Production AI Needs an Evidence Chain

AI governance usually shows up after the model has already answered and the screen looks polished.

A reviewer gets an approval button and a recommendation, with little visibility into how the system arrived at them.

In production AI, the decision record matters as much as the answer. It should show which source records were used, which data or retrieval version was in effect, which prompt and model handled the case, which policy made automation eligible, and which threshold routed it to approval, rejection, or a queue.

That sounds like implementation plumbing until something goes wrong. Then it becomes the difference between a reviewable decision and a confident guess with a timestamp.

This is where AI controls can get thin. They record the outcome, but not the evidence that travels with it: the inputs, policy path, override history, or evaluation results needed to challenge the recommendation.

An approval like that does not reduce risk. It just gives the guess an owner.

The control point is the review record: what the model used, why the case was routed, what the reviewer changed, and whether that feedback shaped the next evaluation.

Human review becomes governance only when the person can see what the system knew, what it assumed, and what would have happened without them.

Originally published on Substack.