Why AI pilots stall, and what production-grade actually means.
Most pilots do not fail on model quality. They stall on the questions nobody designed for: who approved this, what did it see, and who owns it when it breaks.
The pattern
The demo lands. Leadership is excited. Then security asks what data the system saw, legal asks who approved the output, and operations asks who is on call when it misbehaves. The pilot does not fail; it simply never gets permission to matter.
The three questions that stall pilots
Every stalled pilot we have been called into was stuck on some version of the same three questions.
- Permissions: can the system prove a user never sees data they could not open themselves?
- Provenance: can any output be traced to the sources and approvals behind it?
- Ownership: when quality drifts at 2am, whose pager rings, and what runbook do they open?
What production-grade means
Production-grade is not a bigger model. It is permission-aware retrieval, guardrails enforced in the workflow, evaluation running on live traffic, and a team that owns the system with runbooks they wrote. Design for those from day one and the pilot never stalls, because the questions are answered before they are asked.
The operator's test
Ask one question of any AI initiative: if this worked perfectly, would we be allowed to use it? If the answer is not an immediate yes, the work is not the model. It is everything around it, and that is exactly the work we do.