The problem for leaders in 2026 isn’t a lack of AI use case ideas. They have too many. What’s missing is the filter that separates the project worth six months’ work from the project that won’t survive the first in-house demo. Here are the three questions I ask in a scoping meeting. Thirty minutes per use case, no more. If only one of the three remains without a precise answer, we don’t launch.

What specific problem does AI solve?

Not “how AI could improve our productivity”. What specific problem, in what process, with what current volume, what processing time, what quality, what cost?

“Reducing the processing time of customer support requests” is one direction. “Reducing by 40% the processing time for category B requests (simple technical problems), which represent 60% of the volume and are currently processed in 48 hours” is a problem.

The difference is not one of detail. A specific problem defines a measurable success criterion. A measurable success criterion enables you to decide whether the project has worked. Without it, you won’t be able to assess ROI, and you won’t know when to stop.

Why AI and not something simpler?

This is the most uncomfortable question. Before launching an AI project, ask yourself if the problem can’t be solved by a deterministic algorithm (business rule, sort, filter, regex). If not, by a dashboard and a clearer human decision-making process. Sometimes, it’s a process reorganization that solves the problem without touching the technology. Sometimes, it’s just a better-configured existing business tool.

AI has its advantages: it generalizes on unstructured inputs, it adapts to variations, it works on high volumes without proportional cost increase. But it is probabilistic (and therefore fallible), opaque, difficult to audit, and more expensive to maintain than a deterministic rule. Keyword-based document sorting with clear business rules is cheaper, more predictable, and more explainable than LLM for the same task if the categories are well defined.

Who handles cases where the AI is wrong?

That’s the question demos never ask. In a demo, everything works. In production, 3 to 20% of cases (depending on the task and the model) don’t work as expected. Someone has to take care of it.

Who detects errors? How are they reported? Who corrects them? Who is responsible for decisions based on incorrect AI output?

If you can’t answer these questions, your deployment process is incomplete. It’s not a technical problem. It’s a problem of organization and responsibility. AI doesn’t remove responsibility. It shifts it.

On scoring grids

A scoring matrix promises to turn decision into addition. It doesn’t. The decision is made on a case-by-case basis, looking at the context of the organization, the maturity of the team, the opportunity cost of the non-project. Any grid that claims to decide for you is a dangerous simplification.