Which department should pilot AI first? Four questions to ask

You have read your score and chosen one dimension to move this quarter. The very next question, and the one organisations get wrong most often, is which department carries the pilot. Get this choice right and the pilot produces evidence the rest of the organisation trusts. Get it wrong and you spend three months proving something nobody outside that department believed needed proving.

This is a sequencing decision, not a reward. Treat it as one.

The instinct that leads you wrong

The obvious answer is the department that already wants it: the enthusiastic manager, the team that has been asking for a pilot for months. That instinct is understandable and usually wrong, for a simple reason. A department that is already keen has often already found its own workarounds, informally, which means the pilot measures a head start rather than a genuine before-and-after. The other common wrong answer is picking whichever department is under the most external pressure, a funder deadline, a board question, because urgency is not the same thing as readiness.

What actually predicts a good pilot department

Three things matter more than enthusiasm. First, enough task volume and repetition that a pilot produces real evidence within weeks rather than months; a department that does the same kind of task five times a week generates a result faster than one that does it five times a year. Second, a manager who already measures something, anything, because a pilot run by someone who has never tracked an outcome before will produce impressions, not evidence. Third, a genuine baseline of fluency already in place, so the pilot tests the actual intervention rather than accidentally re-running the fluency training from an earlier stage.

The department to avoid first, even if they ask

We have written before about the four-class data table and what counts as Restricted information: casework, safeguarding records, anything that identifies a vulnerable person. A department working mostly with that class of data is a poor choice for a first pilot, not because the work is unimportant but because a first pilot needs room to fail safely, and that room does not exist where a mistake touches a beneficiary directly. Let that department go second or third, once the method has already been proven somewhere the cost of an early mistake is lower.

Four questions to ask before naming a department

• Is there enough repeated volume to produce evidence in weeks, not months? A task done constantly beats a task done occasionally, however important the occasional task is.
• Would a mistake here be recoverable or catastrophic? Choose recoverable for the first pilot. Save catastrophic-risk departments for later, once the method is proven.
• Does the manager already measure things, or would they need to learn how, mid-pilot? A manager without a measurement habit will hand you a story, not a result.
• Is the enthusiasm coming from the department itself, or from whoever is appointing them? Borrowed enthusiasm evaporates the first time the pilot hits friction.

The best pilot department is not the one asking loudest. It is the one that can tell you honestly, in six weeks, whether it worked.

The trade-off in choosing safety over difficulty

Picking the safest, easiest department first is the right call for a first pilot, and it has a real cost worth naming: the evidence it produces may not transfer cleanly to your hardest department, the one everyone actually wants convinced. A finance team pilot that goes well does not automatically persuade a casework team that the same approach is safe for them. The honest answer is that the first pilot buys credibility and a working method, not a universal proof, and the second pilot, in a harder department, is where that method actually gets tested against scepticism worth taking seriously.

The practical takeaway

Choose the department with high task volume, low catastrophic risk if something goes wrong, and a manager who already measures things; treat enthusiasm as a bonus, never as the qualifying criterion.

No external statistic cited; this article presents an internal decision framework rather than third-party evidence.