The AI Output Verification Checklist

Twenty-two checks to run before AI-generated work reaches a client, a reviewer, or the ledger.

A language model writes an answer it is unsure about in exactly the same tone as one it is certain of. Every habit you have for spotting a colleague’s uncertainty fails here, which is why these checks are deliberately mechanical — counts, totals, traces. Mechanical checks are the only kind that catch a confident error.

Covers scoping before you prompt, the arithmetic, the claims, the tone going out, and the paper trail you will want in six months.