The demo went well. Someone on your team, maybe a partner, maybe your operations lead, showed the room what the tool could do with a real proposal or a real client file. People leaned in. Somebody said “we should be using this,” and everyone agreed.
Six weeks later you ask how it is going. Two people use it now and then. Nobody can tell you what it has changed. The license renews next month.
In a firm your size, this is the usual ending for an AI pilot. No decision gets made. It comes up less and less in meetings until it stops coming up at all.
The usual explanation
When a pilot stalls, the explanation tends to land on the tool. It wasn’t accurate enough. It didn’t understand our work. The next version will be better, so let’s wait.
Sometimes that is right. Most of the time it doesn’t survive one question: did the demo work? It did. You watched it do the job in front of the whole room. What it couldn’t do was get itself into anyone’s week.
The best-known study on this also puts model quality low on the list. In July 2025, MIT’s NANDA project reported that 95% of the organizations it studied were getting zero return from generative AI, and that the divide “does not seem to be driven by model quality.” It is a preliminary study, mostly of larger companies, so I hold the number loosely. Its own explanation also blames tools that don’t remember context or adapt to how a company works. What I’d add comes before that, and it is easiest to see in a 30-person firm, where you can name every person who was supposed to use the thing.
What actually stalls a pilot?
Two things are usually missing, and neither is technical. Nobody owns the first job, and nobody has said what “used” means.
Nobody owns the first job
A pilot usually starts as access. Here is the tool, everyone has a login, try it on your work. That sounds generous. In practice it hands the hardest decision, what to use it for, to each person separately, in the middle of client deadlines.
Your account leads try it once on something easy, get a decent result, and go back to the way they have always done it, because the old way is already built into their week. Nobody is wrong to do that. Billable work wins every time it competes with an experiment.
Ownership fixes this in a specific way. One named person is responsible for one job running through the tool, and it sits on their list the same way a client deliverable does. When it breaks, someone notices, because someone is expecting it.
Nobody defined “used”
The second gap is quieter. Ask five people in your firm whether the pilot is working and you will get five different tests. One person means they logged in this week. Another means it wrote a first draft they kept. Your finance lead means it saved money, which nobody measured.
Without a shared definition, a pilot can’t succeed or fail. It drifts. A pilot that drifts for a quarter gets filed as “we tried AI, it wasn’t ready,” and that is much harder to reopen than a pilot that failed a test everyone agreed on.
A smaller start that holds
Pick one recurring job your team already dislikes, give it to one named owner, and decide in advance what “used” looks like in week two. In a firm of 5 to 50 people, each part looks like this.
| Step | What it means | What it is not |
|---|---|---|
| Pick one job | A recurring task your team already complains about: the weekly client status note, first drafts of proposals, turning meeting notes into action lists | ”Use AI for marketing,” or a firm-wide rollout |
| Name an owner | One person whose week includes making it run, with authority to change how the job is done | A committee, or “everyone” |
| Define “used” in week two | A sentence you can check: “every status note this week started from the tool’s draft" | "People are finding it helpful” |
Choose a job people dislike because it takes most of the adoption problem off the table. You are taking something off their plate, and people rarely need persuading to let go of work they resent.
Week two matters because week one is novelty. Everyone tries the new thing once. In week two the job comes back around and the old habit is sitting right there. If the job runs the new way without anyone chasing it, you have something. If it doesn’t, you found out in two weeks instead of a quarter, and you know which of the three parts broke.
What skipping this costs
The license fee is the small part. Most of the cost sits in what a stalled pilot does to the next one.
Your team has now watched AI arrive with a demo and leave without a word. When you raise it again, the partners who were keen last time are slower to put their hand up, and the skeptics feel proven right.
One job that visibly runs every week works in the other direction. People see it, ask about it, and start bringing you the next job themselves.
Where to start this week
Write down three recurring jobs your team complains about. Pick the one that comes around most often. Put one person’s name next to it. Then write a single sentence describing what you would see in week two if it were working.
If you want help choosing that first job and laying out the order of the ones after it, that is the work I do as a fractional CAIO. Here is how that works.
Frequently Asked Questions
- Why do AI pilots stall in small professional-services firms?
- Usually because the pilot starts as access rather than as a job. Everyone gets a login and is told to try it, so nobody owns making one piece of work run through it, and nobody agreed on what counts as using it. Billable work wins every time it competes with an experiment, and the pilot fades without a decision.
- What should a firm's first AI project be?
- One recurring job your team already dislikes, such as the weekly client status note, first drafts of proposals, or turning meeting notes into action lists. A disliked job means nobody needs persuading to hand it over. It should come around often enough that you can tell within two weeks whether it is running the new way.
- How do you know if an AI pilot is working?
- Write one sentence before you start that describes what you would see in week two if it were working, and make it checkable. For example: every status note this week started from the tool's draft. Week one tells you very little, because everyone tries something new once. Week two shows whether the job runs without anyone chasing it.
- Who should own AI in a 30-person firm?
- For the first job, one named person whose week includes making that job run, with authority to change how it is done. A committee or 'everyone' is the same as nobody. Across the firm, someone also has to decide which job comes next and in what order, which is the part most firms that size cannot justify a full-time hire for.