – which may sound modest, but it’s significantly higher than other AI models: OpenAI’s o3 managed 23.3%, Microsoft’s Copilot in Excel about 20.0% on the same test. Moreover, when the agent was allowed to directly edit Excel files (rather than just instructing changes), its score jumped to 45.54%. While still below the 71.3% human expert accuracy on these tasks, the agent outperformed specialized tools like Copilot. It also maintained formulas and formatting, not just values. Investment Banking Financial Modeling On a set of first- to third-year investment banking analyst tasks ( building a three-statement financial model for a company, complete with correct formulas and citations), ChatGPT Agent averaged 71.3% accuracy. This is a strong result considering humans are far from perfect on these complex multi-step problems. It substantially beat the older "Deep Research" agent mode (which scored ~55.9%) as well as the base o3 model (~48.6%) on the same evaluations. This means the AI can perform many financial analysis tasks at a level that approaches (and in many cases exceeds) a well-trained junior analyst – a notable achievement in a field that requires both domain knowledge and attention to detail. BrowseComp (Agentic Web Browsing) This benchmark measures how well an AI agent can find hard-to-locate information on the web. ChatGPT Agent set a new record with 68.9% success on BrowseComp, which is about 17.4 percentage points higher than the previous best achieved by the Deep Research mode (51.5%). It also well exceeds the older o3-based agent’s performance (49.7%) in this category. The agent is much better at using a browser to answer obscure questions or gather information than past models – an indication of its improved search strategies and tool use on the open web. WebArena (Real-World Web Tasks) In the WebArena benchmark (which tests an agent’s ability to complete practical tasks through a web interface), ChatGPT Agent scored 65.4%. This is a few points higher than the previous agent built on the o3 model (62.9%), showing that it has surpassed prior agents in real web task performance. However, it’s still below the human performance level on WebArena (humans score about 78.2% on these tasks). So while the agent is closing the gap in web task proficiency, there remains room to improve before it can match an expert human operator in all cases. API Equivalents ChatGPT Agent can click around websites, scroll, fill forms, run code, and complete end-to-end tasks on its own. The Responses API already provides a computer_use tool (since March 2025) that lets developers programmatically control a headless browser. In other words, developers have had this capability behind the scenes. The flashy on-stage demo is built on tech that third-party devs already use. However, now that it’s packaged neatly in ChatGPT’s UI, users will expect the same seamless automation from all agents. Developers need to match that user experience in their products (or focus on specialized niches where they can do even more). ChatGPT has one-click connectors to Gmail, GitHub, Google Drive, etc., allowing it to directly interact with those services. There’s no official Connectors API for developers (yet). Developers can integrate services by writing their own OAuth flows or using unofficial means (the Deep Research agent or community plugins), but OpenAI hasn’t released a standard connector framework in the API. This is a missing piece in the developer stack. Until OpenAI offers connectors to everyone, custom-agent developers can add value by building these integrations themselves. If you build a domain-specific agent (say, for finance or healthcare), you can create connectors to the services your users care about. You’ll be stepping in where the official platform currently doesn’t go, which can be a competitive advantage. (Keep in mind OpenAI has hinted connectors will come to the API once stabilized, so plan for that future.) ChatGPT now supports scheduled runs (recurring Tasks) with notifications (like a weekly report every Monday). There is no built-in scheduling or cron feature in the API. A developer using the API must implement their own scheduling system (using cron jobs or backend services) or wait for an official "Tasks SDK" that might be in the works. If your agent’s value proposition involves doing things on a schedule (daily summaries, periodic reports, etc.), you currently have to handle scheduling logic yourself. This is extra engineering work, but it also means you can differentiate on reliability and customization of scheduled tasks. The fact that OpenAI showed scheduling in the product hints that a developer solution might come, but until then, consider it a space to innovate. The ChatGPT Agent emphasizes user control, with live narration of its actions, a ‘watch mode’ for sensitive tasks (like sending emails) that requires user approval, and a privacy toggle to wipe data. The Agents SDK and API provide hooks and events