How do you pick your first AI project?

Pick one task that happens every week, is text-shaped, has low stakes if the output is imperfect, and is owned by someone keen to try. Run it for a fortnight with a before-and-after time measure, keep a human on the final check, and let the result, not the technology, decide what happens next.

Each of those filters is doing a job. Weekly, because a task with rhythm produces evidence quickly and builds habit; a quarterly task teaches nothing for months. Text-shaped, drafting, summarising, first-pass admin, because that is where current tools reliably earn their keep. Low stakes, because the first project is a classroom and mistakes should cost embarrassment at most. And owned by a volunteer, because enthusiasm survives the awkward first week that kills mandated experiments.

Candidates that pass the filters in most small businesses: the weekly client update or newsletter draft, meeting notes and actions, first replies to routine enquiries, turning job notes into structured records, or the monthly report nobody enjoys assembling. Candidates that fail: anything that goes to customers without review, anything involving pricing or people decisions, and anything beginning with the words transform or strategy. The first project should be almost disappointingly boring, because its real output is not the time saved; it is what the business learns about where these tools fit.

The honest caveat is that first projects are seductive in both directions: a good fortnight tempts a company-wide rollout, and a bad one tempts writing the whole technology off. Resist both readings, because one task is one data point. The move that compounds is sequential: measure, decide, pick the next task, and keep the foundations, sanctioned tools, a short policy, a human on the output, growing alongside the ambition. If you want the first task chosen and set up properly, call 1800 456 567.