AI in accounting firms: what to automate first, and what not to

Everyone is using it. Almost nobody has rules.

Karbon's 2026 survey found 98% of firms using AI and only 21% with a policy for it. That gap is where the trouble lives. When there is no rule, each person picks their own tool, pastes in whatever helps, and nobody can say afterwards which client files went where.

So the first job is not picking software. It is a one-page list: which tools staff may use, what may go into them, and who checks the output. A firm that writes that page this month is already ahead of most of the profession.

Why this is about capacity, not novelty

The federal Job Bank outlook says Calgary had more openings than accountants to fill them from 2023 to 2025. In Intuit's US survey, 80% of firms said experienced people are hard to hire. When you cannot add a person, the only way to take on more clients is to take hours out of the work you already do.

That is the honest case for AI in a small practice. Not that it is clever, but that it hands back time you cannot hire for.

What to automate first

Pick work that follows rules, repeats every month, and already gets reviewed. That is also where firms have started: payroll, payables and receivables, and data entry topped Intuit's list.

In practice the order is usually: receipt and invoice capture first, because a reviewer already checks every coded line. Then client document chasing and reminders, because a missed reminder costs a deadline and a sent one costs nothing. Then first drafts of routine emails and engagement-letter summaries, which a person edits before anything goes out.

The test for each one is the same. If the AI gets it wrong, does someone see the error before the client does? If yes, it is a candidate. If no, it is not, yet.

What not to hand over

Tax positions, judgment calls, anything signed, and anything that goes to a client without a human reading it first. Language models produce confident text, and confident text is exactly what slips past a busy reviewer.

The Deloitte refund in Australia is the case to remember. A report containing a fabricated court quote and references to research that does not exist reached a government client and cost the firm part of an AU$440,000 fee. If it can get through a Big Four review process, it can get through a three-person office in March.

The privacy part most firms skip

Most AI tools process data outside Canada. Under Alberta's PIPA, when an organization uses a service provider outside Canada it has to tell the individual, and its written policy has to name the country and the purpose. The federal privacy commissioner's guidance adds that you stay responsible for information you hand to a third party for processing.

Your professional confidentiality obligations sit on top of that. Before a client file goes into any tool, check where the vendor stores and processes it, whether it trains on your data, and whether your engagement letter and privacy notice say so. It is a short exercise, and it is the one that matters if something leaks.

What it costs to start

The software is not the expensive part. Dext lists receipt capture at CA$31.50 a month for five users. Karbon's practice management runs $59 a user a month on its Team tier, so a five-person firm is around $295 a month billed annually. Microsoft's Copilot Business is $18 a user a month on a promotion through the end of 2026, on top of a Microsoft 365 plan.

The expensive part is setup: deciding the rules, connecting the tools to the way your firm already works, and checking the first month of output closely. Karbon's respondents say they save about an hour a day each, but that is self-reported. Measure your own before and after on one task, and only then roll it to the next.