Who owns the cost of AI errors, IT or operations?

A lot of AI in payments and finance operations runs on whatever model the firm's cloud provider offers. For a firm on Microsoft Azure, that used to mean OpenAI.

Since June, Claude has been generally available in Microsoft Foundry and hosted on Azure, next to models from Mistral, Meta, DeepSeek and xAI. AWS Bedrock and Google Vertex offer a similar spread. Hence a firm can now pick the model for each task, or each step within a task, without leaving its cloud provider or signing a new contract.

Architects pick models on workload fit, cost, latency and compliance. The cost they look at is usually the token price. The cost of errors, the human checking and the fixing all land in the operations or finance budget, which is where the business case gets judged.

The usual way to optimise is to give the easy steps to a cheap model and keep an expensive model for the steps with messy information or a judgement call. Go too cheap, though, and the money saved on tokens gets eaten by error costs and extra checking, which can sink the business case.

I think testing is the only way to find the right balance, first before you build and then in regular reviews once the use case is live. My own tests so far show that on some payments and finance tasks, a frontier model such as Claude Opus with thinking set low still beats cheaper models running on high reasoning settings. Once the reasoning tokens are counted, the total token cost ends up comparable. The unit economics article shows how the checking and error costs add up per case.

Who owns the cost of AI errors in your firm, IT or operations? I'd be interested to hear how others split it. Let me know on LinkedIn or by email.