How AWS Combines AI Agents and Analytics for Contract Intelligence

AWS just laid out a reference architecture for turning messy contract PDFs into verified, queryable data — and its smartest design choice isn't the AI, it's how the AI checks itself.

The two-agent pattern: extract, then verify

Instead of trusting one extraction pass, AWS splits the job across two agents:

• Agent 1 (Claude Sonnet) reads the PDF and pulls eight key contract fields, each with a confidence score.
• Agent 2 (Claude Haiku) independently re-reads the same PDF and verifies the extraction against the source document.

That second independent read is the whole point — verification gets its own agent and its own access to the document, instead of treating the first extraction as final.

When agents disagree, Textract breaks the tie

For signature detection specifically, AWS routes disputes to Amazon Textract, described as a "deterministic tiebreaker" via computer vision. The scope matters: the fallback is documented for signatures only — not every field disagreement gets an equivalent resolution path.

From checked fields to dashboards

The verified results land in Amazon Aurora PostgreSQL, with:

• Real-time pipeline status over WebSocket
• Amazon QuickSight dashboards for aggregate views
• A natural-language chat agent for plain-language questions about individual contracts or the whole collection

Agents run on Amazon Bedrock AgentCore (built with the Strands Agents SDK); Quick exposes everything to humans.

What's still open

The walkthrough gives no measured extraction accuracy, end-to-end latency, operating cost, or comparison against a single-agent baseline. And it doesn't say how non-signature disagreements are resolved. The design is a concrete architecture to evaluate — whether it delivers production-grade accuracy is still an open question.

Go deeper

Read the full analysis on Agentic Affairs: How AWS combines AI agents and analytics for contract intelligence

---

Originally published on Agentic Affairs (https://agentic-affairs.com)