Why AI pilots stall at professional services firms
The pilot almost always works. That is the confusing part.
A firm runs a proof of concept, something genuinely useful comes out of it, the room is impressed, and then six months later the P&L looks exactly as it did before. Nobody can point to the moment it died. It simply stopped being anybody's job.
Three reasons account for most of it, and none of them are about the technology.
One: the knowledge left when the vendor did
The most common shape is that a supplier builds something, invoices monthly, and keeps the understanding of how it works. The firm has a working system and no ability to change it. So when the process shifts, and it always shifts, the system quietly stops matching the work, and people route around it.
The test is simple and you can apply it before signing anything. If this supplier disappeared tomorrow, could somebody in your building change what was built? If the answer is no, you have not bought capability. You are renting it, and the rent includes never learning how it works.
Accounts opened in your name, in your accounts, is the boring version of the fix. So is having your own people in the room while the thing is built rather than trained on it afterwards.
Two: it was built for people instead of with them
A system that arrives finished is a system somebody has to be persuaded to adopt. A system whose future user helped decide how it works is one they already understand.
This is not a motivational point. It is about where the knowledge sits. The person who does the work knows the eleven exceptions that nobody documented. If they are not in the room, the build handles the clean case and breaks on the fourth invoice, and after that nobody trusts it.
The other half is where it lives. Anything requiring people to log into something new is competing with a habit. Things that run inside the tools already open win by default.
Three: one person was trained, and then they left
Firms of this size usually appoint someone. The keen one. They go on a course, they get good, and every AI capability the firm has now sits with a single employee.
Then they resign, or move teams, or get busy, and the capability leaves with them. This is not a hypothetical risk, it is the normal outcome, because keen people with new skills get offers.
Four is the smallest number that fixes it. Not because four people are twice as good as two, but because at four the knowledge stops being a person and starts being a practice. Four is also the smallest group that argues about the right approach, which is where most of the actual learning happens.
What working looks like instead
The thing that changed for firms who got past this was rarely a better model or a bigger budget. It was that one process went the whole way from manual to running, in production, owned by people who still work there.
One finished thing teaches a firm more than five pilots. It also settles the internal argument about whether any of this applies here, which is the argument that quietly stalls everything else.
If you are choosing where to start, do not start with the most exciting process. Start with the one that is expensive, boring and repetitive, and take it all the way.