GLM 5.2 is nearly as accurate as a human book keeper

4 hours ago

GLM 5.2 prepared a nearly perfect quarterly VAT return for a UK SME at a raw token cost of $2.73, compared to typical human accountant fees of £750–£2,100 per quarter.
The model processed 59 transactions in 68 minutes, with the net VAT position off by only 7 pence (≈10 US cents) relative to ground truth.
Scoring involved 6 criteria per transaction: type, category, VAT treatment, VAT amount, reverse-charge VAT, and receipt attachment, with 20 failures out of 354 checks.
A serious error was misclassifying £10,000 in founder shares as 'Capital Account' instead of 'Unpaid Shares', potentially causing legal and audit issues.
Other errors included confusing 'zero-rated' and 'tax-exempt' VAT categories (14 transactions) and mishandling split currency transactions (3 transactions).
The model excelled at correct account classification, accurate invoice attachment, and disambiguating tricky transactions like same-day same-vendor entries.
Book-keeping is becoming a solved problem, with focus now on scaffolding to deploy AI capabilities for UK SMEs, such as through toot-books.com.

Hasty Briefsbeta