AI now handles the routine side of small business books well, from bank feeds to matching. Four respondents explain where it still goes wrong, and why the riskiest errors are the ones it is most sure about.

AI bookkeeping has moved from promise to daily use for many small firms. Bank feeds arrive, AI bookkeeping software suggests categories, and most transactions match without a person touching them. Still, owners and accountants want to know how far they can trust AI bookkeeping. The honest answer from our respondents is that routine work is largely solved, while judgement still needs a human. For a look at how accounting software firms are adding AI agents, see our report on Intuit Intelligence.

We asked one question: how much of small business bookkeeping can AI now handle reliably, and where do accountants still need to step in? We wanted founders of accounting and bookkeeping platforms, accountants and finance leads at SMEs. Time saved or error rate figures were welcome. Four people answered in this round, and we used all four. None shared hard error rates, so we’ve set their experience against the research.

What the research says about AI bookkeeping

The productivity gains from AI bookkeeping are real. A study by accounting professors at Stanford and MIT surveyed 277 accountants and drew on software data from 79 firms, the Journal of Accountancy reported. AI users shifted about 8.5% of their time from data entry to higher-value work, roughly 3.5 hours a week. They also closed the books 7.5 days sooner.

Accuracy over time is the weak spot. In Penrose Labs’ AccountingBench test, leading AI models closed a year of real company books. They came within 1% of a human accountant after the first month, Public Accountant reported. However, by the end, balances had drifted more than 15% off, about half a million dollars, as early errors went uncorrected and piled up. So AI bookkeeping can start well and still go wrong without review.

Let AI take the volume

Abhinav Gupta leads Profitjets, an accounting and bookkeeping firm that works with US small businesses.

“AI excels at high-volume, rule-based work. It has transformed importing bank transactions, suggesting categories and matching them to existing records, and it can even flag duplicates. That frees accountants for judgement work, such as making sure transactions are categorised in line with tax law. There are many edge cases where a transaction can be read more than one way, and bookkeeping involves a lot of judgement. AI can help with accruals if the bookkeeper fully understands the business, but it probably won’t be able to interpret the intent behind a transaction. AI will usually be faster than people at volume tasks. So the most reliable split is for AI to handle transactional tasks and bookkeepers to handle interpretation and judgement.”

Abhinav Gupta, Founder, Profitjets

His split matches the Stanford and MIT findings, where AI bookkeeping took over data entry rather than decisions. Also, his point on intent is the crux. A payment to a supplier can be stock, equipment or a deposit, and only someone who knows the business can tell. His original answer said AI will always beat bookkeepers at volume tasks. AccountingBench shows AI bookkeeping can still drift on volume work, so we’ve softened that line. He shares no time or error figures from Profitjets.

The confident errors are the risky ones

Ankit Sarawagi curates CFO Matrix, a finance resource, and reviews AI-coded books for clients.

“For regular items like a software subscription, a utility bill or a customer paying an invoice, it’s very reliable. But some situations cause problems. One client bought new equipment, and the tool recorded it as repair costs. It was very confident about it. That changes how it’s taxed, yet nothing in the books suggested a problem.

Accountants still need to decide whether something is an asset or an expense, handle owners’ own money, and make year-end adjustments. I usually check all the low-confidence items plus a sample of the high-confidence ones, because the errors that slip through tend to be the confident ones. I don’t have any data on time saved or error rates. If anyone does, check how they’ve defined an error.”

Ankit Sarawagi, Curator, CFO Matrix

His equipment example shows why AI bookkeeping still needs review. Equipment is usually a capital asset, while repairs are a running cost, so the two are taxed differently. Therefore, a confident miscode can change a tax bill without any warning in the ledger. His habit of sampling high-confidence items is the most practical AI bookkeeping control in this roundup. Still, the example comes from his own work, and we could not verify it.

Time saved is hard to prove

Conor Smart is a director at Arklavo, an ecommerce apparel business that uses AI for its own books.

“Accountants are still needed wherever judgement comes in. I run a small ecommerce business, and while AI can help categorise transactions, such as customer payments, refunds, supplier payments and software subscriptions, check that payouts match sales and flag unusual items, it still isn’t good enough to sign off accounts.

Its suggestions still need checking where transactions relate to each other over time, like refunds on a previous month’s sales. The same goes for anything that depends on events AI doesn’t know about, such as duties paid on behalf of customers or credits from suppliers, and for currency conversions. It will happily categorise things wrongly with complete confidence. For me, AI is great at the mechanical side of bookkeeping, but tax, year-end accounts and judgement calls still need a human in the loop.

I can’t give a percentage for time saved, because we never did this manually, so there’s nothing to compare against. I’d be suspicious of anyone quoting a percentage without having timed the tasks before. AI can help prepare accounts, but not sign them off.”

Conor Smart, Director, Arklavo

His examples show where AI bookkeeping struggles most: items that span months, depend on outside events or cross currencies. In addition, his warning on AI bookkeeping time-saved claims is fair. A firm that never timed its manual process has no baseline, which is why the Stanford and MIT study compared AI users with non-users instead. Arklavo is a small retailer, so his view is an owner’s rather than an accountant’s.

The case for going all in

Ashutosh Juneja is a chartered accountant and managing partner of Crystal Core Consultancy, an accounting and tax firm in Dubai.

“AI can now fully handle bookkeeping. I run my own business and that’s how we do it: our bookkeeping runs through AI, linked to WhatsApp and our other tools. In practice there’s no limit on the bookkeeping itself. The only limit is the AI tool’s token limit, not its ability to handle the bookkeeping.”

Ashutosh Juneja, Managing Partner, Crystal Core Consultancy

His view is the most confident in this roundup, and it comes from a qualified accountant running his own books. However, he gives no error rate and does not describe how anyone reviews the AI bookkeeping output. The AccountingBench results suggest that full automation without checks is where errors compound. Also, as a chartered accountant, he can likely spot a miscoded item himself. So his set-up may include more human judgement than “fully” implies.

What works for AI bookkeeping

The four answers point to a clear order of work. First, let AI bookkeeping tools handle bank feeds, routine coding, matching and duplicate checks. Next, review every low-confidence item, plus a sample of high-confidence ones, as Sarawagi does. Then keep a person on capital purchases, refunds across months, supplier credits, currency and tax treatment. Finally, time your process before and after, so any time-saved claim rests on a real baseline. Banks face the same balance in credit work, as our report on TD’s AI credit memos shows.

What to watch next

Rules on digital records will raise the stakes for AI bookkeeping. In the UK, Making Tax Digital for Income Tax applied from 6 April 2026 to sole traders and landlords with qualifying income over £50,000, GOV.UK confirms. It extends to those over £30,000 from April 2027 and £20,000 from April 2028. Also watch for AI bookkeeping providers publishing real error rates, which none of our respondents could share.

Above all, AI bookkeeping works best as a fast first draft, not a final answer. Firms that check the confident entries, not just the doubtful ones, will get the time savings without the surprises at year end.

FintechBits covers payments, banking and fintech developments for readers in the US and UK. This article is for information only and is not financial or legal advice. Views expressed are those of the FintechBits editorial team.