AI That Survives Contact With Production

Every AI vendor has a demo that looks impressive in a sales call. Far fewer have a system that survives real users, real data, and a compliance review. The gap between the two isn't a bigger model — it's engineering discipline that most AI vendors skip because it doesn't show up in a demo.

Why demos break in production

A demo only has to work once, in front of an audience that isn't trying to break it. Production is different: the model gets asked something it should refuse to guess at, a document from one customer ends up next to another's, and eventually someone with compliance authority asks how a decision was actually made. Most AI builds have no answer to any of that, because nobody built one.

Grounding: refusing to invent numbers

The fix isn't a better prompt — it's a guard between the model and the user. Any number, percentage, or stated fact in the output is checked against what the underlying system actually produced; anything that doesn't match is rejected and replaced with deterministic fallback copy instead of a guess. The check that verifies this runs in CI against the same primitives as the runtime guard, so the two can't quietly drift apart.

Isolation: proving one tenant can't see another's data

"We tested it and it works" isn't proof; a UI check that hides another tenant's data is not the same as data that can't be reached. Real isolation lives at the database layer — Postgres row-level security enforced through a per-transaction session variable, with a bleed-guard so an unset variable reads back as NULL and matches no rows: the failure mode is an empty result, never someone else's data. The application role can't bypass that policy and doesn't own the tables. Every isolation test carries a positive control, so an empty result can't quietly pass as a working query that happens to return nothing.

Why "it worked in the demo" isn't proof

The same discipline extends past grounding and isolation. When the system doesn't know something, it should say so — an honest "we don't know," not a confident guess, and estimates that don't move on data the system was never given. And every departure from the original plan gets written down: a numbered ADR citing the exact clause it deviates from and who approved it, so "why does it work this way" has an answer six months later.

How to ship AI that survives contact with production

Start with the one or two use cases that have a real, nameable payoff — not AI for its own sake. Prototype quickly so you're reacting to something concrete instead of a spec. Build the smallest version that actually delivers the guarantees above, not the biggest version that fits the budget. Two senior architects use AI leverage to move at this pace, with a security review across every surface before handover and the ADR trail behind every decision — which is what makes the speed something you can audit rather than something you have to take on faith.