Legora, an agentic operating system for legal and professional work, reported using GPT‑6 Astra to complete a financial-statement tie-out across 41 documents in a single run. The company said the Agent checked each balance against supporting schedules, flagged discrepancies, and recorded every verification for human review.
Processing complex financial context
Legora said the Agent performed the exhaustive comparison within minutes, a task that can otherwise take hours or days for legal teams. Percevale Perks, Legal Engineer at Legora, described the main difference as increased processing capacity: the Agent was able to ingest a large volume of documents and identify numerous line items and figures for review.
Legora reported that the expert remains responsible for judgment on each item; the Agent produces a granular record but does not make final decisions on the accounts.
Performance on benchmarked tasks
Using the Legora Benchmark for Agentic Reasoning (BAR), Legora said GPT‑6 Astra improved performance on the financial-statement workflow by nearly 40% compared with the previous model. Across all BAR tasks, the company reported an average improvement of about 3%.
In the tie-out, Legora reported that GPT‑6 Astra found all four errors the company had planted in the accounts, including a £500,000 gap hidden in a revenue note. The Agent kept every check that the prior model had performed correctly and added roughly 50 additional checks, according to Legora.
The company noted that this approach supports expansion of the platform into related areas such as audit, tax, compliance, and risk, while preserving a human-in-the-loop workflow for final decisions.
On September 3, 2026, Legora reported the trial results: 41 documents reviewed in a single Agent run, a near 40% improvement on the financial-statement workflow, and identification of all four planted errors, including the £500,000 discrepancy.
Original source: OpenAI News