← All articlesCoaching & Education

How to pick your first AI use case, before someone picks it for you

September 5, 2026 · 3 min read

After a workshop, the whiteboard holds forty ideas for AI. Most will disappoint. A simple scoring method, and a few ground rules, helps you start with the one that will actually succeed.

  • AI Agents
  • Data & RAG
  • Modernization
  • Education
  • Trust & Safety
  • Coaching

Forty ideas and no way to choose

Run an AI workshop with any team and the result is the same: an energetic list of ideas. Summarise every customer call. Predict churn. Generate all product descriptions. Automate the monthly report. Answer HR questions. Everyone is excited, and then nothing happens, because nobody knows where to begin.

Choosing wrong is costly in a particular way. A first project that fails, or simply fizzles, teaches the organisation that AI does not work, and the lesson sticks for years. The first use case matters less for its direct value than for what it proves to everyone watching.

Your first AI project is mostly an argument. Pick one that you can win.

Score ideas on four questions

Give each idea a score from one to five on four questions, then compare:

  • Frequency: how often does this task happen? Daily tasks give fast feedback and visible savings. A yearly task will not build confidence.
  • Pain: how much time or irritation does it cost today? Ideas people already complain about have willing users.
  • Data readiness: is the material the AI would use available, reasonably clean and legally usable? Without it, the project becomes a data project.
  • Cost of a mistake: if the AI is wrong occasionally, what happens? A weak summary is harmless. A wrong contract clause or payment is not.

Score the first three high and the last one as 'low risk is good'. You do not need a perfect formula. The exercise forces a conversation that otherwise never happens.

The sweet spot: frequent, tedious and forgiving

The ideas that tend to win share a profile. They are frequent, people find them boring, the output is checked by a human before it matters, and a mediocre result is merely annoying.

  • Drafting first replies to routine customer emails, which an agent reviews and sends.
  • Summarising meetings, calls or long documents into short notes.
  • Extracting fields from invoices, forms or orders into a structured format for review.
  • Searching internal documents to answer staff questions, with sources shown.

They are not glamorous. They are exactly what builds trust for the more ambitious projects that come next.

Define 'good' before you start

Choose one measure and write down today's baseline. If the task is drafting replies, measure the average time per reply now. Collect around twenty real examples, including the awkward ones, to test against. Agree what result would count as a success, for example 'a third faster with no more corrections'.

Without this, the project ends in opinions: some people love it, others hate it, and there is nothing to settle the discussion. With it, the result is a number everyone can see.

Set the rules before the tool arrives

Decide, in plain language, a few rules for use. What kind of data may be pasted into an AI tool and what may not? Which outputs must a person review before use? Who is accountable if something goes wrong? Keep it to one page.

People are far more willing to use a tool when they know where the boundaries are. Unclear rules lead to two bad outcomes at once: some staff avoid the tool for fear of mistakes, and others use it carelessly.

Run it for a month, then decide

Time-box the experiment, with a small group of willing users and a regular short check-in. At the end, compare against the baseline and decide openly: keep, improve, or stop. Stopping is a legitimate result, and it should be recorded as one. Share what was learned, including what did not work.

The ideas you did not pick are not wasted. Keep the list with the scores. As the first project builds confidence and data improves, the second and third candidates will already be ranked and waiting.

Keep reading