What can you actually prove about your AI?

Finance is now being asked to stand behind numbers that an AI model touched. Not produced entirely — touched. A forecast the model adjusted, a classification it made, an accrual it suggested. Somewhere downstream, someone signs.

So ask the question an auditor will eventually ask: Can you show me what happened on the eleventh of March?

What most organizations can produce is a system log. Latency, token counts, error rates, maybe a truncated prompt. That data was built for an engineer debugging a problem at two in the morning, and it is very good at that. But it was not built for attestation, and it does not become evidence because you need it to be.

Evidence should be captured at the moment of inference — the AI’s momentary response — because it cannot be reconstructed later without losing audit credibility. At a minimum: which model and version answered, which prompt and policy version governed it, what inputs it saw, who or what authorized the call, what came back, and a tamper-evident binding of those facts to each other and to the clock. That is what an effective control includes.

The reason this is missing is in the design, not a technical constraint. Engineering logs are built for debugging. Business risk flows through that architecture with controls applied as band-aids. Finance inherits the attestation problem downstream, when the only remaining move is to write a memo describing controls that were never built.

Here is the test I would apply. Pick one inference from ninety days ago. Can you produce a record a third party can verify without taking your word for it?

If not, you do not have controls. You have logs, and a memo.