AI for accountants: what it's good at, where it goes wrong, and what to train for

Everyone's using it. Almost nobody's trained.

Karbon's State of AI in Accounting 2026 report puts AI use among accounting professionals at 92%. The same report says 46% of firms are investing in AI training. Those two numbers sit a few pages apart in the same survey and nobody seems bothered by the gap.

The picture at firm level is thinner still. Thomson Reuters' 2026 AI in Professional Services report found organisation-wide AI use had reached 40%, up from 22% the year before, and only 18% of respondents knew their firm was tracking any return on it. So most of that 92% is people using ChatGPT on their own, for their own tasks, with nobody checking what it's doing to the work.

That matches what I see in practices and finance functions. The usage number is real. What it mostly measures is people drafting emails faster. Karbon's own data says 77% of AI use is communication tasks. Useful, but it's not the reason a partner signs off a budget.

Where AI earns its keep in a practice

The good uses share a shape: the accountant owns the number and the model handles the words around it.

• Client communication. Turning a technical position into an email the client will read and act on. This is the 77%, and it's a fair use of the tool.
• Long documents. A 60-page lease, a set of loan covenants, a board pack. Summarise it, extract the terms, then check the extraction against the source.
• First-pass commentary. Management accounts narrative, variance explanations, the notes nobody enjoys writing. The model drafts, the accountant corrects and owns every figure.
• Research with receipts. A question about a standard or a tax treatment, answered with the source alongside so you can check it. Never the answer on its own.
• Meeting notes and process documentation. The month-end steps that live in one person's head, finally written down.
• Spreadsheet help. Explaining a formula somebody else wrote, tracing an error, cleaning up a data export.

Karbon's figures say the average firm saves 21 hours a month per employee, and that firms which invest in training save more. I'd treat the exact number with some caution, since it comes from a survey run by a company that sells practice software. The shape of it is right, though. The time comes back from the reading and the writing, not from the arithmetic.

Where it goes wrong: the numbers

In June 2026 Wall Street Prep asked four AI tools to build a three-statement model for Apple from SEC filings and consensus estimates, to investment banking standards. Claude, with its Excel integration, scored 5.5 out of 10. Shortcut scored 5.9, Copilot 4.4, ChatGPT 2.5. For comparison, their top-tier analysts score 9.4 and their lower-level analysts 6.4.

The detail matters more than the ranking. Both of the top two tools invented historical data. None of them handled circularity or shares outstanding properly. Most used plugs instead of properly integrated statements. And every one of them had a first draft up in 15 to 25 minutes, against one to two hours for a person.

That's the whole problem in one test. Fast, plausible, nicely formatted, and wrong in places you'd only find by auditing every line. A junior who worked like that would be managed out within a quarter. The tool doesn't get managed out, because it's so fluent that people stop checking.

So the rule I teach is short. AI drafts the language around numbers. The accountant owns the numbers. Anything that crosses that line gets verified at source before it goes anywhere near a client.

Claude for Excel, specifically

The tool accountants ask me about most this year is Claude for Excel, Anthropic's add-in for Excel on Windows, Mac and the web. ICAEW's Excel Community published an early look at it in January 2026 and called it "a powerful illustration of how AI is moving from the periphery of analytical work into its core tools". That's about right, and so are the caveats that come with it.

What it does well, going by Anthropic's own documentation and by using it: it reads a whole workbook and answers questions with cell-level citations, it traces how a number was derived, it changes an assumption while keeping the formula relationships intact so the dependent cells recompute, and it finds the root cause of a #REF! or a #DIV/0. That last one is worth the licence on its own for anyone who inherits other people's models.

What Anthropic says not to use it for is just as useful. The documentation lists final client deliverables without human review, audit-critical calculations without verification, and models containing highly sensitive or regulated data without proper controls. It doesn't do data tables, macros or VBA. And there is a warning about prompt injection: a spreadsheet from outside the firm can carry hidden instructions, so you only point it at files you trust.

Put that next to the Wall Street Prep result and the use case writes itself. Let it explain, audit and flex the models you already have. Don't let it build a model unsupervised and then trust the history tab. Data handling is the other check. Anthropic says inputs and outputs are deleted within 30 days, chat history stays in your browser, and the add-in doesn't inherit your organisation's custom retention settings. Your data protection lead will want to read that page before anyone opens a client file in it.

Advisory, insolvency and restructuring are a different job

Most AI-for-accountants advice is written for compliance work: bookkeeping, tax returns, year-end. Advisory practices, and insolvency and restructuring teams in particular, have a different shape. The work is time-sensitive, it's billed on time and materials, the client data is sensitive and the process is regulated.

Two things follow. First, the driver isn't headcount. Nobody is trying to run a restructuring with fewer people. The prize is velocity: getting the business review, the cash forecast or the creditor analysis out in days rather than weeks, at the same quality, with room to take the next engagement. Second, the tool has to work inside the firm's approved environment and its confidentiality rules, or it won't get used at all. That's a readiness question before it's a training question, and I'd settle it before booking anyone into a room.

What AI training for accountants should cover

ICAEW, ACCA and AAT all run AI courses now, and they're a sensible place to start. What they can't do is train your people on your files, in your tools, against your controls. That part is the training. If you're comparing providers, this is what I'd insist on:

• Verification habits, practised. Every figure, citation and extracted term is unverified until checked at source. Drilled on real work rather than stated on a slide.
• Data boundaries, written down. What can go into which tool, what an enterprise agreement changes, what never leaves the firm. Karbon found 83% of accounting professionals are worried about data security. The fix is a rule people can follow.
• Prompts that become assets. The variance commentary prompt, the covenant extraction prompt, the client email prompt, saved and shared so the practice keeps them when the person who wrote them leaves.
• Excel work on your own models. Auditing and flexing real workbooks, with the failure cases from the Wall Street Prep test shown live rather than described.
• A build standard. Most practices already have custom GPTs and Claude Projects appearing on their own. A one-page standard for how they're documented, tested and owned stops that turning into a mess. We compared the two in Claude Projects vs custom GPTs.
• A measure. Baseline the hours on two or three recurring tasks before the training, then again a month after. It's the only honest answer to the partner who asks what it was for. The method is in how to measure AI ROI.

That's how we build AI training for accounting and finance teams at Fautons: on your reports and your spreadsheets, sized from a single team up to a whole firm. The sister piece on AI training for finance teams covers the in-house FP&A side, and Claude for data analysis goes deeper on the spreadsheet work. For neighbouring professions, AI for lawyers covers the legal side and AI for HR the people team.