The pattern repeats across sectors. A function leader gets enthusiastic about generative AI, runs a small proof-of-concept with a vendor, demonstrates an impressive demo to the executive team, and then nothing happens for nine months. The pilot quietly dies. Budget is reabsorbed. Trust in AI as a credible investment area erodes by another increment.
Across more than fifty engagements, the same four root causes appear. None of them are model failures. All of them are answerable up front. We have come to insist that no pilot kicks off until the sponsor can answer these four questions in writing.
The four questions
- Who owns the outcome — not the tool? A pilot owned by the technology function will deliver a working prototype and stop there. A pilot owned by an accountable business leader — with a P&L, an operating metric, and a reputation at stake — will push it into production. Name the owner before you name the use case.
- What is the value pool, in pounds? "Productivity uplift" and "time saved" are not value pools. They are inputs. Force the conversation to: which budget line shrinks, which revenue line grows, or which risk loss reduces, and by how much? If no one can answer this in a single sentence, the pilot has no path to capital approval for scale-up. Most stalled pilots die here, in month four, when the funding case cannot be made.
- What governance must be in place before scale? The temptation is to defer governance until "we know it works." This is the most expensive deferral in enterprise AI. The controls a high-risk AI system needs under the EU AI Act — risk management, logging, transparency, human oversight — cannot be retrofitted in three weeks. Start them at week one of the pilot. They are cheap then and ruinous later.
- What is the path from pilot to production? Most pilots are built on infrastructure that cannot serve production traffic, with data that has not been cleared for production use, and with vendor terms that cannot be extended to production scale. None of these are surprises if the pilot architecture and commercial model are designed to scale from day one. They almost never are.
What good looks like
Pilots that pass these four tests reach production within nine to twelve months. Pilots that do not pass them — and most do not — typically consume six to nine months of effort, deliver a presentable demo, and quietly disappear from the executive scorecard by the next planning cycle.
The questions are simple. The discipline of answering them up front, before a line of code is written or a vendor is selected, is what separates the organisations scaling AI from the organisations talking about it.
The cost of asking these questions in week one is half a day of executive time. The cost of not asking them is six to nine months of programme effort and the credibility of the AI agenda within the firm.
For the executive sponsor
Before signing off the next pilot, require a one-page note from the proposed owner answering these four questions specifically. Reject any answer that retreats into adjectives. If the sponsor cannot answer them, the pilot is not ready to start — and starting it anyway is how organisations end up with twenty-three running proofs-of-concept and zero scaled use cases.
The boards we work with that are furthest ahead on AI are not the ones that approve the most pilots. They are the ones that approve the fewest, but get them to production.