for similar guardrails. Developers get callbacks or flags when the agent is about to do something sensitive, and can build in confirmations. The Moderation API and safety guardrails are also available for content filtering. However, these are not automatically enforced in a custom UI — you as the developer must make use of them. OpenAI has put a lot of work into a friendly interface that keeps users in charge (narrating steps, asking permission, etc.). If you’re building your own agent app, you should mirror these trust features. By using the guardrail hooks in the SDK, you can show users what the agent is doing, ask for confirmations on critical steps, and provide easy ways to pause/stop. This will not only match user expectations set by ChatGPT, but also protect you legally and ethically. Safety and transparency should be part of your UX just as they are in ChatGPT’s product. OpenAI’s ChatGPT Agent launch includes quotas: Pro subscribers get 400 agent actions per month, Plus users get 40 per month (with more purchasable via credits). The API’s usage is metered by tokens and processing time, not by a fixed number of actions. There’s no concept of a fixed monthly action count in the API - it’s pay-as-you-go. Developers pay per token and for tool usage seconds, with their own limits set by organization quotas or rate limits. If you build an agent for end-users, you have flexibility in the business model. You could offer unlimited use (within fair limits) if the cost in tokens is manageable, or set your own pricing tiers. The ChatGPT quotas might make some users feel constrained, which could drive them toward specialized third-party solutions. Keep an eye on whether OpenAI aligns API pricing with the ChatGPT-style quotas in the future. No new model family, but major performance jump The big boost does not come from a brand-new model but from smarter coordination — specifically, how the agent chains together browsing, coding, and file tools during execution.  These tools have been available in the API since March, but combining them in a continuous agent loop enables the system to think ahead and reuse context.  This orchestration now leads expert benchmarks and, in many cases, even surpasses human performance on high-value tasks. The takeaway from July’s launch: With the right setup, you can get much more out of existing models. User expectation has jumped Before, an average user might have been delighted if a third-party GPT-based app could, say, give a good answer with references. Now, after seeing ChatGPT Agent, users will expect AI agents to take action and complete tasks autonomously.  Agents are now expected to not just chat, but act. For developers, this means that simply providing an answer isn’t enough to impress — your agent may need to book the restaurant, format the spreadsheet, or orchestrate the workflow just like ChatGPT can.  OpenAI Agents SDK 0.2.x Release On July 17, 2025, OpenAI shipped version 0.2 of its Agents SDK. The release folds three production-grade capabilities into the library we already use to build AI assistants. Guardrails now run alongside every model or tool step. They apply the same policy filters OpenAI uses in ChatGPT, so unsafe content or prompt injection attempts are blocked before they reach users. A persistent session object is included. By default, it stores chat history in a local SQLite file, but the interface lets us plug in Postgres, Redis, or any other store that meets our data governance requirements. With sessions in place, the agent automatically “remembers” prior turns - developers no longer need to stitch conversation history into each prompt ourselves. Tracing records every loop iteration — each model call, tool invocation, and agent handoff — and exposes the log through a simple UI and JSON feed. Developers can replay a conversation after the fact, pinpoint slow or expensive steps, and provide auditors with a full trail of agent decisions. Together, these features replace much of the custom glue code developers maintain today for safety checks, state management, and debugging.  Migrating an existing prototype agent is expected to take one to two developer-days: change the import paths, enable guardrails, choose a session backend, and turn on tracing. Areas that still need solutions, which developers can address Connectors not open to API This gap means a developer can create a custom agent with connectors to niche services or internal systems today, while OpenAI’s own connectors are limited or beta. If you have industry-specific knowledge (say an agent for medical records or for enterprise CRM systems), you can integrate that now. By the time OpenAI releases an official connectors API, you could already have a reputation or customer base for that integration. No built-in scheduling for API If your use-case benefits from periodic or background tasks, you can build that infrastructure and offer something ChatGPT can’t (yet) do outside its interface. It’s more