Why most AI pilots die in month three.
The model is almost never the reason. It's one of three things every time, and each one has a fix.
We have watched a lot of AI pilots die. Almost never because of the model. The model usually did what it was asked. Something around it gave way.
It happens at roughly the same point every time. The demo was in month one. Leadership was pleased. Then month two is quiet, and by month three someone asks who's still using it, and nobody is sure. The Slack channel has gone silent. The vendor has moved to the next client. The pilot isn't cancelled. It is just not there anymore.
When we go back and look, it is one of three things. Usually two of them.
1. Nobody owned it after the demo.
A pilot has a champion. The champion gets it approved, sits in the sessions, and shows it to the boss. Then the champion gets promoted, or goes on leave, or is pulled onto the next urgent thing. The vendor's engagement ends the same week. Now the pilot belongs to everyone, which means it belongs to no one.
The fix is dull and it works: one named owner on the client side who's still there in month six, with the pilot written into their actual job. It has to be one person with a calendar, because a steering committee belongs to no one either.
2. It never touched real data.
The demo ran on a clean sample. Twenty tidy records, a handful of well-formed questions, a happy path. It looked wonderful. Then it met the real CRM, with the duplicate accounts, the field someone has been using as a notes box since 2019, and the fifteen percent of records with no email address at all.
Real data isn't a phase you get to later. If the pilot can't run on a live copy of the real system by the end of week one, the pilot is a slideshow with extra steps. We put real data in on day three, when the product is still ugly enough that nobody minds it breaking.
3. No one planned for Monday.
Who opens it at nine in the morning? What do they do when it is wrong, because it'll sometimes be wrong? Who fixes it that afternoon? What does it cost per month, and whose budget is that? These are boring questions and they decide everything.
A pilot that has an answer to each of them turns into a product. A pilot that doesn't turns into a story people tell at the next vendor pitch.
Fix those three before you write a line of code, and the AI part becomes the easy bit.
What we do about it
Our ten-day build starts with two days of exactly these questions, answered in writing, with the people who can say yes. It is the least glamorous part of the calendar and the only part we refuse to skip. Then real data on day three, real users on day six, and a demo on day ten that ends with a decision: go, change course, or stop.
If you have a pilot that's already stuck in month three, that's a normal place to be, and it is usually recoverable. We have an offer built for exactly that.