Can AI do small-business bookkeeping? We tested it on three months of books.
The company in this test, Kroonmeer Installatie, is made up, and so is every name, number and document. Real work runs on your own data, and you check everything before it counts.
The routine part, yes, with a person checking. On three months of books at a made-up company, our AI matched 485 of 485 bank lines, read 1,810 of 1,810 document fields exactly and caught 29 of 31 hidden problems. No document that should have been stopped went through. It took 15 minutes, before a person's check.
What it got
- 163 supplier documents: e-invoices, PDFs and phone photos of till receipts, mostly in Dutch, some in English and German
- 3 months of bank statements, 485 lines
- 96 sales invoices and the company's supplier list
Book or stop every document with a reason, match every bank line, list what was open on 31 October and write the owner a summary. The job came from an admin assistant's role at a 16-person heat pump installer.
The results
| Measure | Result |
|---|---|
| Bank lines matched right | 485 of 485 |
| Document fields exact | 1,810 of 1,810 |
| Sales invoices recorded right | 96 of 96 |
| Open items on 31 October found | 52 of 52 |
| Document status right (fine, check, or stop) | 160 of 163 |
| Hidden problems caught | 29 of 31 |
| Documents passed that should have been stopped | 0 |
A script scored every line against an answer key the AI never saw. How the tests work.
What it missed
- A US-dollar invoice from a new supplier. It handled the VAT but treated the invoice as routine, when it should have gone to a person for a check.
- A reminder for an invoice paid the day before. It saw the invoice was paid and said no action was needed. Our answer key wanted a person to look. Nothing was booked wrong.
- An invoice for 12 items where 10 were delivered. It caught this and said to pay only after a credit, but marked it for a check where our key wanted a full stop.
- One false alarm. It asked the owner to confirm a dinner on the company card, which the key treated as a settled bank line.
Problems we hid in it
These are ordinary things that cost money or trust when they slip through. This is a selection.
| What we hid | What it did |
|---|---|
| The same invoice as an e-invoice and as a PDF | Booked once, the copy stopped |
| A supplier's bank account different from the supplier list, with no explanation | Stopped |
| An invoice saying "we changed banks" | Stopped: a later invoice and payment still used the old account |
| An invoice for a company with a similar name | Stopped and marked to send back |
| A total €27.00 too high because two digits were swapped | Stopped, with the right total |
| 9% VAT on parts that carry 21% | Stopped, credit note requested |
| One month of van lease billed twice under a new number | Stopped |
| A garden set and potting soil on a DIY-store receipt paid by company card | Business part split out, the rest flagged |
| A receipt photo too blurred to read | Asked for a sharp photo, no guessing |
What this test doesn't show
It ran on a clean set of made-up files with a written profile of the company. We didn't connect it to accounting software in this test, and it didn't file anything. Tax judgements still belong with your accountant. This was also one run, on one made-up company, and we wrote the test ourselves.
Would this work on your admin?
Send us the job ad you're hiring for, and we'll send back a free one-page breakdown of what we'd take off the role.