Article
Why do most AI pilots at small companies never reach production?
Bill Coombes · · 8 min read
Most small-company AI pilots stall for organisational reasons, not technical ones. The tool works. What's missing is a defined piece of work, an owner, a decision about data, and a reason for anyone to change how they spend Tuesday morning. The nine patterns below are the ones that recur: buying logins before defining the work, no named owner, no baseline, unresolved data questions, a pilot with no path to production, training that ends when the session does, an executive sponsor in name only, success measured in enthusiasm, and a stack nobody consolidated.
1. Paying for logins before defining the work
The licences arrive first because they're easy to buy. The work they're meant to change is never written down, so nothing is measurably different a quarter later — and the renewal conversation has no evidence in it either direction.
2. Nobody owns it
Interest is not ownership. A pilot with three interested people and no accountable one moves at the speed of everyone's spare time.
3. No baseline anyone agreed to
If you didn't record how long the task took before, you cannot claim it takes less time now. Agree the baseline while there's still nothing to argue about — ideally with the person who will later question the number.
4. The data question was deferred
Someone eventually asks what happens to client information typed into the tool. If that question is answered for the first time in month three, the pilot stops while it gets answered. A short written position — data handling, disclosure, review — costs a fraction of the delay it prevents.
5. The pilot had no route to production
A demonstration built on a copied spreadsheet proves capability and nothing else. If the production version needs access, budget and a policy decision, then the pilot's real deliverable was a list of things nobody has approved yet.
6. Training ended when the session did
People leave a workshop able to do the thing and, three weeks later, back to their old sequence. Adoption dies in the gap after the room empties, which is why a check-in afterwards belongs in the design rather than in the optional extras.
7. The sponsor was a name on a slide
Executive sponsorship means visibly changing how the sponsor works and asking about it in meetings that already exist. A quote in the kick-off deck is not that.
8. Success was measured in enthusiasm
"People love it" survives no budget review. Two honest numbers beat a survey: who signed in, and what happened to the specific work you defined at the start.
9. Four tools doing the same job
Different teams buy overlapping tools. Nobody consolidates, so the spend rises, the policy has to cover four surfaces, and no single tool gets enough use to look worth keeping.
The pattern underneath
Every one of these is a sequencing failure. The tool arrived before the decision. Fixing it is unglamorous: define the work, name the owner, write the baseline, settle the data position, then build. That order is why the first month of an engagement here produces a plan rather than a prototype.