Reviewing financial statements can become a long, repetitive task: comparing figures, looking for differences, and confirming that every amount matches its supporting documentation. Legora says GPT-6 Astra completed that process in minutes, reviewing 41 documents in a single run.
A review that used to take days
Legora is a work platform for legal professionals and other regulated industries. More than 100,000 professionals use it across over 1,800 legal departments and firms in more than 50 markets.
One of its most tedious processes is the financial-statement tie-out, a review in which every figure in preliminary accounts is compared with trial balances, consolidation schedules, and financial statements from the previous year.
In simple terms, it means checking that everything adds up. The problem is that doing this manually can take an entire afternoon or even several days, especially when dozens of documents and thousands of figures are involved.
With GPT-6 Astra, Legora’s agent reviewed all 41 documents in a single run. It compared each balance with its supporting documentation, identified differences, and left a detailed record of every check so the professional could review it.
It detected all four errors introduced into the accounts
To evaluate the system, Legora planted four deliberate errors in the financial statements. GPT-6 Astra found all four, including a £500,000 discrepancy hidden in a revenue note.
The company also says the model preserved all the checks that the previous version had completed correctly and added nearly 50 additional checks. The result was a faster, more complete first review, with clearer traceability.
AI handles the comparison and organizes the evidence. The final decision remains in the hands of the professional.
This point matters. The idea is not to let a model approve accounts on its own or issue a legal conclusion. The agent reduces mechanical work and presents the findings, while an expert interprets each result and decides what to do.
Nearly 40% better performance on this task
Legora evaluated GPT-6 Astra using its Legora Benchmark for Agentic Reasoning, known as BAR. This test measures performance on complete legal tasks based on real-world situations.
In the financial-statement review workflow, the company reported an improvement of nearly 40% compared with the previous model. Across all the benchmark’s tasks, the average improvement was approximately 3%.
The difference between these results helps explain something important: a model can improve significantly in one specific process without all its capabilities increasing by the same amount. In this case, the biggest benefit appeared in a task that requires reading many documents, maintaining context, and checking relationships between figures.
Beyond legal work
Legora is expanding its platform into areas such as auditing, tax, regulatory compliance, and risk management. These are fields where reviewing information accurately matters just as much as doing it quickly.
For a professional team, this could mean that a first pass that used to take hours is completed in minutes. But speed does not eliminate responsibility: any significant difference still needs context, judgment, and human review.
The lesson is less futuristic than it may seem. AI does not have to replace a specialist to be useful. In many jobs, its greatest value lies in moving through large volumes of information, detecting potential problems, and leaving a clear map for a person to make the right decision.
Original source
https://openai.com/index/legora-financial-statement-review-with-astra