The four decisions that move AI from pilot to production.

Most enterprise AI programmes don't fail in the model. They fail in the four architectural decisions made before the first prompt is ever written. Here is what separates the 5% that reach production from the 95% that stall.

1 · Decide the platform before the use case

The instinct is to start with a flashy use case and find a platform that fits it. That produces a vendor zoo: one stack per pilot, none production-grade, none reusable. We invert it — one reference architecture on your cloud, governed end to end, that every subsequent use case rides on. The second use case ships in a third of the time of the first because the platform already exists.

2 · Design governance first, model second

A model with no eval framework, no risk register and no human-in-the-loop policy is a liability the moment it touches a customer. We design the governance — risk tiers aligned to the EU AI Act, golden datasets, automated eval gates — before selecting the model. Counter-intuitively, this is what makes shipping faster: the DPO and the platform engineers build one runbook, not two adversarial ones.

3 · Make data flows legible

If your security team cannot draw the data-flow diagram from memory, the system will not pass review — and it shouldn't. Every movement of data should be an explicit, documented contract: no silent multi-provider relays, no shadow copies, full lineage on every decision. Legibility is not bureaucracy; it is the precondition for trust at enterprise scale.

4 · Productionize behind real users in weeks, not quarters

The gap between a pilot and a platform is operational, not algorithmic. CI/CD for prompts and models, observability, rollback, on-call. We ship the first system into production behind real users in eleven weeks — then hand the runbook to your team. The measure of success is not a demo that impresses; it is a system your own engineers operate on Monday morning.

The pattern

Platform before use case. Governance before model. Legibility before scale. Production before polish. None of it is exotic. All of it is the discipline that turns AI ambition into a number on the P&L — which is the only place it counts.