How accurate is AI bookkeeping? A method to measure it on your own books

Accuracy depends on the data. A software agency with twenty software subscriptions and a landlord will see near-perfect coding from day one. A building contractor with 140 suppliers, half of them invoicing in Arabic on hand-stamped paper, will see more corrections at first. A single headline percentage hides that spread.

It also hides what counts as 'right'. Reading the total correctly is one thing. Choosing the right expense account, the right VAT treatment and the right customer or project is harder. A vendor can report 99% on the first and say nothing about the rest. So we suggest you test what you actually care about.

Take one recent month and pull 100 documents: supplier bills, receipts and expense claims. Do not cherry-pick. Include the faded fuel receipt, the bilingual DEWA or SEWA bill, the foreign software invoice in USD and the handwritten delivery note from the small supplier in Deira. If your month has fewer than 100 documents, use them all.

Have someone who knows the business code the same documents by hand first, or use the coding your accountant already approved. That is your answer key.

Run the documents through HelloBooks and compare field by field. Score each separately so you know where the weak spots are.

• Extraction: supplier name, date, invoice number, net, VAT and total all read correctly.
• TRN: the supplier TRN captured where one is printed.
• Account: the expense or asset account matches the answer key.
• VAT treatment: standard-rated, zero-rated, exempt, out of scope or reverse charge, per line.
• Flags: did it flag the documents a careful person would query, such as a missing TRN?

Say the results are 97 of 100 extractions right, 88 of 100 accounts right and 93 of 100 VAT treatments right. The useful work is in the twelve account misses. In our experience they cluster. Often eight of them are one supplier the AI has never seen, coded to general expenses because nothing told it otherwise. Correct that supplier once and those eight become right next month.

Some misses are genuinely ambiguous. Is a AED 3,400 laptop for a new hire office equipment or a fixed asset? That depends on your capitalisation policy, which the AI cannot guess until you tell it. Write the policy down, set it as a rule, and that class of error disappears.

Run the same test on the following month. This is the number that matters, because AI bookkeeping is meant to learn your patterns. A good result is a clear improvement in account and VAT accuracy, with the remaining errors being new suppliers or truly unusual items. If the same supplier is wrong two months running after correction, raise it with us; that is a configuration issue worth fixing.

Accounting firms can run this test across several clients and compare. Firms that want custom categorisation rules for a sector, for example a standard mapping for all their restaurant clients, can have the HelloBooks team build it; ask via hello@hellobooks.ai.

Accuracy is only useful if it changes how you work. When the AI is right on 95 of 100 lines and tells you which five it is unsure about, you can review by exception instead of re-checking everything. Take a firm handling 1,200 client documents each month: if review time drops from two minutes a document to twenty seconds for confident items, that is roughly 30 hours back each month. Keep spot-checking a random 5% even when confidence is high; that is how you notice drift.