effort, but for certain enterprise clients or power users, that could be a deciding factor. Session length and persistence ChatGPT’s UI keeps the virtual computer session alive 30 minutes, whereas API sessions time out after 5 minutes idle. If long-lived sessions matter (monitoring something continuously or lengthy workflows), developers might soon get tools for it, but if not, you could engineer a workaround (perhaps by intelligently re-initializing context) and advertise that stability. Geographic and compliance constraints The ChatGPT Agent is geofenced out of the EU (EEA) at launch, likely due to regulatory concerns. This means EU-based enterprises or users can’t use ChatGPT Agent yet. If you are a developer or company operating in Europe, this is a moment to pitch custom solutions that run on the API (which can be configured for data residency and compliance). You can provide AI agent services without the geofence, addressing privacy or legal requirements in ways the main product might not initially. Advice for developers Since the core capabilities (browse the web, execute code, handle files, etc.) are becoming commodities available to anyone using OpenAI’s platform, the strategic advice is to differentiate in other ways. Proprietary data or knowledge Incorporate company-specific data or industry expertise that a general ChatGPT agent won’t have by default. For example, a financial agent that knows your accounting system or a medical agent that has access to up-to-date research in a specialized field. User experience & integration Design a user interface or workflow that fits a particular job better than ChatGPT’s general UI. Maybe a browser extension, or integration directly into a team's Slack/SharePoint, or a voice interface for use on the go. Provide a streamlined experience for specific tasks. Trust and custom guardrails As mentioned, use the guardrail primitives to build trustworthy agents. Enterprises might prefer an agent whose every action is logged, auditable, and with admin controls — something you can build with the API on top of OpenAI’s models. Branding and relationships Companies might prefer a branded solution (their own AI assistant) for psychological or business reasons. As a developer, you can offer white-label agents using OpenAI under the hood. You maintain the client relationship and can tailor the solution exactly to their needs, which the general ChatGPT cannot do for each client. Beyond the Hype: The Hidden Costs of "Mostly Correct" AI Automation Today’s AI "agents" promise to handle whole jobs for us, but leaders should view them less as veteran employees and more as unreliable interns: fast, tireless, and eloquent, but prone to costly blunders if left unsupervised. Even a two-percent error rate sounds tiny until you chain dozens of automated steps together. At that scale, the math virtually guarantees something will break. A mistyped year in a financial report, a stray decimal in pricing, or a logic bug in a data pipeline can ripple outward and cost millions—and chasing those hidden flaws often erases any time the AI saved. Security risks compound the problem. Because an agent blindly interprets any text it is fed, a malicious instruction buried inside an email, PDF, or even hidden HTML can hijack the system—what researchers call "prompt injection". If that agent also holds privileged credentials, one invisible line of text can trigger data exfiltration or an unauthorized purchase, turning the model into an unwitting insider threat. Handing high-powered tools to a model that can be tricked so easily is why regulators and risk teams are starting to worry. Regional regulations complicate deployment: Europe’s stricter privacy rules already delay some advanced features, forcing companies to navigate a patchwork of compliance before rolling agents out at scale. Yet squeezing the next few "nines" of accuracy from these systems is ruinously expensive. Improving a model from 98 percent to 99.99 percent reliability can demand orders of magnitude more data, compute, and money—much like the stalled quest for full self-driving cars. Hardware limits, energy costs, and a shortage of fresh, high-quality training data all hint that brute-force scaling is hitting a wall, and future gains will come from new architectures: smarter algorithms, not simply bigger models. History shows businesses tolerating small failure rates when the efficiency gain is significant, and AI will likely follow that pattern. Routine, rules-based office functions are already being automated, shifting human roles toward exception handling and firefighting. Far from ushering in leisure, faster tools usually raise management expectations. Employees often find their workload grows as they supervise the AI in addition to their original duties. Oversight remains the best defense. The most successful teams break work into micro-steps, subject every result to automated tests or human review, and immediately pull the agent off the task if something looks odd. Treating the model like a junior colleague who drafts at lightning speed but needs line-by-line checks prevents "tech debt" from snowballing. Without that discipline, whatever time you saved on first drafts you will pay back—often with interest—during debugging and clarification. The safest technical blueprint emerging today isolates privileges: one sub-agent can read untrusted text but has no authority to act, another holds the keys to sensitive tools but never sees raw external data. "Sandwiching" these models behind strict guardrails, capping their spending power with prepaid cards, and inserting human approval for anything high-impact are becoming basic.
