GPT-6 Astra: Legora Financial Statement Review Performance
OpenAI has announced a case study involving Legora, an agentic operating system for legal and professional work, using GPT-6 Astra to automate the financial-statement tie-out process. This integration allows legal professionals to process dozens of complex financial documents in minutes rather than days, significantly reducing the manual effort required for figure verification.
High-Scale Processing of Complex Financial Context
GPT-6 Astra enables Legora's Agent to complete a financial-statement tie-out across 41 documents in a single run. The system checks every balance against its supporting schedule, identifies discrepancies in amounts, and records each check to provide a granular record for professional review.
According to Percevale Perks, a Legal Engineer at Legora, the primary advancement is the model's processing power and its ability to ingest and digest a large volume of complex information and line items simultaneously.
Performance Gains via Legora Benchmark for Agentic Reasoning (BAR)
Legora measured the performance of GPT-6 Astra using the Legora Benchmark for Agentic Reasoning (BAR), which utilizes real-world end-to-end legal tasks. The results indicate a significant uplift in the financial-statement workflow:
- Workflow-Specific Improvement: GPT-6 Astra improved performance by nearly 40% over the previous model for the financial-statement tie-out task.
- General BAR Performance: Across all tasks within the BAR, the model showed an average improvement of approximately 3%.
Accuracy and Reliability in Financial Auditing
In practical testing, GPT-6 Astra demonstrated increased accuracy, completeness, and reliability. In a controlled test where Legora planted four errors in the accounts, the model successfully identified all four, including a £500,000 gap in the revenue note.
Furthermore, the model maintained all the correct checks performed by the previous model while completing approximately 50 additional checks. This ensures a more comprehensive first pass of the data, while the final judgment and decision-making remain with the human legal expert.
Human-in-the-Loop Integration
Legora employs a human-in-the-loop approach where the AI agent handles the exhaustive comparison of data, but the expert remains responsible for the final judgment call on each result. This operational model is being used to extend Legora's platform into audit, tax, compliance, and risk management.