Plenty of organisations feel pressure to do something with AI without a clear sense of where to begin. The temptation is to pick the most impressive idea on the list. A better first project teaches you how AI behaves in your environment while delivering something people genuinely value.
Five things to weigh up
Before getting excited about a use case, it's worth scoring it honestly against five questions. None of them need a technical background to answer.
- Business value: if this worked, what would actually change, and for whom?
- Data permission: is the information available, and are you allowed to use it this way?
- Error consequences: what happens when the output is wrong, and who would notice?
- Human-review effort: how long does it take a person to check the output properly?
- Operating cost: what will it cost to run, monitor and support once the pilot's over?
Use cases that score well on all five, like drafting responses from approved material, summarising long documents or finding answers in internal procedures, tend to deliver early value without exposing you to unnecessary risk.
When AI isn't the answer
Take a hypothetical example: invoices arrive as structured PDFs from the same five suppliers, and someone re-keys them into the finance system. AI could read them. But the layouts never change, the rules are fixed and a mistake means paying the wrong amount. Conventional automation, a supplier portal or simply asking suppliers for a data file will usually be cheaper, more predictable and easier to support.
And sometimes the honest answer is no change at all. If a task takes ten minutes a week and nobody minds doing it, the effort is better spent elsewhere.
Keep people in the loop by design
Human review isn't a sign the technology has failed. It's how you build confidence, catch the odd cases and learn where automation can safely go further. Designing that review step from day one, with clear logging and an easy way to flag problems, is often the difference between a pilot that stalls and one that grows.