Bloomberg's AI is Live. It Just Isn't BloombergGPT
Bloomberg's AI is Live. It Just Isn't BloombergGPT
BloombergGPT Is in Production or Not? In March 2023, Bloomberg published a paper about BloombergGPT, "a 50 billion parameter language model that is trained on a wide range of financial data". The training set was "a 363 billion token dataset based on Bloomberg's extensive data sources, perhaps the largest domain-specific dataset yet, augmented with 345 billion tokens from general purpose datasets". Bloomberg claimed that BloombergGPT "outperforms existing models on financial tasks by significant margins without sacrificing performance on general LLM benchmarks". After the paper, Bloomberg gave no official updates about BloombergGPT. You could not find a GitHub repo or a Hugging Face model card with a downloadable checkpoint, no public demo was available, and the price was not announced. However, only two months after Bloomberg's paper, a general model available through an API beat BloombergGPT on most of the tested financial tasks. Researchers from Queen's University and J.P. Morgan AI Research tested ChatGPT against BloombergGPT on financial tasks. According to their findings, ChatGPT performed "better on financial NLP tasks than BloombergGPT, which was specifically trained on financial corpora". On question answering, it was ahead of BloombergGPT by more than 15 percentage points. Despite this criticism, Bloomberg kept silent. Nine months later, Doug Levin, founder and first CEO of Black Duck Software and now an Executive Fellow at Harvard Business School, wrote a review in February 2024: "I recently had the opportunity to use BloombergGPT on the Bloomberg Professional service. The way I understand it, BloombergGPT was developed by Bloomberg's ML Product and Research group collaborating with the firm's AI Engineering and data teams." His experience was very positive: the AI was able to quickly generate high-quality financial reports, market summaries and trading ideas or analysis that incorporated real-time data from the platform. He wrote that "the potential for augmented intelligence in finance is clear", and that "BloombergGPT has disrupted the Bloomberg terminal with a new technology that is likely to replace the established one". Directly mentioned use cases from Doug Levin's article: Financial report generation Market summaries Trading ideas or analysis Financial data analysis Market trend predictions Sentiment analysis Automated report generation Language translation Financial document text generation Risk assessment Real-time market updates Support in client communication Support in regulatory compliance S-1 analysis and modeling Search functionality via Bloomberg Search (SEAR) It seemed that, despite the skepticism of the researchers, Bloomberg had rolled out its model in its products and the clients were very happy. Bloomberg had a custom model, and this was now the only way to build a financial LLM that stands out. This opinion held in the community until 2026. However, in April 2026, Wayne Barlow, Global Head of Terminal Products at Bloomberg, told A-Team Insight that "BloombergGPT was a research model and is actually not used at all in any of our products". In fact, their products "are built on a combination of models". They have "some commercial models, some open-weight LLMs, some smaller, custom language models that we build ourselves. Because we have an agentic system, different layers may use different models, each solving specific needs. Some of it relates to latency and speed, in which case you may need a smaller model. In other cases, it's simply which model performs best for a specific task." Shawn Edwards, Bloomberg's CTO, confirmed this. He told Fortune that the system uses "a variety of AI models, including some built by Bloomberg itself and others from frontier AI model companies such as Anthropic", routing queries "to the cheapest model that can handle a given task". So Bloomberg trained its own large financial LLM (around 1.3 million GPU hours, a core team of about nine people), but this particular model never went into a product. They learned how to build one mega custom model, and then applied this knowledge to build small custom models for specific tasks, and only for the tasks where the