What an AI Evidence Trail Has to Look Like Before an Auditor Sees It

There is no AI clause in SQF Edition 10 or in FSMA. There does not need to be. The existing rules already decide the question: records must be accurate, attributable, and verified by someone competent. If AI drafted the record, those requirements land on the evidence around the draft. We run AI inside a certified facility, so we had to answer this before it was a blog topic.

The four things the trail must hold

• What the model saw: the records, documents, and requirements in context when it drafted

• What it produced: the draft as generated, before any human touched it

• What the human did: the diff between draft and approved version, and who approved

• When and under what version: model, prompt, and configuration, timestamped

The draft is not the record

The load-bearing design decision: the AI output is an input to a human decision, and the record the program stands on is the approved version with a named approver. The trail exists so that the approval is inspectable, not so the machine can take responsibility. Machines cannot hold responsibility in a food safety program; name-and-date can.

Append-only or it does not count

An evidence trail someone can edit after the fact is a liability with extra steps. Ours is append-only: corrections are new entries pointing at what they correct, the way a good lab notebook works. When we handed the structure to an auditor familiar with electronic records, their questions took about ten minutes, because every answer was a lookup, not an explanation.

What we would tell a facility evaluating AI tools

Ask one question first: show me the record of what the model saw, what it wrote, and who approved it. If the vendor cannot produce that in under a minute, the feature is a demo, not a program tool. Everything else, accuracy included, is negotiable by comparison, because a wrong draft with a real review trail is a caught error, and a right draft with no trail is an unverifiable record.