Most projects do not fail technically. They fail because they solved something that did not need solving, or something that could not be solved that way.
Five filters, in order
1. Is there a decision or an action at the end? If the result changes nothing in what someone does, the project produces a report, not value. \u201cWe would like to understand our data better\u201d is not a use case.
2. Is the frequency sufficient? Automation has a fixed cost. A task done three times a year does not justify it, however tedious. A task done three hundred times a day justifies it even if it saves only a few minutes.
3. What does a mistake cost? If an error is cheap and easily corrected, you can accept an imperfect system. If an error produces a legal consequence, you need very high accuracy, real oversight and documentation — which changes the budget entirely.
4. Do you have the data? Not \u201cthere is data somewhere\u201d, but: accessible data, of reasonable quality, with a legal basis for this purpose, in sufficient volume. This is where most projects stop — and it is better that they stop here than three months in.
5. Is there an agreed success criterion? What figure, measured how, from what value, has to be reached for the project to count as successful? If there is no answer in advance, there will be an argument afterwards.
Good signals
- People already do the task by hand, repetitively, and complain about it.
- There is a \u201ccorrect\u201d result that someone can recognise.
- The decision has a pattern; it is not unique every time.
- Historical data exists about how the task was handled until now.
- An approximate result is still useful.
Warning signals
- The use case came from a supplier, not from inside.
- The framing starts with the technology, not the problem.
- Nobody can say who will use the result, daily.
- \u201cWe will prepare the data as we go.\u201d
- Success is defined as \u201chaving AI too\u201d.
What a proper pilot looks like
A good pilot is small, time-boxed, has a declared stopping criterion and a real user who actually uses it. Its purpose is not to prove that it works, but to find out whether it works — including when the answer is no.
A pilot that cannot fail is not a pilot. It is a demonstration.