The most dangerous word in SME AI adoption is "still."
The pilot is still running. The team is still getting used to it. It's still early days. I sat with a founder recently whose AI pilot was in its fourteenth month. When I asked how it was going, he said — without irony — that it was still promising. Fourteen months. A promise that old isn't a promise anymore. It's a subscription with hope attached.
He's in good company. Anthropic's enterprise playbook — the guide they've distilled from their customers' deployments and their own internal teams — opens with a statistic worth sitting with: 92% of companies plan to invest in generative AI, and 1% believe their investments have reached maturity. Ninety-two and one. That gap isn't a technology gap. It's a discipline gap, and most of it lives in one specific failure: pilots that were never given permission to end.
Eight to Twelve Weeks. Then a Decision.
The playbook's own numbers are blunt. A well-run pilot shows its first measurable signal by week four to six. Adoption patterns and quality effects are clear by week eight to ten. And the line I'd frame and hang in every SME boardroom: if you're not seeing results by week twelve, adjust your approach or your use case — not your timeline.
That's what a kill date is. Not pessimism. A decision, scheduled in advance, with only two exits: scale it, or stop it. The pilot that can run forever can't fail — which, as I wrote in the last essay, is another way of saying it can't succeed.
An experiment that isn't allowed to end isn't an experiment. It's a habit with a login.
A CFO recognises this instantly, because it's capital discipline wearing new clothes. Every month a zombie pilot shuffles on, it consumes two budgets: the subscription line you can see, and the credibility line you can't. The second one is dearer. When the pilot that was never measured finally gets cancelled, the story the team remembers is "we tried AI and it didn't work" — and the next attempt, the properly designed one, inherits that scepticism at compound interest.
Pick a Pilot That's Cheap to Kill
The kill date only works if killing is actually affordable, which is why pilot selection matters more than pilot enthusiasm.
The playbook's guidance matches what I've seen in every engagement: never pilot on customer-facing work or mission-critical processes, no matter how tempting the upside. Pick something contained, repetitive, and expensive — the internal report nobody enjoys writing, the data entry that eats a Thursday, the first draft that always starts from a blank page. If the pilot dies at week twelve, the business shrugs. If it lives, you scale it with evidence instead of adrenaline.
And when the pilot surfaces five new ideas by week three — it will; good pilots are generative like that — write them into a next-phase list and keep walking. Scope creep is how a twelve-week pilot becomes a fourteen-month promise. The backlog isn't bureaucracy. It's where good ideas wait their turn instead of murdering the current one.
The Post-Mortem Is the Product
Here's the part almost everyone skips: the pilot's real output isn't the time saved. It's the learning — and the learning has to be collected deliberately, at the kill date, while it's fresh.
The numbers matter, of course: hours before, hours after, error rates, who used it and who quietly didn't. But the stories matter more. The workaround someone invented in week five is next quarter's process design. The friction that made one team resist is the objection you'll meet again at scale, and now you have the answer. A one-hour honest post-mortem — what worked, what didn't, why — is worth more than the pilot's entire subscription cost. It converts a twelve-week experiment into an institutional asset, whichever way the decision went.
Stopped pilots aren't failures. Unmeasured ones are.
Put a Date On It
So: one process, contained and expensive. A baseline taken before day one. A target written down. An owner with a name. And a kill date on the calendar — week twelve, with a decision meeting booked now, not "when we get a chance."
That's the entire method. It's also, not coincidentally, how the AI Transformation Programme runs every deployment — because the firms whose AI spend shows up in the P&L aren't the ones that avoided killing pilots. They're the ones that killed them quickly, learned deliberately, and scaled the survivors with the confidence of a decision already made.
If you're weighing whether your business is ready to run a pilot properly — honestly ready, not conference-keynote ready — start with the three-minute readiness check. It's adapted from Anthropic's enterprise playbook, translated into questions an SME owner can actually answer.
The tools will keep improving without your permission. The discipline won't.