Before it touches real work
Deployments fail for organisational reasons far more than technical ones. This is the list worth working through before go-live, in roughly the order the omissions hurt.
1–3. Scope and authority
One workflow named narrowly, the exclusions written down, and the decisions it may make alone separated from those it must route. That is the job description, and everything else derives from it.
4–5. Knowledge and contradictions
The thirty questions it must answer have current written answers, and conflicts between documents are resolved. Contradictions produce confident wrong answers, which are worse than gaps.
Almost every incident traces back to something on this list that felt optional at the time.
6–7. Permissions and prohibitions
Its own account with scoped access, and the list of things it must never say encoded as tests rather than hoped for.
8–9. Escalation and disclosure
A working route to a named human during defined hours, triggered on stakes rather than confidence — and a plain statement to customers of what they are talking to.
10–11. Evaluation and review
Twenty real inputs scored before launch, and a scheduled sampling routine afterwards with the rate written down.
12. An owner, by name
One person responsible for it being right. If that question cannot be answered on the day you launch, it will not be answered later either — and an unowned system decays.