AI news
AI model reviews
AI tools
How-to guides
Our staffAboutContact

Will AI take over accounting? What the evidence says in 2026

AI is taking over routine accounting tasks like data entry and reconciliation, not the profession. Here's what the job data, studies and benchmarks show.

Hand using a calculator on a desk
Photo: Towfiqu barbhuiya on Unsplash

Will AI take over accounting? No, but it's already taking over a big slice of the work accountants used to do by hand. Data entry, transaction coding and first-pass reconciliations are moving to software fast, while judgement, sign-off and client advice stay with people. If you're an accountant, I think the real risk isn't losing the profession; it's being stuck in the part of it that's shrinking.

That split shows up clearly in the US job projections. Accountant and auditor jobs are expected to grow about 5% between 2025 and 2035, from roughly 1.6 million to 1.67 million. Bookkeeping, accounting and auditing clerk jobs are projected to fall about 6% over the same decade. Same field, opposite directions.

Will AI take over accounting? The task-level picture

Here's the short version. Accounting isn't one job. It's a stack of tasks, and AI is very uneven across them.

At the bottom of the stack sits high-volume, rule-based work: pulling numbers off invoices, matching bank lines to ledger entries, suggesting the right account code. That's where software has always pushed, and large language models have made it cheaper to automate the messy edges (a PDF bill in a strange layout, a vague bank description). Further up, you've got work that needs context and accountability: deciding how to treat a lease, explaining to a client why their margin dropped, signing an audit opinion that a regulator can hold you to. Models can draft that work. They can't own it.

So the question "can AI take over accounting?" is really two questions. Can it do the tasks? Increasingly, for the routine ones. Can it carry the responsibility? Not under any licensing regime I know of, and I don't expect that to change soon.

TaskExposure to AIWho's likely doing it in five years
Data entry from invoices and receiptsHighSoftware, with spot checks
Transaction coding and bank reconciliationHighSoftware first pass, human review
Month-end closeMediumAccountant running and checking AI output
Audit evidence gatheringMediumAI agents, reviewed by experienced auditors
Judgement calls on accounting treatmentLowAccountant, possibly with AI research help
Signing opinions and tax filingsVery lowLicensed professional
Client advice and difficult conversationsLowAccountant

The exposure column is my own read of the evidence below, not an official rating.

Real examples from firms and software

The clearest example is EY. The firm has launched a multi-agent framework inside EY Canvas, its audit platform, which its 130,000 auditors use daily. Its goal is to have 100% of audit activities supported by agents by 2028. The first agents search and summarise documentation and handle admin; two more are coming, one to review auditors' work papers and one to match invoices and other records to audit samples.

That last one is exactly the kind of job a first-year auditor used to spend weeks on. Marc Jeschonneck, who leads EY's global assurance transformation, was candid that juniors will "need to have quite a level of experience" to review what those reconciliation agents produce. That's the awkward bit. The agents remove the repetitive work that used to teach people the job.

Hiring numbers back it up, at least in the UK. The Big Four cut graduate intake by as much as 29% over two years; KPMG's dropped from 1,399 to 942, Deloitte's by 18%, EY's by 11% and PwC's by 6%. Offshoring and a consulting slump played a part too, so it'd be wrong to pin all of that on AI. And the firms haven't stopped hiring: each of the Big Four plans to take on at least 1,000 junior people this year, with EY expecting 1,600.

On the small-business side, the change looks less dramatic and more everyday. Xero's assistant, JAX, answers plain-English questions like "What was my total income in the last six months?" and helps create invoices and quotes. QuickBooks has its own AI features in the same vein. None of this replaces an accountant at year-end; it mostly removes the reasons a business owner used to call one mid-month.

Benefits and limitations

The best evidence on benefits comes from a Stanford and MIT study that surveyed 277 accountants and analysed hundreds of thousands of transactions at 79 small and midsize firms. Accountants who used generative AI closed the month-end books 7.5 days sooner. They moved about 8.5% of their time (roughly 3.5 hours in a 40-hour week) from data entry to work like client communication and quality checks, and they reported 21% higher billable hours. The authors' own conclusion was that AI is "potentially augmenting, rather than replacing, professional accounting expertise."

That's a good result. It's also a study of AI as an assistant, with a human checking everything.

The weak spot: errors compound

Accounting has a property most text tasks don't: errors carry forward. A wrong balance in March is still wrong in April, and every later entry builds on it.

A benchmark called AccountingBench, built by Penrose Labs on a year of real data from a SaaS company, shows this neatly. The strongest models finished the first month's close with balances typically within 1% of the human CPAs' figures. Over the following months, every model piled up compounding errors, and overall balances drifted by more than 15%, about half a million dollars. The models sometimes booked transactions before understanding the source data, then struggled to fix their own mistakes.