ready-made LLM models did not have enough capabilities. Where a big commercial model did a task at a higher cost than a cheaper model, the task went to the cheaper model. In fact, they built the conversational AI interface ASKB, which uses different models depending on the task, and this matches the way financial AI products are built now. BloombergGPT-like LLM: Train from Scratch or Use the APIs of Public Commercial Models? In July 2023, four months after the paper, David Rosenberg, who leads the Machine Learning Strategy team in the Office of the CTO at Bloomberg and is a co-author of BloombergGPT, told The TWIML AI Podcast with Sam Charrington how the model was built. This is the only detailed description of the custom LLM development process from Bloomberg. Then Rosenberg explained why Bloomberg trained its own model this way. "Using an API like OpenAI’s is not suitable for us: we have data we don’t want to send out. So for internal and sensitive use, in-house models are preferable." — David Rosenberg, The TWIML AI Podcast It seems that in 2026 Bloomberg's position was clarified and adjusted, and in some cases they are no longer against using external LLM APIs and do not worry about security. Nevertheless, it looks like the most critical data still stays inside anyway, and it is processed by custom self-hosted models based on what the team learned from BloombergGPT. The results of BloombergGPT are outdated, but the cutom development process is not, and it is still useful for any company that plans its own financial LLM. Backstage of Bloomberg's Custom Financial LLM Development Decision Behind Creating a Custom Financial LLM BloombergGPT is an example of a project where a team inside an enterprise trained and built a custom large language model specializing in financial language. The enterprise made a strategic decision to invest money, time, and human resources into this machine learning effort. In 2026, as Barlow said, the model itself is not used in any product, and what they got from this investment is the experience of training a large model, the data pipeline and the team, which they now use for small custom models. "It was clearly a big investment. We decided it was worth to start building our own model. I'm not sure we knew exactly at that time what it would be used for." — — David Rosenberg, The TWIML AI Podcast Financial LLM use cases What the BloombergGPT development team actually spent the most time on was thinking about financial LLM use cases from a variety of angles: could the model help them solve problems they already had solutions for, but in a better way, or with less investment in training data? They explored several use cases, including the following ones: Natural language to BQL. Bloomberg Query Language is used inside Bloomberg Terminal to pull structured financial data, and the idea was to build a kind of financial code assistant that translates human language into Bloomberg-specific queries. Interaction with a large document. The user inputs a document and asks what information it contains. In fact, these particular two use cases became products much later. In April 2025, Bloomberg launched Document Insights, where users "ask questions in everyday conversational language" about a company document, such as an earnings call transcript, and get "transparency links to highlight relevant excerpts in the original document". In February 2026, their financial AI product ASKB started to generate BQL. According to The TRADE, "where queries involve data analysis, the system generates the associated Bloomberg Query Language code, allowing traders to extend outputs directly into Excel, BQuant Desktop or BQuant Enterprise for further modelling and workflow integration". LLM Hallucination problem "As for production use, we need to be very cautious. No one has really solved the hallucination problem yet. These language models can say wrong things, do strange things, so there needs to be a process around making them safe to use, either internally or, eventually, with clients." — David Rosenberg, The TWIML AI Podcast Moreover, if an LLM product seems reliable at first, the users who rely on it for everyday internal tasks may become less critical of its output in the future, and if the LLM eventually returns something wrong to the people who trust it, this can end in disaster. So the hallucination problem has to be solved at the system level. However, hallucinations were the second priority at the start of the custom financial LLM development. They started with internal use, paying less attention to safety or reputation and more to the development and testing of the functionality. The main question was: did it do the job? That was their focus at the time. Nevertheless, the question of solving hallucinations was still there. The simple solution seemed to be that the user just checks the output for hallucinations. But after this came another strategic question: how much of the responsibility for checking