Against the pilot project
A pilot is a way of spending real money to avoid making a decision. Do one small thing properly instead.
The standard shape of a corporate AI engagement is a pilot. Three months, a limited scope, a steering committee, and a report at the end. The report is almost always positive, because a pilot designed by the people who want the project is not an experiment, it is a demonstration.
Then the rollout stalls, because the pilot deliberately avoided every hard part. It ran on clean data that somebody prepared by hand. It ran indoors. It ran with the vendor present. None of the integration was real, none of the edge cases were in the sample, and nobody had to own it at three in the morning.
The pilot succeeded. It always does. That is the problem with pilots.
The alternative is not a bigger pilot. It is to pick the smallest thing that would actually be used in production, and put it into production, with the ugly data and the real environment and a named person who owns it. One camera on one line. One document type from one supplier. One cell, one shift.
It is less impressive in a steering committee and it tells you more in six weeks than a pilot tells you in six months, because the thing you learn from a real deployment is what breaks, and a pilot is specifically constructed not to break.
When we are asked to quote a pilot we will usually propose this instead, and occasionally we lose the work over it. That trade is fine. A pilot that leads nowhere is worse for us than no engagement at all, because we end up in a case study nobody can point to.