Perspectives

Most AI never reaches production.

21 November 2026 · Neurai

The model is rarely the reason. What is missing is an account of how the organisation works, and people who stay until it does.

A pilot answers a question in a meeting room. Production answers it at four on a Friday, when the order is late, the record is wrong and nobody who built the pilot is in the building. Most AI work is judged in the first setting and fails in the second.

A demonstration knows the document. Production needs the business.

A demonstration works on the material it was shown. A business runs on relationships that no single document states: which order waits on which delivery, who may approve a credit, which of two stock figures is the one to trust.

A system that has not been told these things guesses. A guess that sounds confident is worse than no answer, because someone will act on it.

What has to exist first

Before any model is chosen, someone has to write down how the organisation operates: its entities, its processes, its decisions, and the source each fact comes from. We call that the ontology.

It is slower than a prototype, and it is the part that lasts. The models will be replaced several times over its life; the understanding of the business will not.

Who stays

The other reason is simpler. Systems reach production when the people who build them stay inside the business until the result shows in the numbers: cost, speed, capacity. A prototype and a slide are how pilots end.

Production is the test. We build for it from the first day.

Bring us one process. We start there.