Last month we sat in a review where everyone liked the answers. The model was fluent. The retrieval looked tidy. The product owner asked, reasonably, when it could go live. Nobody in the room could name the thing that would stop it.

That silence is expensive. Not because the feature was careless — it was not — but because “it looked good” is not a criterion. Demos are designed to look good. They do not tell you what happens when the index is a week stale, a tool call times out, or last Thursday’s prompt tweak changes tone for a regulated customer.

Write the no before the yes

A release gate for an AI system is a written agreement. What “good” means. On which examples. Which failure modes are blockers, and who is allowed to hold the line. Until that exists, every review is taste. Taste does not survive a model upgrade, a new corpus, or a Friday afternoon.

• Twenty examples that must pass. Five that must fail. Named, versioned, and boring.
• A score next to latency — not in a slide, in the same place the pipeline already looks.
• A person who can refuse the release. If that seat is empty, the gate is decoration.

“

A pipeline that only knows how to say yes is a launch button with extra steps.

The move is small on purpose. Pick one journey that matters commercially. Score it the same way every time a prompt, a model, or an index changes. Put the number where shipping decisions are made. That is not research. That is how a commercial system stays honest.

If you want help drawing the gate, that is the work we do. Request a call and we will talk about the journey, not the demo.