Kode-1
optimise

Why AI pilots stall after the demo

21 November 2026

The pattern is familiar enough that people in the room can usually predict it. The pilot works. The demonstration lands. Someone senior says the organisation needs to scale it. Then very little happens for the better part of a year, and by the time anyone asks, the pilot has been quietly retired, the use case has been renamed, and the budget has moved to the next idea. It is tempting to read this as a technology problem. It almost never is. The model did what it was asked to do. The data was good enough to demonstrate with. The platform team delivered what it promised. What failed was everything around the pilot, and it fails in the same few places so consistently that it is worth naming them. The first is ownership. A pilot has a sponsor and an enthusiastic team. A production service needs an owner: a named person accountable for what it does, what it costs and when it should be switched off. Pilots are usually run by people who are measured on launching things, not on running them. When the pilot ends, nobody's role description includes keeping it alive, so nobody does. The second is governance that exists on paper but not in the system. Most organisations now have an AI policy, signed off and filed. Very few can show that the policy is enforced on every request the system handles. So when risk or audit asks the reasonable questions (what did the system do, under which rules, and who approved the exceptions), the honest answer is that nobody knows. A pilot can survive that answer. A production service in a regulated organisation cannot, and the people who would have to sign off on scaling it know that. The third is cost. Pilot costs are small and usually absorbed by an innovation budget. Production costs are larger, variable and land on someone's cost centre. If spend is not attributed to the use case and the business unit that benefits, the finance conversation becomes an argument about a growing cloud bill rather than a decision about value, and arguments about growing bills tend to end in a freeze. The fourth is measurement. If the organisation never agreed what success looks like in business terms, then the question of whether to expand becomes a matter of opinion. The sponsor is optimistic, the sceptics are sceptical, and without an agreed scorecard there is nothing to settle it. Decisions that should be mechanical become political, and political decisions about new technology default to waiting. None of these is solved by buying another platform. The tooling has matured quickly: tracing, evaluation, cost reporting and policy enforcement are all available now, often from the platforms organisations already pay for. But a platform that emits traces is not the same thing as an organisation that reads them and acts. The missing piece is a set of agreements: who owns each live service, which rules run where, how cost rolls up, what gets recorded and kept, and who looks at the scorecard every month with the authority to expand, hold or stop. That work is unglamorous. It makes a poor demonstration and wins no innovation awards. It is also the difference between AI programs that compound and AI programs that keep restarting. It also changes the funding question. Most organisations do not need more pilots. They need the pilots they already have to either earn production or be retired on evidence. Funding a fifth experiment while four successful ones sit in limbo is the most common way AI budgets are spent without anything changing. A practical place to start: pick the single most valuable pilot you already have and give it what production needs before you fund anything new. Name its owner. Enforce its policy in the system, not the document. Attribute its cost. Keep a record of what it does. Agree the three numbers that decide whether it grows, and book the monthly meeting where someone with authority looks at them. Then, and only then, fund the next one.
Field notes

The work, in your inbox.

Occasional notes on strategy, systems, delivery, and risk, written by the partners, sent when there’s something worth saying. Unsubscribe anytime.