How Accurate Is AI Bookkeeping, Really?

If you are considering automated bookkeeping, you have probably run into two claims. Vendors say it handles the vast majority of routine work. Accountants say it cannot be trusted with judgment.

Both are true, and the interesting question is not the headline accuracy figure. It is what happens to the transactions the system gets wrong, because that is where the risk actually lives.

The number, with caveats

Published figures for automated transaction categorisation cluster somewhere in the high eighties to mid nineties percent. Roughly 80 percent of routine bookkeeping work is described as automatable across coding, transfer detection, invoice matching and posting.

Treat those numbers carefully. They are vendor-reported, measured on transactions the system has seen before, and heavily dependent on how clean your bank feed and your chart of accounts are. A business with consistent suppliers and a tidy chart will land at the top of that range. A business with messy historical data and forty ad-hoc suppliers will not.

So: most of your volume, most of the time, with a meaningful tail.

The metric that matters more

Here is what the accuracy percentage obscures. Automation is not just fairly accurate, it is perfectly consistent, and consistency is where manual bookkeeping actually fails.

A human bookkeeper handling your AWS invoice will code it correctly. Probably every time. But across a year, with holiday cover, a staff change, and forty other clients, the same supplier gets coded three different ways. Now your cloud costs are split across three accounts and your gross margin is wrong in a way nobody notices because each individual entry was defensible.

A system with vendor memory codes it identically every single time. If the rule is wrong, it is wrong consistently, which means you find it once and fix it everywhere. That is a far better failure mode than being scattered.

For most SMEs, consistency improves the books more than the accuracy figure suggests.

Where it is confidently wrong

The tail is not random. It concentrates in predictable places, and all of them are judgment calls.

Capital or operating. A large one-off purchase that should be capitalised and depreciated. The system sees an expense.

Revenue recognition. An annual contract invoiced upfront. The ledger sees cash. Correct treatment is deferred revenue released monthly. Get this wrong and your ARR is overstated, which is the single most checkable error in a diligence process.

Transfers dressed as transactions. Moving money between your own accounts, a director loan, a founder reimbursement. These look like income or expense and are neither.

Anything requiring context about your business. The payment to a supplier who is also a customer. The invoice that is actually a refund. The December spike that is one annual renewal rather than a trend.

The dangerous property is that automation is not uncertain about these. It codes them with the same confidence as everything else. There is no flag, no hesitation, nothing to draw your eye.

What this means practically

The realistic model is not automation instead of a person. It is automation doing volume and a person doing judgment, which is a better use of an accountant than data entry was.

Three things make that work.

A review step on anything unusual. Value thresholds, new suppliers, and anything the system has not seen before should surface rather than post silently. If your tool posts everything without ever asking, that is a warning sign rather than a feature.

A human on the month-end. Not re-keying. Looking at the output and asking whether it reflects the business. A person who knows you will spot a wrong number that reconciles perfectly.

Your accountant on anything with a filing attached. Tax positions, statutory accounts, SST or GST treatment. This needs someone licensed and accountable, and no accuracy percentage changes that.

The question to ask a vendor

Not what is your accuracy rate. Every answer will be a number between 90 and 99 and none of them are comparable.

Ask instead: what happens to a transaction you are not confident about. Does it post anyway, or does it surface for review. Can I see, in one place, everything that was auto-coded this month without human confirmation.

A tool that can show you that list is one you can audit. A tool that cannot is asking you to trust a number in a marketing page.

The honest summary

Automated bookkeeping is good enough that manual categorisation is no longer a sensible use of anyone's time, and its consistency will probably improve your books rather than degrade them.

It is not good enough to run unsupervised, and the errors it makes are concentrated in exactly the areas that matter most for investor reporting. That is not an argument against using it. It is an argument for knowing which 10 percent to look at.

See how categorisation and review work together