Most stalled AI projects didn't fail on the model. They stalled at a sign-off: security couldn't approve the access, leadership couldn't see evidence it worked, or nobody owned it after the pilot team moved on.
The production checklist
- Its own identity and least-privilege access.
- Connected to the systems people actually work in, not a separate tab.
- An evaluation set from your data, with a baseline and a pass mark.
- Review on every output that leaves the building: a person or a calibrated LLM judge.
- Tracing, cost limits and alerts.
- A runbook, a rollback path and a named owner.
Sequence matters
Do identity and access first: they take longest to approve. Build evaluation while that review runs. Integrate last, once you know the agent's output is good enough to put in front of people.