Starting from the technology
“We should use AI” produces projects looking for a problem. They demonstrate well and never quite justify themselves.
The fix is to start from an expensive, repetitive problem and ask what would solve it. Sometimes that is AI; often it is a rule or a form field.
No agreed definition of correct
If nobody can state what the right answer is for a hundred specific cases, there is no way to measure the system, no way to improve it and no basis for deciding it is ready.
This is the single most common cause of a project that runs and never lands.
Skipping the review interface
Cut to save budget, and it is what determines whether the system saves time. Without it, output that is 85% right is unusable, because checking is slower than doing.
Automating the exception path first
The hard 20% is tempting because it is where the pain is loudest. It is also where automation is least reliable and most expensive.
Automate the routine 80% and route the rest to people. That is where the return is.
The remaining four
- No owner after launch — quality drifts and nobody notices
- No cost cap — a loop spends the quarterly budget over a weekend
- Launching to everyone at once — no chance to correct before reputation is set
- Measuring nothing — so nobody can say whether it worked, and it gets cut at the next budget review
What the successful ones share
A narrow first scope, a domain expert who was available, a review interface that was fast, a named owner, and a measurement agreed before launch.
None of that is about the model, which is the point.