What does it take to make an AI system auditable?
Record enough per request to replay it: the input, the retrieved context, every tool call and its result, the model and prompt version, and the final output. If you cannot reconstruct why the system said what it said on a given day, you do not have an auditable system regardless of what the policy document says.
Version the prompts in source control and stamp the version into every trace. Prompt changes are behaviour changes, and an audit trail that omits them cannot explain a behavioural difference between two dates.
Retention is a design decision with legal weight. Traces contain the retrieved enterprise data, which means the trace store inherits the sensitivity of everything the system can read — and needs the same access controls.
Build the replay path early. A trace format nobody has ever replayed usually turns out to be missing a field at exactly the moment it matters.
Written by Binary AI Labs · Reviewed