Start from the task, not the tool
The projects that work begin with a sentence like: every quote takes forty minutes because the pricing lives in three places. That is a problem with a measurable cost and an obvious finish line.
The projects that fail begin with: we should use AI. That has no finish line, so it never finishes.
A useful test: if you cannot say how you would know it worked, it is not ready to build.
Three shapes that reliably pay off
Answering the same question repeatedly. Where the answer exists in your documents and a person is retrieving it by hand, dozens of times a week.
Moving data between systems that refuse to talk. The quiet tax on most small operations. Rarely glamorous, almost always worth it.
Triage. Sorting incoming work by urgency or type so a person starts with the right thing. Getting this eighty percent right is often enough, because the failure mode is a mis-sorted item rather than a wrong answer to a customer.
Where to be careful
Anything where a confident wrong answer reaches a customer without review. Put a person between the model and the customer until you have evidence you can remove them.
Anything touching regulated records. Compliance data wants deterministic rules and an audit trail, not a probabilistic answer.
Anything where the process is not written down. Automating an undocumented process automates whatever the last person did, including the mistakes.
What we do in the first conversation
Ask what took too long last week, and listen for the task that comes up twice.
Ask what would have to be true for it to be worth doing. Usually a number: hours saved, quotes turned around faster, enquiries answered outside hours.
Then say honestly whether AI is the right tool. Often the answer is that a better form, or one system talking to another, gets most of the value with none of the risk. We would rather say that than sell the interesting version.