Where to Start Your First AI Project (Hint: Not With the Interesting One)

Someone in your business has an AI idea. Maybe it's you. It's usually a good one — clever, ambitious, the kind of thing that would genuinely change how the place runs if it worked.

That's exactly why it's the wrong place to start.

Not because the idea is bad. Because a first project has a second job on top of doing the work: it has to prove to everyone watching that this stuff actually lands. And ambitious ideas are the worst possible thing to prove anything with.

The first project isn't about the work — it's about trust

Think about what happens after a first project succeeds. Your team stops bracing. The sceptic in accounts starts asking whether you could do the same for their pile. Nobody has to be convinced any more, because they've watched it happen.

Now think about what happens after a first project stalls. It doesn't just fail — it becomes the story. We tried that. It didn't work. Every future suggestion gets measured against it. You don't lose one project; you lose the appetite for the next five.

So the first one has to land. Which means picking for certainty, not for upside.

The four-part test for a good first job

Run any candidate through these. A good first project passes all four. If it fails even one, park it — it's a second or third project, not a first.

• It happens a lot. Dozens of times a week at least, ideally daily. Rare tasks make bad first projects, because you can't tell within a month whether anything improved. Frequency is what makes the result obvious.
• It's the same job every time. Not identical, but the same shape. Answering "when's my delivery?" is the same shape every time. Handling a supplier dispute isn't. Same shape means you can describe what "done properly" looks like, which means you can tell if it's being done properly.
• Someone can name what it costs. "Louise spends about two hours a day on this." "We lose maybe six enquiries a week to slow replies." You need a number, even a rough one, because it's the thing you'll point at afterwards.
• A mistake is survivable. If the worst realistic outcome is an awkward reply that a human then fixes, you're fine. If the worst realistic outcome is a wrong invoice going to a client or a safety-critical detail getting missed, that job needs to wait until you've built up some confidence in the setup.

Notice what's not on the list: how interesting it is, how impressive it sounds, how much of the business it touches. Those are the qualities of a good third project.

What this looks like in real businesses

A plumbing firm. The exciting idea was AI-assisted job scheduling across seven vans. The first project was answering the phone after 6 p.m. — which took twenty minutes to explain, ran the same way every night, and recovered four or five jobs a week that used to go to whoever the customer rang next. Scheduling came later, and went in far more smoothly because by then nobody was arguing about whether the technology worked.

An accountancy practice. The exciting idea was drafting client advisory notes. The first project was chasing missing paperwork in the run-up to a filing deadline — same message, hundreds of times, the sort of thing everyone hated doing and no one wanted to own. It gave back the better part of a week per partner in January.

A wholesaler. The exciting idea was demand forecasting. The first project was pulling order details out of emailed PDFs so nobody had to retype them. Deeply unglamorous. It also removed the single most common source of order errors in the business, which turned out to matter more than the forecasting would have.

In all three, the pattern is the same: the boring job went first, worked, and bought permission for the interesting one.

How to actually pick yours this week

You don't need a workshop or a strategy day. You need about an hour and three questions asked of the right people.

Ask what people do that they'd describe as "the same thing over and over." Don't ask what should be automated — that's a question about technology and people will guess. Ask what's repetitive. Everyone knows their own answer to that immediately.

Ask where things go quiet. Where does work sit and wait? Enquiries that don't get answered until tomorrow, quotes that take three days, invoices that only get chased when someone remembers. Silence in a process is nearly always a job nobody has time for.

Ask what people are afraid of getting wrong. This tells you where the boundary is. Whatever comes up here goes on the "later, with a human checking" list — not the first-project list. Knowing what to exclude is half of picking well.

Then take the shortest item that passes all four tests, and start there. Resist the urge to widen it. If the job is "answer after-hours calls," it is not also "book the appointment, take the deposit, and update the CRM." Those come next month, once the first bit is boringly reliable.

The uncomfortable part

The right first project usually feels too small. That's the feeling of doing it correctly. You will be tempted, roughly two weeks in, to bolt on the interesting idea because things are going well. Don't. Let it run plainly for a month, gather the number you promised to gather, and then expand from a position of proof rather than optimism.

Businesses that get AI working tend to have an unremarkable first project and a very good second one. Businesses that don't tend to have one ambitious project that's still nearly finished.

If you'd like an outside read on which job in your business passes the four-part test, our free AI-readiness audit walks through it in a few minutes and tells you straight. You can also see how we work if you'd rather someone ran the first one with you.

At BuildPulse, we'd rather ship your dullest problem in six weeks than your most exciting one never.