How to Choose Your First AI Project

The audit is done. You have a ranked queue of things AI could help with. And now the most consequential small decision of the whole programme lands on the table: which one do you actually build first?

The instinct is to reach for the exciting one. The client-facing feature, the sales tool, the thing you could put in a pitch. Resist it. Your first AI project is not really about the return on that project, it is about what it buys you: the credibility and the political runway to do the next ten. Get that wrong and the whole programme stalls on a stalled moonshot. Get it right and you earn permission for everything after. This is how to choose it on purpose.

Why Your First AI Project Should Be Boring

The most quietly correct instruction in every AI rollout map is that the first real build should be a quick win to some back-office process, with attributable hard ROI, where cultural pushback is expected to be low. Boring, in other words. And boring is the point.

Here is the logic. Nobody in your business has seen AI work here yet. They have seen the demos and the hype, but not a real result on real work in their own company. The first build's job is to convert that abstract belief into a concrete fact everyone can point at: this saved us a day a week, and here is the day. Once that fact exists, every future conversation gets easier, budgets loosen, and the sceptics quiet down. A boring back-office win produces that fact cleanly. An ambitious client-facing project, even when it works, produces argument, because the stakes are high enough that everyone has an opinion and something can always go wrong in front of a customer.

You are not optimising for the biggest win. You are optimising for the most certain, most visible, least contested one. Those are different things, and confusing them is the classic first-project mistake.

The Four Tests of a Good First Build

Run every candidate from your queue through four tests. The one that passes all four is your answer.

• Attributable ROI. Can you point at the hours or the dollars afterward and say, plainly, this is what it saved? If the benefit is real but diffuse ("better decisions", "more insight"), it is a bad first build, because you cannot prove it worked and proof is the whole point.
• Low cultural pushback. Will the people affected be relieved, or threatened? Automating the task everyone hates is a gift. Automating the task someone's identity is built on is a fight you do not want for build number one.
• Reversibility. If it goes wrong, can you quietly turn it off with no damage done? First builds should be easy to unwind. Anything that writes irreversibly to a client-facing system fails this test.
• Data readiness. Is everything it needs already reachable, ideally read-only? If the project depends on connecting three systems you have not touched yet, that is a level-three data project wearing a first-build costume. Pick something that runs on data you already have.

Most teams have a candidate that passes all four sitting in plain sight, usually in finance or operations. Invoice reconciliation, debtor-chase drafting, report assembly, proposal first drafts, onboarding checklists. Unglamorous, measurable, and safe.

Back-Office Before Front-Office

There is a reason the map says back-office, and it is worth making explicit. The back office is where the work is repetitive enough to automate well, private enough that mistakes stay internal, and measurable enough to prove the win. The front office is the opposite on all three: bespoke, public, and hard to attribute.

This lines up with where agents are actually earning their keep in the real world. Anthropic's usage data across 1.2 million Cowork sessions puts business-process operations, reporting, reconciliation, checklist work, at 33.4% of all agent usage, far ahead of anything client-facing. The largest pool of proven, low-drama agent work is exactly the back-office category you should be fishing in for build number one. Start where the technology already has a track record, not where you wish it did.

Your flashy front-office ideas are not cancelled. They are build number six, funded and de-risked by the boring wins that came first.

How to Size the Win

Small enough to ship in weeks, big enough that someone notices. If your first build will take a quarter, it is too big; the runway you need arrives too late and the risk of it stalling is too high. If it saves twenty minutes a month, it is too small to change anyone's mind. Aim for something that lands in two to four weeks and removes a chunk of work a real person will feel the absence of.

And measure it properly before and after. Time the task by hand for a week, ship the agent, time it again. That before-and-after is the artefact you will use to unlock the next project, so treat the measurement as part of the build, not an afterthought. Hold the output to a defined quality bar too, because a first build that ships fast but produces work people have to redo is worse than no build at all. The method for that is in how to evaluate AI agents.

What to Avoid for Build Number One

Steer clear of four things specifically. Anything client-facing, because the stakes are wrong for a first attempt. Anything that needs to be right every single time, because early agents fail occasionally and your first build should tolerate that gracefully. Anything that depends on data you have not connected, because you will spend the whole project on plumbing and never reach the win. And anything politically loaded, where success threatens someone with influence, because you do not have the credibility yet to win that fight.

Save all four for later. By build number five or six you will have the proof, the tooling, and the standing to take them on. For the first one, boring and certain beats bold and contested every time.

Where to Start

Pull your audit queue back out and run the top few candidates through the four tests. The one that is attributable, low-friction, reversible, and ready is your first build, even if it is the least exciting thing on the list. Measure it before, ship it in weeks, measure it after, and tell the whole company the number. That number is the runway for everything else. If you have not done the audit yet, start there: how to run an AI audit.

If you want help picking the right first build and shipping it to a quality bar that holds up in front of your team, that is what our Deep Dive is for. We find the boring win hiding in your operation, and we help you land it.