Before choosing a model, choose the task.

Name the job, not the ambition.

“Use AI in the business” is too broad to test. “Find the delivery condition in an approved product manual” is specific enough to examine. So is “turn these meeting notes into a draft action list”. A narrow task does not make a project unambitious. It gives the project a starting point that people can agree on.

Bring a small, varied set of examples.

Choose some ordinary cases and a few difficult ones. Include incomplete material, ambiguous wording and a question the source cannot answer. Keep the same examples for each candidate. Otherwise, a polished demonstration may tell you more about the choice of examples than about the tool.

Write down what a good answer includes.

Correctness is not the only consideration. A result may need a source reference, a consistent format or a clear indication that the information is missing. Decide which mistakes make an answer unusable and which ones are easy to correct. Ask the person who will do the work to take part in that decision.

Count the work around the answer.

Notice the effort spent preparing inputs, checking outputs and dealing with exceptions. A tool that produces a quick first draft is not automatically a useful workflow. It becomes useful when the surrounding steps are manageable and the result can be handed over with confidence.

Keep the first decision reversible.

Begin with a small prototype and a simple comparison record. Record the model version, settings and examples used. These notes make later changes easier to explain. The first choice does not need to be permanent; it needs to teach you enough to make the next choice well.