should be moved to the user, and will it turn out in the end that the users spend as much time or more on checking, and there is no efficiency gain that is expected from using an LLM? At that time, they had no answer to this question, so they kept things internal. However, by 2026 it became known how Bloomberg solved the hallucination problem and built "a useful, trustworthy system" (Shawn Edwards, Bloomberg's CTO). Evaluations at the product development level. As Edwards told Fortune, using AI models as evaluators "can work for easy cases", and "for everything else, human assessors are required". During development, Bloomberg pulled bond covenant experts, equity analysts, market structure specialists and journalists off their day jobs both to write benchmarks for sub-agents and to help evaluate entire workflows". Checking at the user level, which is still cheaper fro business users than collecting the information yourself. According to Bloomberg's press release, every ASKB answer comes with "transparent attribution to original sources and the underlying Bloomberg Query Language code for any data analysis".  Training Dataset for Bloomberg AI The training dataset was a mix of standard general-purpose data and Bloomberg's proprietary financial data. About half of the dataset (51.27%) came from Bloomberg's curated collection called FinPile, built over many years starting in 2007. Most of FinPile (298 of 363 billion tokens) was financial content scraped from the web. The rest was news articles, filings, press releases, and Bloomberg's own content. There was a problem with the tokenization of numerical data, and they solved it. In 2026, this problem is solved at the tokenizer level in most open-weight models, so a company that fine-tunes a ready-made model does not need to do this work again. The BloombergGPT model did not go into a product, but the data did. In 2026, ASKB draws on "hundreds of millions of company documents like earnings call transcripts, presentations and filings", and Shawn Edwards told Fortune that the data is the hard part: "You have to buy all those sources, do all the validation work, build benchmarks, and tokens aren't cheap." Custom BloombergGPT Training Process At that time, Meta had just released the OPT model together with its training logs, and Hugging Face had published the detailed logs of BLOOM. The team used these logs as a roadmap. "We copied the BLOOM model architecture fairly closely, with some small tweaks. "— David Rosenberg, The TWIML AI Podcast The training took three attempts: v0, curriculum learning. FinPile data was sorted chronologically by month, because the newer data was more similar to the future documents, and the validation set was the month after the training set. After about a week, the model was stuck on both training and validation loss, and they stopped this run, because they "deemed this to be too risky to catch any other potential problems with the training that might require early intervention". They restarted with fully shuffled data. v1, gradient norm growth. After about 4.5 days (12k steps) the gradient norm started to grow, and the validation loss jumped. They made four attempts to fix it (a rollback to an earlier checkpoint, reshuffled data, a lower learning rate, gradient clipping, FP32 in the LM head), and none of them worked. Then they found that out of 70 layers, the layer norm weights of the first layer shrank until steps 11-12k and then started to grow. During this investigation, they also found a bug: weight decay was applied to the layer norm weights that should stay around one. v2, the final run. They started from scratch with conservative hyperparameters (a lower learning rate, gradient clipping 0.3, a longer warm-up, an extra layer norm at the embedding layer, no weight decay on layer norm weights, fully shuffled data) and additional monitoring of the first layer. The training was "very smooth and uneventful" for about 42 days. Around day 48, the validation loss stopped improving for a week, they tried six more runs with a lower learning rate, a rollback and reshuffled data, and stopped training after 53 days at about 80% of the dataset, because "the loss on our held-out development set was no longer improving". Resources Used to Train BloombergGPT The core team was about nine people (four on implementation, three on ML and data, one on optimization and compute, and the rest on evaluation, literature review and support). The compute was 512 40GB A100 GPUs on Amazon SageMaker (64 p4d.24xlarge instances with 8 GPUs each) with the SageMaker Model Parallelism library. Bloomberg pre-purchased around 1.3 million GPU hours at a negotiated rate. They took this compute budget and calculated from it, by the Chinchilla scaling laws, how big a model (50 billion parameters) and how much training data (569 billion tokens) fit into 1.3 million GPU hours. So the whole project was about $1.3 million at the low end of cloud pricing ($1 per hour for reserved