AWS outlines a two-model system for extracting and checking contract data

AWS has described an AI contract platform that turns information in PDFs into structured records, then lets users ask both portfolio-wide and document-specific questions. Its key design choice is to use AI for extraction and checking, but a database for the sums and comparisons that retrieval-based chat can miss.

AWS AI Watch analysis

What happened

The proposed system processes contracts stored in Amazon S3, extracting eight fields with Claude Sonnet 4.6 and checking them with a separate Claude Haiku 4.5 model. Verified results go into Amazon Aurora PostgreSQL, where dashboards and a natural-language interface can answer questions such as total contract value, alongside lookups into an individual contract.

For disagreements about whether a contract is signed, the design calls on Amazon Textract to inspect the page visually. AWS describes the architecture in its contract intelligence platform walkthrough.

Why it matters

A question about one contract can be answered by finding a relevant passage. A question about the value or expiry date of hundreds of contracts needs the system to account for the whole portfolio. AWS’s approach makes that distinction concrete: extract fields into structured data for aggregation, while keeping the original documents available for clause-level questions.

AWS says its model-selection test used 20 contracts and 160 labelled field values. That is a small, directional evaluation, not evidence of accuracy across every contract format or business. The walkthrough gives no overall accuracy figure in the supplied material, so the architecture is more informative here than any claim of proven performance.

Our read

This is a useful pattern for document-heavy AI: let models handle reading, then use conventional data systems for arithmetic and comparisons. The second model and visual signature check are sensible attempts to catch errors, not guarantees that every extracted field is right. Teams considering the approach should test it against labelled examples from their own contracts before letting a tidy dashboard settle a costly question.

What to watch

• Whether AWS publishes detailed results from its 20-contract evaluation.
• How often the two models disagree, and how those cases are sent for human review.
• Whether the system performs reliably across different contract layouts and field types.
• How access policies work in practice for sensitive contract data. AWS has laid out a practical alternative to asking a chat tool to total up whichever contract passages it happens to retrieve. The full story looks at how extraction, verification and database queries fit together, and where the evidence remains limited. The useful question is whether a second model and a targeted visual check are enough, or whether people should review every extracted contract field. When would you trust this kind of system with a real portfolio?

Discussion spark: For contract data, is a second-model check plus targeted human review enough, or should every extracted field be checked by a person before it informs business decisions?

not affiliated with or endorsed by Amazon Web Services (AWS)