GLM 5.2 is nearly as accurate as a human book keeper

	GLM 5.2 is nearly as accurate as a human book keeper(toot-books.pages.dev)
	208 points by adamkurkiewicz 17 hours ago \| 116 comments
	tl;dr: GLM 5.2, an open-weights model, prepared a quarterly UK VAT return for a small business (59 transactions) in 68 minutes at a token cost of $2.73, with the net VAT position off by just 7 pence from the human-prepared ground truth. Out of 354 scored checks, it failed 20, with only one serious error (misclassifying £10,000 in founder share capital); most other mistakes involved confusing zero-rated with tax-exempt VAT categories. The authors argue bookkeeping is becoming a solved problem and are building tooling (toot-books.com) to deploy this to UK SMEs.
	HN Discussion: ↓Benchmark is unfair because humans did broader work like sourcing invoices while the LLM got pre-packaged context ↓Legal liability and accountability concerns make LLM bookkeeping risky since responsibility falls on the user ↓Skeptical of trusting an unknown startup with sensitive financial data despite the results ↑Bookkeeping is a well-constrained problem that LLMs can handle reliably, confirming the article's thesis ↓'Essentially correct' isn't good enough for tax authorities, and the remaining errors are concerning