Most organisations have a long list of ideas for AI. The temptation is to start with the most exciting one. A better rule is to start with the one most likely to work, and that people will notice when it does.

Three tests for a first project

It happens often. A task done hundreds of times a month gives you enough examples to test against and enough volume for small improvements to add up.

You can measure it today. If you already know how long it takes, how often it goes wrong or what it costs, you can prove the result. If you cannot measure it now, you will struggle to defend it later.

A mistake is recoverable. Drafting a reply that a person approves is a good first project. Approving a loan on its own is not. Start where a human review is natural and cheap.

What usually passes

Intake and triage of enquiries, drafting routine correspondence, pulling fields out of documents, answering staff questions from internal policy, and preparing reports that someone currently assembles by hand.

What usually waits

Anything customer-facing without review, anything where the data is scattered across systems nobody owns, and anything where the business case depends on a number you have never measured.

The first project is not the biggest opportunity. It is the one that earns you the right to pursue the biggest opportunity.