Here's a worked example of why that matters. Say a small firm handles 40 clients with 500 bank transactions a month each, so 20,000 lines. A model that codes transactions correctly 97% of the time sounds great, but it still leaves 600 miscoded lines every month. If nobody catches them, they land in VAT returns, management accounts and tax filings. If someone has to check all 20,000 lines to find those 600, you've saved a lot less time than the headline accuracy suggests. The economics only work when the software flags the lines it's unsure about, so humans review maybe 5% instead of 100%.

Classification itself is harder than it looks. Digits, which sells AI bookkeeping software, benchmarked 19 frontier models on 17,792 real transactions from over 100 small businesses and found no general-purpose model cleared 70% accuracy. (Digits has an obvious interest in showing generic models fall short, so I'd take the exact figure with some salt.) Its fix was strict controls, like blocking any suggestion outside the client's chart of accounts.

And the market's been brutal to companies that promised full automation. Bench raised US$100 million, signed up nearly 40,000 customers, then shut its platform in December 2024. Honestly, that tells you more about how hard the last 10% of bookkeeping is than any demo does.

Practical takeaways for accountants and students

If you work in accounting, the useful move is to shift your weight up the task stack before your employer does it for you.

  • Get good at reviewing AI output: spotting anomalies, tracing a wrong balance back to its source, knowing when to distrust a confident answer.
  • Learn the tools your firm or clients already use, whether that's an audit platform's agents or the AI features in Xero and QuickBooks.
  • Pick up some data skills. Power BI, Alteryx and Python now show up in 63% of US accounting programmes' AI and analytics teaching.
  • Build the client side of the job. Explaining numbers, managing expectations and handling bad news are the parts software is weakest at.
  • If you're a bookkeeper or clerk, plan a route toward review, advisory or a professional qualification, because that category's the one projected to shrink.

For students, the picture's mixed but not grim. US accounting enrolment rose 12% in 2024, and there are still about 115,000 projected openings a year for accountants and auditors. Still, first jobs will probably ask more of you on day one. Robert Half's David Shogelola says juniors are already "spending less time on repetitive data entry and more time on reviewing outputs."

If you run a small business, AI tools can likely replace some of what you paid a bookkeeper for. I'd still keep a qualified person on year-end accounts and tax, at least until the error rates above look very different. Our guide to AI accounting software covers the main options, and if you're mostly worried about filing, can AI do my taxes? goes into that.

The next five years

My best guess is that three things happen at once. Routine processing keeps moving into software, and the number of pure data-entry roles keeps falling. Audit and tax firms keep hiring juniors, but fewer of them, and train them through simulated cases instead of years of repetition (EY has already said it'll train new hires this way). And the demand for people who can check, explain and take responsibility for AI-produced numbers grows, partly because regulators will want a named human behind every opinion.

The wild card is the training gap. Firms that froze graduate hiring after 2008 struggled for managers about five years later, and some are worried AI could repeat that. If the profession gets this wrong, the shortage won't be of software; it'll be of experienced reviewers.

That would be a strange outcome for a technology sold as a way to need fewer accountants.

Accounting isn't unusual here. Lawyers face a very similar split between automatable groundwork and accountable judgement, which we cover in will lawyers be replaced by AI?, and the wider pattern across jobs is in what jobs will AI replace.

FAQ

Will accounting be replaced by AI?

Not as a profession. AI is replacing specific tasks, mainly data entry, transaction coding and first-pass reconciliation. US projections still show accountant and auditor jobs growing about 5% over the next decade, while clerk roles shrink about 6%. Licensed work like audit opinions and tax sign-off still needs an accountable human, and I don't see that changing soon.

Will AI take over?

Not in the science-fiction sense. Today's AI systems do narrow tasks people point them at, and they still make errors that need checking. The realistic change is AI absorbing routine work inside jobs. Fears about AI harming people are a real policy debate, but a separate one.

Will AI take accounting jobs?

Some, yes, mostly at the entry level and in bookkeeping. UK Big Four graduate intakes fell by up to 29% over two years, partly because AI handles junior work. But firms still plan to hire thousands of juniors, and accountant roles overall are projected to grow. The jobs most at risk are the ones made almost entirely of repetitive processing. For more, see our piece on whether AI will replace accountants.

Is AI taking over accounting?

It's taking over parts of the workflow. EY wants agents supporting 100% of its audit activities by 2028, and small-business tools like Xero's JAX answer financial questions in plain English. But the accountant still reviews the output and signs off. A Stanford and MIT study found AI users closed their books 7.5 days faster and reported 21% more billable hours. To me, that looks like augmentation. Firms are getting more out of each accountant, and so far that's showing up as faster closes and more client work, not empty desks.

Can AI take over accounting?

Not reliably on its own, at least with current models. In the AccountingBench test, top models started within 1% of human CPAs but drifted more than 15% off after a few months of compounding errors. Accounting needs every number to be right, every month. AI can handle a lot of the routine work, but someone has to catch its mistakes and own the result.

Keep reading