When the answer is not AI
Perspective
This is an inconvenient thing for an AI company to write, but worth saying plainly: a fair share of the AI projects that get started are solving a problem that did not require AI.
Not because the technology does not work, but because the question was asked backwards. Somebody decided to use AI, then went looking for somewhere to use it. You always find somewhere. Rarely where the money is.
When the problem is that nobody has decided
A team asks for a system that reads incoming cases and suggests which department takes them. Sit down with them and it turns out there is no agreed rule. Four people in the room give three different answers on the same example.
No model fixes that. What is missing is a decision that takes an afternoon.
A model trained on disagreement produces disagreement, only faster.
The tell: you cannot agree on the right answer for 20 examples. If you cannot, the system will not either, and you will never be able to measure whether it works.
When the problem is that the data is not reachable
A request for an assistant that answers questions about customers turns out to be about information sitting in four systems that do not talk to each other, two of them updated by hand.
An AI layer hides this for a while. The assistant answers confidently based on something partly out of date, and the error gets harder to spot than before, because now it arrives in complete sentences.
The tell: when you ask where a number comes from, you get the name of a person rather than the name of a system.
When the rule is known and stable
Some tasks have a precise answer already written down: a rate, a deadline, a threshold. A model can learn to reproduce it, and will get it right most of the time.
Most of the time is a bad trade when the alternative is always. Spend the model on what is ambiguous, and let the rule be a rule.
Where AI genuinely is the answer
To be clear: there are plenty of tasks where nothing else works. Language that varies freely, documents with no fixed structure, and work where the rules exist but nobody has managed to write them down.
What separates the projects that succeed is not the model they picked, but that they spent a week understanding the task first.
The question we ask first
What would happen if we solved this without AI? Often it would take longer or not scale, and then the case is clear. But sometimes it would take 2 weeks and cost a fraction. Worth finding out before, not after.