An AI bookkeeper earns trust by saying when it is unsure

Almost every company that sells accounting software to small businesses now promises that artificial intelligence will do the books. The promise is usually expressed as an automation rate: the share of receipts, bills and bank transactions that the software reads and categorises without a person touching them. Ninety per cent is common in marketing material. Some vendors claim more.

I run product at a company that builds these features, and I understand why the number is attractive. It is easy to measure, easy to compare and easy to print on a sales page. It also describes the wrong thing. The owner of a hardware shop in Seremban does not want her books to be ninety per cent automatic. She wants them to be right, and when they are not right she wants to know which tenth to check before her accountant files anything in her name.

The automation rate is the wrong headline

Consider what an automation rate hides. A system that posts every transaction automatically reports an automation rate of one hundred per cent, whatever its accuracy. A system that posts eighty per cent and asks about the rest looks worse on the sales page, even if the items it posts are almost always right and the items it asks about are exactly the ones it would have got wrong. The headline rewards the product that is more confident, not the product that is more correct.

The cost of a wrong categorisation also falls unevenly. Filing a fuel receipt as office supplies changes nothing that matters. Treating a customer's deposit as revenue in the wrong period, or a loan from a relative as income, can change the tax a business owes and the story its accounts tell a lender. A small business rarely has a finance team to catch such errors. Its owner finds out months later, from an accountant, a letter from the tax authority or a declined loan application.

At Rimbunara we read about 3.1 million receipts and supplier bills a month, photographed on phones in shops, kitchens and site offices across four countries, in four languages and several scripts. The photographs are often taken in poor light, folded or half covered by a thumb. Our models are good at this work and improve every quarter. They are not good enough, and no model I know of is good enough, to decide silently what every one of those documents means.

Confidence the owner can see

The rule we adopted in 2024 is simple to state. Every categorisation the model makes carries a confidence score, and only those above a threshold are posted automatically. The rest wait in a short queue for the owner or her bookkeeper to confirm with one tap, with the model's suggestion already filled in. Today 71 per cent of items are posted automatically and 29 per cent wait. Thresholds are set per category, so that anything which changes tax or borrowing faces a stricter test than routine expenses.

We check the posted items as well. Each month, accountants in our partner network audit a random sample of automatically posted transactions in each country. The error rate on those items is currently 1.2 per cent, and we show it to customers inside the product, next to the share of items waiting for them. When the error rate in a category rises, we raise its threshold before we look for a better model. That choice lowers our automation rate. I would make it again.

We also tell the owner why an item is waiting. A receipt may be held because the total was unreadable, because the supplier is new, or because the same amount appears twice in a week. Owners told our researchers that a reason makes the queue feel like a colleague's question rather than a fault in the software. The share of waiting items confirmed within a week rose from 58 per cent to 81 per cent after we added reasons, and that matters, because an unconfirmed item is a gap in the books.

There was internal resistance, and some of it was fair. Our sales team pointed out that competitors' claims made our product look slower in demonstrations. One of our largest accounting-firm partners wanted everything posted automatically and corrected later by its own staff, because that suited its workflow. We now let accounting firms choose a lower threshold for clients they manage, with the firm's name recorded against the setting, while owners without an accountant keep the stricter default.

What the assistant will not answer

The same principle applies more sharply to the assistant we launched in 2024, which answers questions in English, Malay, Indonesian and Filipino about a business's own figures. A language model can produce a fluent answer to almost any question, including questions it has no business answering. Fluency is the risk. An owner who asks whether a payment to her brother is deductible will receive a confident paragraph from any general-purpose model. It may even be correct in one of our four countries.

Our assistant answers questions about what is in the books: why profit fell last month, which customers are paying late, what was spent on transport this quarter. Every answer lists the transactions it used, so the owner can open them. It will not give tax or legal advice. When a question needs a judgement about tax treatment, the assistant says so plainly, explains what an accountant would need to know and offers to send the question, with the relevant figures, to the customer's accountant or to a firm in our partner network. About 14 per cent of questions end in that hand-off.

We review a sample of conversations every week, in each language, with reviewers who speak it. The commonest failure is not a dangerous answer. It is an answer that is technically right but uses a figure the owner did not mean, such as sales including tax when she wanted them without. The assistant now asks a clarifying question when a request could mean two things. That makes it slightly slower and noticeably more useful.

None of this settles how machine learning should behave in small-business finance. Our practice is documented against Singapore's Model AI Governance Framework, and the board's Audit and Risk Committee reviews our model risk twice a year, but frameworks describe principles rather than thresholds. Thresholds are product decisions, and somebody has to own them.

My suggestion to other product leaders, and to the directors who oversee them, is to ask for three numbers instead of one: the share of items the system posts on its own, the measured error rate on those items, and the share of questions it declines to answer. A product that can report all three is being honest about what it knows. A product that reports only the first is asking to be trusted without saying what for.