The post is a benchmark from Toot Books, a startup building AI bookkeeping software, claiming GLM 5.2 can prepare a UK VAT return nearly as well as a human. In the test, the model got the key refund number within 7 pence of a careful human-prepared filing. The author clarified that the benchmark covered only the accounting step after invoices had already been found and relevant context had been written down as notes. That limitation became the center of gravity. People did not dispute that a constrained VAT workflow is automatable. They disputed what that proves about real bookkeeping.
What emerged is a much narrower and more believable picture of where LLMs fit. They look useful for the messy intake layer that traditional finance software handles badly, like pulling data from emails, PDFs, odd invoice formats, and bank feeds, then drafting classifications or assembling a return for review. Several people already do exactly that with
FreeAgent,
Beancount, Claude Code, or custom scripts and say it saves real time. Even skeptics conceded that small businesses and solo operators can get value here, especially when the alternative is expensive accountants or brittle manual work.
The harder part is everything outside the benchmark. Invoices are often incomplete artifacts in an ongoing business relationship, not clean facts on a page. Charges live in footnotes, side agreements, mismatched purchase orders, late freight bills, and judgment calls about tax treatment or accounting periods. Large-company accounts payable already automates the deterministic three-way-match part without LLMs. Humans stay because exceptions drive real financial risk and sometimes require decisions that break the rules on purpose. That led many people to frame the right comparison as "
LLM plus controls plus reviewer" versus "junior bookkeeping labor," not "AI replaces accountants."
A second theme was accountability. Some objected that "nearly correct" is the wrong standard for tax filings. Others pushed back that taxes are not perfectly deterministic either, especially around estimates,
materiality, and ambiguous treatment, and that tax agencies usually distinguish mistakes from fraud. The practical consensus was less dramatic than either extreme. AI output is acceptable when it is boxed into software guardrails, reconciled against known transactions, and escalated on anything unclear. It is not acceptable as a free-ranging agent with authority to invent missing evidence, approve payments, or silently resolve exceptions. The benchmark convinced people that bookkeeping is moving into the same category as coding and support work: strong automation on narrow, well-bounded tasks, with the ugly edge cases and the legal exposure still sitting firmly on the human side.
toot-books.pages.dev