AI consulting has found its favourite product: the workshop. User interviews, journey mapping, ideation, use-case prioritisation, then a prototype that impresses everyone in the meeting room. Every step is useful. None of it is the work.
The work is a system that runs on a Monday morning, wired into your real tools, with an owner, controls, and a number that moves. Everything before that is a promise.
The graveyard nobody puts on a slide
The numbers of this economy are public and brutal. MIT measured that 95% of enterprise generative-AI initiatives show no impact on the P&L. Gartner projects that over 40% of agentic AI projects will be cancelled by 2027. The pattern is not a technology problem. The models work. The projects die between the demo and the operation.
Almost every one of those dead projects has something in common: it was launched beautifully. There were workshops, personas, a polished readout, an applauded prototype. Then the prototype met reality, the existing systems, the compliance review, the teams who have a job to do, and nobody had signed up for that part.
An applauded prototype is not a system in production. It is a promise with the budget already spent.
Why the pilot economy persists
Because it is comfortable for everyone. The workshop is sellable, pleasant and risk-free: nobody was ever fired for running a discovery phase. Production is dated, measurable, and can fail in public. So the profession organised itself to sell the part that cannot fail and leave the client the part that counts.
You can see the result in the org charts: whole teams for user research and design, and nobody whose mandate ends at "the system runs, the teams use it, the number moved".
What we would do instead
- Start from the production conditions, not the ideation: which systems, which data, which compliance review, which operational owner. If a use case cannot meet them, kill it before the workshop, not after the prototype.
- Size discovery in weeks, not quarters. The real problem shows up on the floor, inside the actual workflow, faster than it shows up in a workshop room.
- Govern by construction: controls, traceability and accountability set from the first workflow, so leadership can sign the deployment.
- Measure adoption at 30 days, not satisfaction at the end of the workshop. If usage does not hold, the project did not happen.
If you already have the prototype
Many companies are exactly there: a paid pilot, a successful demo, and months of silence since. The good news is that a stuck pilot usually has a fixable problem. It was never wired into the systems that carry the work, it has no owner, or it cannot pass the compliance review. That can be audited, that can be repaired, and it costs less than starting over with another workshop.
Discovery has value. We practise it too, and a week on the floor beats every sticky note in the world. But it is a means, not a deliverable. The deliverable is AI in production. As long as the profession sells the journey instead of the destination, 95% of projects will keep dying to applause.