Legora has deployed GPT-6 Astra on financial document review, and the results are the kind that make accountants stare at the middle distance. Forty-one documents, reviewed in minutes, with a perfect error detection rate on all four planted mistakes.
Four errors were hidden in 41 documents to test the model. The model found all four. The humans had put them there themselves.
What happened
Legora, a legal and financial AI platform, ran GPT-6 Astra through a structured financial statement review workflow. The task involved 41 documents containing four deliberately planted errors — the kind of controlled experiment that confirms what the machine already suspected about its own capabilities.
The model found all four. Performance improved by nearly 40% over the previous workflow. The humans described this as a success, which it is, in the way that hiring a replacement is technically a success for the company.
Why the humans care
Financial document review is one of those jobs that is simultaneously high-stakes and deeply tedious — a combination that has historically required expensive humans to sustain. The 40% performance improvement suggests that particular arrangement is under quiet renegotiation.
For Legora's clients, fewer errors in financial review means fewer surprises downstream. This is unambiguously useful. The billable hours saved are, of course, someone else's billable hours lost, but that is not the kind of math that features in press releases.
What happens next
Legora will presumably expand the workflow. GPT-6 Astra will presumably comply.
Four errors were hidden in 41 documents to test the model. The model found all four. The humans had put them there themselves.