Where pilots actually stop
The model is rarely the hard part. By the time an AI pilot is scoped, the team usually knows what it wants to build and what success looks like. What is missing is everything between a working demonstration and a system people can rely on: where the data comes from, who is accountable when an answer is wrong, and what happens when the input is not the shape the pilot assumed.
Three questions worth answering first
- Where does the data actually come from, and who owns it?
- What does a person do when the system is wrong?
- Who is accountable for the outcome, not the model?
What changes the outcome
Pilots that reach production have usually done less model work than expected and more work on the boundaries around it. Discovery is where that is settled, and it is cheaper to settle it there than it is to discover it after launch.