Most companies have an AI pilot that worked in a demo and never reached the business. Two weeks, fixed fee: we establish why it stalled, whether it can be recovered or has to be rebuilt, and what it would take to turn it into a number your finance director recognises.
The gap between those two is where most AI budgets are sitting right now. A demo needs a good day and a willing audience. Production needs the process rewritten around it, a person accountable for the output, data it can reach every morning, and someone able to say what it saved. Pilots rarely fail on the model. They fail on everything downstream of it.
The pilot belonged to an innovation team. Nobody in the line organisation was accountable for it working on a Tuesday.What breaksIt cannot be escalated, so it quietly stops.
Nowhere for a person to intervene, so the first bad output became a reason to switch it off rather than a case to handle.What breaksTrust goes before the value arrives.
It ran on an export. Nobody built the permissioned route it needed to run continuously.What breaksIt works once and cannot be scheduled.
Nothing was measured before it started, so there is no before to compare against and no case to expand on.What breaksFunding has nothing to point at.
A short document with a decision in it, and the arithmetic behind the decision.
Named, specifically, against your pilot rather than a general list. Usually two or three reasons, and rarely the model.
Whether what exists can be carried into production, whether it is cheaper to rebuild on proper ground, or whether the use case was wrong.
What has to be measured, from when, and which figure the business will accept. Agreed with whoever owns that figure before anything is built.
The work between here and production, sequenced, with an honest view of effort. Including the parts your own team should do.
A figure chosen after the fact is a figure nobody believes. So the review ends with one measure, written down, that the person who owns that part of the business has accepted as fair. Cycle time, cost per case, throughput, error rate, whichever is theirs.
That is what turns an AI project into something a board can fund again. Not a demo that impressed people, a movement in a figure that was already being reported.
The pilot carried into production in four weeks, on a proper data path, with gates and the record switched on.
Where the pilot failed because the operating model underneath it cannot hold anything. Six to eight weeks, installed in your team.
Sometimes the answer is that the use case was wrong. Two weeks and a fixed fee is a cheap way to find that out.
Two weeks, fixed fee, a decision at the end of it. Bring us the pilot that has been nearly ready for a year.