Where to start with AI in a small business: one person, or one shared process

The usual starting point is not a problem but a feeling: competitors are doing something with AI, and the team is not. That feeling produces vague briefs and stalled pilots. A sharper question gets further – when AI lands in this business, who is it for?

There are two honest answers, and they are different projects.

One person’s workload

Put a good model in front of one person and let them use it for the drafting, summarising and lookup they already do. Nothing about the business changes – no process, no handover, no new system to own. The person gets hours back, this week.

This is where most firms should start, and it is worth doing properly: an hour on what the tool is good and bad at, a short list of the tasks it is allowed to touch, a rule about what never goes into it. The limit is also clear. The gain stays with that one person. Nothing compounds.

One process the whole team shares

The second kind of project picks one process that everyone touches – enquiries in, quotes out; exceptions that need a human; the weekly numbers – and rebuilds it so the routine part runs on its own and only the exceptions reach a person. The leverage goes to the whole team, not one desk.

It is also harder, and the reason is rarely the model. A shared process needs the facts it runs on to live in one place. In most small firms they live in email threads, three spreadsheets and someone’s memory. The first real step is often not an agent at all: one table, one agreed shape for the records, one owner. Build the agent before that and it automates the confusion.

Picking the first shared process

Two criteria, in this order.

• –Lowest build risk – clear inputs, a known-correct answer for most cases, a person who already handles the exceptions and can keep doing so.
• –Most visible daily pain – something the team notices when it stops hurting. A reconciliation nobody enjoys beats a dashboard nobody opens.

The first working piece should go live early, on real work, so it is judged on what it does rather than what it might do. Quality is measured against checks agreed before launch, not assumed. When it holds, the next process is easier: the data is already in one place and the team already trusts the pattern.

If you are not sure which kind of project you are looking at, that is a normal place to be. A written audit answers it in plain language – start with one person’s workload, start with one shared process, or wait – and says what has to be true about your data first. Our three-step process is described on the how it works page, and the first conversation is a 30-minute call.