The AI pilot-to-production checklist

Every production-readiness checklist in circulation tests accuracy, latency, hallucination rate and security posture. Those matter, and they are almost never the reason a pilot dies. The pilot dies because the organisation could not absorb it — so the gates below test that instead.

Gate 1 · A named workflow with a P&L owner

Pass condition: one person can be named who reports a number that changes if this works. Remedy on failure: stop and find the workflow. Nothing downstream is worth doing until this exists, because without it success is unattributable and therefore unfundable.

Gate 2 · A baseline measured before the pilot

Pass condition: you can state today's cost, time or error rate for the workflow. Remedy: spend a week measuring before you build anything. Teams that skip this can never prove the improvement, which is how a working deployment loses its budget.

Gate 3 · The eventual operators were present during the pilot

Pass condition: the people who will use it daily were in the room, not represented by a manager. Remedy: rerun the pilot with them. A tool designed around a proxy for the user gets worked around by the actual user.

Gate 4 · A human gate on the failure modes

Pass condition: you can name what happens when the output is wrong, who catches it, and what it costs. Remedy: instrument the failure path before scaling the success path.

Gate 5 · Governance sign-off before scale

Pass condition: legal and risk have reviewed at the tier the workflow actually sits in, not at the tier of the pilot's twelve users. Remedy: bring them in now — the review takes weeks when it precedes scale and quarters when it follows it.

Gate 6 · A handover date with a named owner

Pass condition: a date exists after which your team runs this without the vendor. Remedy: contract it. If handover is not an event with a name and a date, it does not happen.

Six gates, and only one of them is about the software.

Where partner selection sits

Between gates 1 and 2. Naming the workflow tells you which partner is even relevant; measuring the baseline is something the right partner will insist on before building. If you are choosing a supplier before gate 1, you are shopping without a specification.