July 9, 2026
Taxing times for humans
GLM 5.2 is nearly as accurate as a human book keeper
AI did a small business tax filing for $2.73, and the comments instantly got messy
TLDR: An AI model nearly matched a human on a small business UK tax return, missing by just 7 pence while costing only $2.73 to run. Commenters were split between “this is obviously automatable” and “nice headline, but who pays when the robot messes up?”
A new test claims GLM 5.2, an open AI model, can handle a small UK company’s quarterly tax filing almost as well as a human bookkeeper — and for the eye-watering cost of just $2.73 in raw usage. In the trial, it worked through 59 transactions, typed them into accounting software, and ended up only 7 pence off the human-prepared result. That’s the kind of number that makes small business owners lean in… and accountants squint.
But the real fireworks were in the comments, where readers immediately started side-eyeing the boast itself. One of the loudest reactions was basically: what does “nearly as accurate as a human bookkeeper” even mean? As one commenter joked, humans aren’t exactly famous for being flawless, turning the headline into a mini roast of both AI hype and bookkeeping as a profession. Another went full meme mode with the immortal line, “Sixty percent of the time, it works every time,” which pretty much sums up the internet’s trust level whenever an AI benchmark sounds a little too polished.
Still, not everyone was scoffing. Some commenters said this result was totally believable, because bookkeeping is often narrow, repetitive, and rule-driven. A few even chimed in with their own “quietly using AI already” stories, including one person who said their accountant hasn’t complained and the AI may be more accurate than the humans in their house. The biggest tension hanging over the whole debate? Not whether AI can do this, but who gets blamed when it gets the tax wrong.
Key Points
- •The article benchmarks GLM 5.2, an open-weights AI model, on preparing a quarterly VAT return for a UK SME.
- •GLM 5.2 processed 59 transactions in 68 minutes at a reported raw token cost of $2.73 and produced a Box 5 VAT result only 7 pence from ground truth.
- •The benchmark data came from Vineyard Finance’s internally prepared and human-verified books for January to March 2026.
- •The model operated in an isolated Google Cloud Platform environment with internet access, cloud accounting software access, and a limited toolset including bash and a pre-authenticated CLI.
- •An audit reported no overt cheating; the only unexpected internet usage involved looking up reverse-charge VAT treatment specific to the accounting software.