Evaluating AI deliverables you cannot fully verify

Ask for the test set. Any serious specialist can show you the inputs they used and what the system produced for each one.

Spot check the failure cases, not the highlights. The demo always works; the edge cases tell you whether the work is real.

Request a short handover document: how it runs, what breaks it and what to do when it does.

If the deliverable is content, run a plagiarism and factual check before acceptance. It takes minutes and is the most common reason for revisions.