capacity). Custom Financil LLM Validation and Performance Evaluation During training, they used the month after the training data for validation. Later, they added a second validation set from the last 105 million tokens of the shuffled training data. They also monitored downstream tasks like MMLU and BBH (Big Bench Hard). After training, they did a full evaluation and compared BloombergGPT to other models. On financial tasks, it outperformed the open models of that time. They also had internal benchmarks: Sentiment analysis on financial news and social media Named entity disambiguation (linking "Apple" to its stock ticker, etc.) Natural language to BQL, which is like SQL for Bloomberg Terminal Even though the model wasn't trained on BQL directly, it performed well in few-shot scenarios. They also experimented with headline generation and other generative tasks. In 2026, the evaluation is the main part of the work. The benchmarks are now written by domain experts, and not only by the ML team. What’s Next for the BloombergGPT Team? Already in 2023, with a working large model, they understood that they needed to move to small-scale experiments and test tokenization strategies, data mixtures and architecture choices in a more disciplined way. "We're more interested now in smaller models. They're easier to use: you can run inference on a single GPU. Our 50B model requires multi-GPU infrastructure. Inspired by the LLaMA paper, we're exploring what we can achieve with smaller models, longer training, and careful design. We want both small and large models for practical use."— David Rosenberg, The TWIML AI Podcast They managed to move in this direction and stick with it. In 2026, Bloomberg's products run on "smaller, custom language models that we build ourselves" together with commercial and open-weight models, as Barlow said. The 50B model never went into a product. They also continued the small-scale experiments. On July 2, 2026, Bloomberg's AI researchers presented 8 papers at ACL 2026, including "a new training recipe that enables smaller models to match or outperform larger specialized guardrails on unseen policies" and "a novel approach to complex text-to-SQL". Your own Financial LLM: A Mirage? Financial LLMs are not a fantasy. There are already startups in this space that have gone to production, raised funding, and landed their first clients. The most well-known are those backed by YCombinator, for example, Truewind ($17 million raised, Series A in January 2025, over 100 customers in 2026). However, Truewind did not train its own model and uses models from OpenAI, as well as JPMorgan's LLM Suite, launched in 2023, which uses models from OpenAI and Anthropic. Goldman Sachs built its GS AI Assistant so that it can interact with different LLMs from OpenAI, Google and Anthropic, as well as open source models. The model vendors came to finance themselves. On May 5, 2026, Anthropic released "ten ready-to-run agent templates for the most time-consuming work in financial services". Today, Claude connects to dozens of market data, research platforms, and financial companies' internal systems, including FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, Chronograph, LSEG, and Daloopa. The financial models became small. Fin-R1 (latest version March 2026) has "a compact size of 7 billion parameters" and, despite it, achieves competitive performance across financial benchmarks. The mirage is one big custom financial LLM. Bloomberg trained one, and in 2026 this model is not used in any of its products. LLMs also have core limitations by design, and you can't expect a financial LLM to do what it simply wasn't built to do. So, one big custom financial LLM is a mirage. Small custom models for specific tasks, together with pre-built commercial models, are not. How Belitsoft Can Help Building a financial LLM is not a technical challenge. It's a product challenge. Belitsoft helps financial firms build AI products on a combination of models, as Bloomberg did with ASKB, where each task goes to the model that handles it at the lowest cost. This means: Training small custom models for specific tasks, where ready-made models do not have enough capabilities Fine-tuning existing open-weight models with in-house financial content Connecting commercial models (OpenAI, Anthropic, Google) for the tasks where they are cheaper or better Routing tasks between these models depending on latency and cost Evaluations at the product development level by your domain experts Checking at the user level: every answer comes with its sources and the underlying query A conversational AI interface for your users or an integration into your dashboards and query layers You get your own financial AI product: an internal assistant, a decision tool or a document engine. The most critical data stays inside and is processed by custom self-hosted models. The other tasks go to pre-built commercial models through their APIs. Frequently Asked Questions

