The pilot decides the program.

Most AI initiatives now die in the pilot.

The pattern usually starts with good intentions and a deadline. Someone picks the most annoying task in the building, automates it with the quickest possible fix, and for two weeks everything is faster. Then the side effects arrive: outputs nobody checks, handoffs that no longer line up, and a team that quietly concludes AI makes work worse. The initiative gets shelved, and — this is the expensive part — the next initiative starts with an audience that has already seen this movie.

The numbers say this is now the norm, not the exception. In S&P Global Market Intelligence's 2025 Voice of the Enterprise survey of more than 1,000 respondents across North America and Europe, 42% of companies reported abandoning most of their AI initiatives — up from 17% a year earlier. The average organisation scrapped 46% of its proofs-of-concept before they reached production. The models did not get worse between those two surveys. The selection discipline just didn't keep up with the spending.

That is the case for treating pilot choice as a strategic decision rather than a procurement one. A pilot has two jobs: prove measurable value, and earn the team's belief that the next one is worth doing. Choose it well and both compound. Choose it badly and you pay twice — once in budget, once in credibility.

Pick a process, not a task.

The most common selection mistake is looking at tasks in isolation. A task viewed alone always looks automatable: the follow-up email, the CRM entry, the report. But tasks are steps in a process, and a process has a purpose — usually a customer at the end of it. Automate one step without mapping the whole and you optimise a fragment: the step gets faster while the process stays broken, or gets faster at being broken.

So zoom out before you zoom in. Map the workflow end to end — where it starts, what feeds it, who touches it, what the customer receives — and only then decide where a machine belongs. This is also why the research on successful deployments keeps finding the same shape: MIT Project NANDA's 2025 research found the rare winners typically bought purpose-built tools from specialised vendors — external partnerships reached deployment about twice as often as internal builds — customised them for one specific workflow and embedded them deeply, instead of scattering licences across the org and hoping. One process, done properly, beats ten tasks done quickly.

There is a second reason to start from the process: tools bend teams, or teams bend tools. A pilot chosen tool-first forces your team to work the vendor's way, and the cost of that bending shows up as non-adoption. A pilot chosen process-first starts from how your team already works — which is the version people actually use.

Frequency × value tells you where to start.

With the process mapped, you will have more automation candidates than budget. Two axes rank them: how often the work happens, and how much a good outcome is worth. That gives four boxes, and a sequence.

The reason the first pilot almost always comes from the high-frequency row is not just the maths of time saved. Frequency is where learning lives — for the system, which improves on volume, and for the team, which builds the habit of working the new way daily instead of remembering it monthly. A brilliant automation of a quarterly report teaches nobody anything; by the next quarter, everyone has forgotten how it works. Frequency compounds. Rarity decays.

Five checks before you commit.

A candidate that survives frequency × value still has five ways to fail. Check all five before a euro moves. Each check is cheap; skipping any of them is not.

1. 01 — It supports a goal the business already has. "We should do something with AI" is not a goal. "Cut proposal turnaround because deals stall in week two" is. A pilot attached to an existing priority inherits its urgency and its sponsor; a pilot attached to a technology has neither.
2. 02 — A number is agreed before you start. Minutes per brief, response time per lead, cost per proposal — one metric, measured for a couple of weeks before the pilot, so the after has a before. This is what makes ROI provable instead of arguable, and it is the difference between "the team feels it helps" and a business case a CFO signs. The metric doubles as a filter: work whose output you can score quickly and objectively — meeting booked, field correct, minutes saved — makes a far better pilot than work where judging correctness is itself expensive.
3. 03 — The data exists and is usable. An AI workflow built on a CRM that was last cleaned two years ago produces confident output from stale facts — proposals quoting old pricing, outreach referencing a contact who left. Check what the system will actually read, not what it theoretically could. If the data needs fixing first, that is not a blocker; it is the real first step, and it belongs in the plan.
4. 04 — The team can actually run it. Fits the existing stack rather than adding a silo, works inside the tools people already have open, and has a named owner — one person accountable for the pilot working, not "the team". Skills count here too: if nobody can adjust a prompt or judge an output, build that in before launch, not after the first bad week.
5. 05 — The risk is survivable. For a first pilot, internal-facing beats client-facing: a mediocre internal brief costs an eye-roll, a mediocre client email costs a relationship. Check the boring failure modes now — privacy, bias, compliance, what happens when the output is wrong and nobody notices. You want a pilot where the worst day is recoverable.

Design the pilot to be seen.

One selection criterion is systematically undervalued: visibility. Pilots that deliver an immediate, concrete result that colleagues can see with their own eyes get integrated fastest — and pilots whose value needs a quarterly dashboard and an explanation do not. This matters because a pilot's second job is belief. The people watching your pilot are deciding, quadrant by quadrant, whether the next rollout deserves their effort. A before-and-after that fits in one sentence — "the meeting brief took thirty-five minutes, now it takes twelve" — recruits more support than any launch deck.

The discipline continues after the pilot works: hold, measure, and only standardise once usage is stable and the number has moved — the 75–85% adoption band covered in N°01. A pilot is not a small rollout. It is the proof that a rollout deserves to exist.

Diagnosis comes before prescription. That's where every engagement starts.

1. S&P Global Market Intelligence · Voice of the Enterprise: AI & Machine Learning, Use Cases 2025 (1,000+ respondents, North America & Europe; reported March 2025). 42% abandoned most AI initiatives, up from 17%; the average organisation scrapped 46% of proofs-of-concept before production.
2. MIT Project NANDA · The GenAI Divide: State of AI in Business 2025 (2025). The successful-deployment pattern: purpose-built tools embedded in one specific workflow. Released as preliminary, non-peer-reviewed findings; cited here for the qualitative pattern.
3. Harvard Business School Online · AI Essentials for Leaders. Frequency × value prioritisation; use-case selection criteria; early-piloting characteristics; 75–85% adoption benchmark.