Production AI

Most enterprise AI does not fail at the model. It fails in the distance between a prototype that impresses a room and a system an organization can depend on. That distance is engineering: the data path, the permissions, the failure handling, the observability, and the ordinary software that puts a model in front of the people doing the work.

We write about that gap because it is where our engagements live. A pilot that never reaches production is not a cheap experiment. It spends the sponsor's credibility, and it makes the next proposal harder to get signed.

The failure has a pattern. A pilot is scoped as an experiment, so it is built like one: data pulled by hand, credentials borrowed, no owner for the systems it touches, correctness judged by whoever is watching the screen. It works, because the person demonstrating it knows which inputs to use. Then the request arrives to run it for real, and every question that was deferred arrives at once. Which system of record it writes to. Who is accountable when it is wrong. What happens when the model provider changes something underneath it. None of those are model problems, and none of them are solved by a better prompt.

We take the opposite approach and build the production version of the problem from the outset, in a deliberately narrow slice. The system runs inside your own cloud accounts and identity systems wherever the architecture allows, against real data with real permissions applied, and it is judged by an evaluation suite written in the language of the business rather than by a demonstration. The slice is small because the point is not to show breadth. The point is to prove that one path through the system can be trusted end to end, and to surface the integration and access problems while they can still be designed around rather than after the wider rollout has been committed to.

That is why production readiness is a design question here, not a hardening exercise once the demonstration has been approved. Observability, failure handling, rollback, permissions and the runbook belong in the first architecture. It makes the early work look less impressive than a prototype would, and it is the only version of the work that survives contact with an organization. A system your team can operate, change and defend to an auditor is the deliverable. Anything short of that is a demonstration with a maintenance obligation attached.