AWS describes a contract platform that extracts and verifies document fields with AI agents, then combines portfolio analytics and document-level answers.
Amazon Web Services published an architecture for analyzing large collections of vendor contracts, pairing AI-based document extraction with structured database queries and natural-language access. The September 29, 2026 post describes a reference design, not a newly announced commercial product.
In the example, uploading contract PDFs to Amazon S3 starts an automated workflow. An extraction agent, using a Claude Sonnet-series model, identifies eight fields and assigns confidence scores. A second agent, using a Claude Haiku model, checks the same document independently. If the agents disagree about whether a contract has a signature, Amazon Textract supplies a computer-vision-based tie-breaker. Verified records are stored in Amazon Aurora PostgreSQL.
The design addresses a limitation AWS identifies in retrieval-based chat systems: retrieving a few relevant passages can help answer questions about one contract, but it may not provide the complete dataset needed to total values or compare expiration dates across a portfolio. Storing extracted fields in a database lets analytics handle those portfolio-wide calculations, while the original documents remain available for clause-level questions.
A React application presents the results through embedded Amazon Quick dashboards and a natural-language chat interface. Users can query structured records for aggregate information and consult the underlying documents for specific terms. The post says typical processing can take seconds per contract and describes a serverless setup intended to process many documents in parallel. It also includes real-time status updates and views for extraction confidence, field matches, and costs.
AWS estimates monthly costs of about $349 for a configuration handling 1,000 contracts, including $290 for Amazon Quick, $44 for Aurora, $12 for Bedrock, $2 for Textract, and less than $1 for several supporting services. The post puts AI inference at $0.014 per contract. Its model comparison used a hand-labeled sample of 20 contracts and 160 field values, which AWS describes as a small, directional test; it advises teams to evaluate models on their own data. For unresolved disagreements, the design calls for human review rather than automatic resolution.