The honest no

Every enterprise AI programme has evaluation criteria. Almost none of them have ever stopped anything. That gap — between criteria that exist and criteria that can kill — is where most of the waste in enterprise AI lives.

The mechanics are predictable. A pilot is commissioned with enthusiasm and a budget line. By the time results arrive, the programme has a steering committee, a slide template, and a sponsor whose name is attached to it. The evaluation is then performed in an environment where a negative result embarrasses specific people. Under those conditions, thresholds soften. Success is redefined mid-flight. The pilot ‘shows promise’ and rolls into scale on momentum rather than evidence.

This is not a technology failure. The models did what models do. It is a governance failure: the organisation never created a moment where ‘no’ was a legitimate, career-safe answer.

Gates only work if they predate the work

The fix is structural, and it is cheap. Evaluation thresholds must be written before the pilot begins — specific numbers, on specific measures, on production data, agreed by both the sponsor and the delivery team. Not because the numbers will be perfectly chosen, but because a threshold set in advance belongs to the programme, while a threshold set afterwards belongs to whoever needs the result.

The second half of the fix is making the negative outcome a deliverable. A pilot that misses its thresholds should close with a written verdict: what was tested, what the evidence showed, why the recommendation is stop — and what would have to change for the answer to change. That document has real value. It prevents the same idea from being re-piloted eighteen months later by a different team with the same result and a fresh budget.

The commercial logic of saying no

A consultancy is structurally tempted to never recommend stopping, because stopping ends billable work. We think that maths is wrong over any horizon longer than one engagement. An advisor who has demonstrably killed workstreams against their own commercial interest is the only advisor whose ‘yes’ contains information. Everyone else’s yes is a pricing signal.

So we put the honest no into the operating system rather than leaving it to individual courage: thresholds in writing before code, verdicts in writing after evidence, and a delivery model where no phase advances without its gate. It is a small amount of process. It is most of the difference between an AI programme and an AI budget.