Our Clients' Feedback

Any software development, programming, or design needs we have had, Belitsoft company has always been able to handle this for us

We have been working for over 10 years and they have become our long-term technology partner. Any software development, programming, or design needs we have had, Belitsoft company has always been able to handle this for us.

We are very happy with Belitsoft

Belitsoft has been the driving force behind several of our software development projects within the last few years. This company demonstrates high professionalism in their work approach. They have continuously proved to be ready to go the extra mile. We are very happy with Belitsoft, and in a position to strongly recommend them for software development and support as a most reliable and fully transparent partner focused on long term business relationships.

We can recommend Belitsoft and are looking forward to start a follow-up project

It was a great pleasure working with Belitsoft. Software Development Company. New requirements and adjustments were implemented fast and precisely. We can recommend Belitsoft and are looking forward to start a follow-up project.

We are definitely looking at Belitsoft as a long-term partner

They use their knowledge and skills to program the product, and then completed a series of quality assurance tests. We were working in an agile way with them. Belitsoft performed very well throughout our project. We are definitely looking at Belitsoft as a long-term partner.

I highly recommend Belitsoft for website design and development

I highly recommend Belitsoft for website design and development. We were up against a tight deadline to launch the project. The work was delivered on time and within budget! I will continue working with Belitsoft as a valued partner for our web development!

We would recommend them to anyone who is in need of custom programming work

We have worked with Belitsoft team over the past few years on projects involving much customized programming work. They are knowledgeable and are able to complete tasks on schedule, meeting our technical requirements. We would recommend them to anyone who is in need of custom programming work.

Belitsoft seems to know what they're doing, which I appreciate

Belitsoft company is able to make changes instantly. One of our internal engineers has commented about how clean their code is. Belitsoft seems to know what they're doing, which I appreciate.

We continue today to augment our existing staff with great developers from Belitsoft

Belitsoft company has been able to provide senior developers with the skills to support back end, native mobile and web applications. We continue today to augment our existing staff with great developers from Belitsoft.

We highly recommend that you use this company if you want the same benefits

Belitsoft company delivered dedicated development team for our products and technical specialists for our clients time to time custom development needs. We highly recommend that you use this company if you want the same benefits.

They were able to convert it into a multi-platform software solution

We approached BelITsoft with a concept, and they were able to convert it into a multi-platform software solution. Their team members are skilled, agile and attached to their work, all of which paid dividends as our software grew in complexity.

Belitsoft guarantees first-class service through efficient management, great expertise, and a systematic approach to business

Having worked with Belitsoft as a service provider, I must say that I'm very pleased with the company's policy. Belitsoft guarantees first-class service through efficient management, great expertise, and a systematic approach to business. I would strongly recommend Belitsoft's services to anyone wanting to get the right IT products in the right place at the right time.

If you are looking for a true partnership Belitsoft company might be the best choice for you

If you are looking for a true partnership Belitsoft company might be the best choice for you. They have proven to be most reliable, polite and professional. The team managed to adapt to changing requirements and to provide me with best solutions. I strongly recommend Belisoft.

I expected and demanded a lot of you at Belitsoft company, but you exceeded my expectations

I expected and demanded a lot of you at Belitsoft company, but you exceeded my expectations. You acted pro-actively, challenged me at the right moments. Thanks!

Let's Talk Business
Do you have a software development project to implement? We have people to work on it. We will be glad to answer all your questions as well as estimate any project of yours. Use the form below to describe the project and we will get in touch with you within 1 business day.
Contact form
We will process your personal data as described in the privacy notice
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply
Contact us

USA +1 (917) 410-57-57
700 N Fairfax St Ste 614, Alexandria, VA, 22314 - 2040, United States

UK +44 (20) 3318-18-53
26/28 Hammersmith Grove, London W6 7HA

Poland +48 222 922 436
Warsaw, Poland, st. Elektoralna 13/103

Email us

[email protected]