Your first AI project should be the dullest job in the firm

"Where do I start?"

Billie Mcloughlin hears it from most of the firms she works with. She runs Aiccountancy, trains accountants on Microsoft and Claude, and spent over a decade in practice before that. Behind the question, she said at our AI agents panel on 1 October, is usually a fear that everyone else has already transformed their practice. Mostly, they haven't. It has only recently become possible. (watch)

The panel's answer was consistent, and deliberately unexciting. Start with the admin nobody wants to do.

Boring is the point

Billie's example: a piece of client information is missing from one of your systems. An assistant connected to that system and to Outlook spots the gap and drafts the email chasing the client for it, say their UTR. (watch)

"Those are the time savings. They're not glamorous. They're not sexy. They're not sparkly." But they are where firms should start, she said, because they compound.

Kevin Lord of Firmcheck, who hosted, did the sums. Save five minutes per client and it doesn't sound like much. Across 100 clients, it's over eight hours, roughly a working day, every time that job comes round.

Four steps, in order

Billie's method works wherever a firm is starting from. (watch)

Audit first. Find out which tools people in the firm actually use and how much time goes into each. Billie runs this as an amnesty, so people own up to the AI they're already using without permission. That is a security question as much as a productivity one.

Then prioritise. Don't try to fix everything. Pick the tasks that take a lot of time and carry little risk. Onboarding admin is the classic case: slow, repetitive, low value, and low risk to get slightly wrong as long as a person checks it.

Then build. Set up the assistant, the connections and the instructions for those few tasks only.

Then embed. Make it the way the job is done, not a pilot someone remembers to run. The value of the first project is that it becomes a habit.

Her warning was against doing it the other way round. Plenty of accountants tell her they don't want to automate, while describing AI writing their emails. "Stop trying to automate your entire firm," she said. Do the basic admin first.

Write it down before you automate it

Marcus Batson, Head of Customer Success at Sodium, made the point that decides whether the four steps work. The basics are the same whether you use AI or not: what your processes are, what kinds of clients you work with, where the data lives. If that lives in someone's head, the firm is hard to scale in any direction, and an assistant has nothing to follow. (watch)

The survey data agrees with him. In Xero's May 2026 research, 87% of UK practices using AI in their daily work said their processes were well documented and regularly updated. Among firms with no plans to adopt AI, the figure was 18%.

Writing a process down also tells you something before you automate anything. Once it's visible, Marcus said, you can see whether that job should be scaled with a system or with a person. Kept in someone's head, you can't tell.

Do it once, then make it repeatable

The step from a one-off to a habit is smaller than most firms expect. In the session, Kevin asked Claude in a single prompt to bring Firmcheck up to date with the client list in Sodium and run any outstanding ID checks. Having done it once, he said, he could save it as a skill, a stored set of instructions the assistant can run on request, and have it run every Friday afternoon. (watch)

Marcus already works this way. A skill he built handles his post-call wrap-up: summarise the meeting, add the notes, update the record. What used to be a chunk of admin after every call is now low-touch. (watch)

Celso Pinto, who runs Connector Central, offered a simple test of whether the habit has stuck. When someone wants to know what to work on this morning, do they start in the assistant, which can see their email, the practice management system, and anything blocked in Firmcheck? Or do they still open each system in turn? (watch)

This week

Ask everyone in the firm one question, and make clear there's no wrong answer: what job did you do last week that you'd most like never to do again? Pick the one that comes up most often and costs least if it goes slightly wrong. Write down, step by step, how it's done today. That document is your first AI project.

The full panel recording is here.