A company's first AI use case carries a lot of weight. It sets the teams' confidence, it serves as an example, and it often reveals the real obstacles: scattered data, unclear responsibilities, poorly documented processes. Before launching the project, an executive can ask five questions. None of them requires technical expertise.

1. What precise problem are you trying to solve?#

"Using AI" is not a problem. "Preparing the first draft of an investment memo from the data room" is one. The right question is about an identifiable task, carried out today, by identifiable people, with a cost in time or a risk of error.

If nobody can describe the task in a few sentences, the use case is not ready.

2. Which data is needed, and where is it?#

A use case depends on its data. You need to know where it is, who controls it, in what format it exists and whether it can be used. Some projects fail at this stage, not for lack of technology, but because the information is scattered across inboxes, spreadsheets and personal folders.

The question to ask: if we wanted to automate this task tomorrow, what information would we start from?

3. Who decides whether the result is good?#

An AI output must be assessed by someone who knows the subject. Before launching the project, name that person. Also define what a good result is: a complete memo, a list of points to check, a reliable classification.

Without a shared quality criterion, everyone judges by gut feeling, and the project loses credibility at the first mistakes.

4. What is the measured starting point?#

To know whether the use case adds anything, you need to know the situation before. How long does the task take today? How many people do it? What error rate is observed? These measures can be simple, but they must exist.

A project without a starting point cannot prove its value. It ends up being judged on the impression it gives.

5. What happens if the result is wrong?#

Every AI makes mistakes. The useful question is therefore: what does an error cost, and who detects it? A poorly worded internal summary is easy to correct. A flawed analysis sent to an investor cannot be corrected the same way.

The answer determines the level of human oversight: systematic review, sampling, or use restricted to preparatory tasks.

Pitfalls to avoid#

  • Choosing the tool before the problem.
  • Launching the project on the most accessible data rather than the most useful.
  • Confusing an impressive demo with everyday use.
  • Not planning training for users.

Key takeaways#

A first use case should be small, measurable and reviewed. It should address a real problem, on available data, with a clear owner. The answers to these five questions are often worth more than the choice of tool.

Further reading#