How to pick your first AI use case, and what to avoid

The most common way an AI programme dies is that the first project was chosen to be impressive. It targets the hardest process in the company, runs for two quarters, and produces a demo nobody can put into production. By then the budget conversation has moved on.

The four questions we ask

• ✳How often does this run? Weekly is fine, daily is better, twice a year is not a first project
• ✳Do we already measure it? If not, you will have no way to show it worked
• ✳What happens if the output is mediocre? If the answer involves a regulator or a customer, pick something else
• ✳Who owns the process? A project without a named owner has no one to argue for it

What to avoid first

Avoid anything where a wrong answer is expensive: hiring decisions, credit, medical or legal advice, anything with a compliance obligation attached. Those are legitimate use cases eventually, but not while your team is still learning what the tool does badly. Avoid processes that need three departments to agree, too. The technology will not be what delays you.

A good first project looks boring

It is usually something like drafting responses to a category of routine enquiry, or turning a recurring report into a first draft. Frequent, low-stakes, already measured, and owned by one person who wants it to work. That combination is what produces a number you can take to a board, which is what buys the second project.