The request usually arrives as a comparison. Which model, which vendor, which platform, and can we have it as a matrix. We can make a matrix. It will not tell you whether AI belongs in the operation. That answer is in the work, and the work is inconvenient to put in a cell.
So we look at the work first, and we say so at the start, before anyone is attached to a pilot.
four questions, in order
What is the job, performed, not described. Where does a person stall, retype, or wait on someone else's knowledge. What would go wrong if a machine did the easy part wrong — money, a customer, a record you cannot quietly fix. And what data do you actually have, in a system someone can query, as opposed to a spreadsheet on a laptop we will not be allowed to see.
If the data is not ready, a dashboard is a drawing of a wish. We will say that. Getting the data fit to use is sometimes the whole engagement, and it is more valuable than a chatbot on top of a mess.
If the honest answer is do not build it, that answer is the deliverable.
the readout is a sequence
What to do in the next ninety days, what to leave alone, and what would have to be true before a bigger bet makes sense. Build when the workflow is yours and the off-the-shelf tool would bend the work out of shape. Buy when a tool already does the job and your people can live in it. Do neither when the process itself is the problem.
You should be able to hand the readout to the person who does the job and have them recognize their week in it. If they do not, we assessed the slide, not the company.