To create AI software, you need developers who can build an app or website around the AI and run it in production, prepare company data so the AI can use it, and write instructions the AI follows. If you augment AI developers, the cost depends on the hourly rate and the time your project takes to complete. If you need to train your own model, it's required to use specialized servers that are also rented by the hour, and how many hours are needed depends on your project.
You do not usually build AI entirely from scratch. You buy AI models from OpenAI, Google, or Anthropic. You pay per token you send to and receive from the AI, where each 1M tokens has its own price depending on the AI provider. As you get more users, this API bill scales with your traffic.
You also have to pay for moving your AI data out of your cloud provider to your users over the internet. This costs cents per gigabyte of outbound traffic, but if you stream large files or audio, this is a major recurring cost.
You must also budget for proving your AI is not hallucinating or violating GDPR or HIPAA, protecting your software against prompt injection attacks (where hackers trick your AI into giving up private data), and paying for expensive software to manage the infrastructure (like Datadog for monitoring or Snowflake for databases).