AWS describes a contract intelligence architecture that turns PDFs into verified fields and makes those results available through Amazon Quick dashboards and natural language questions. Its central design choice is to separate extraction from verification: one AI agent reads the contract, a second checks the extraction against the same document, and Amazon Textract resolves a specific class of disagreement about signatures.
For cloud architects, the useful pattern is the division of responsibility across document storage, agents, verification, a database, and analytics. Amazon Bedrock AgentCore hosts the agents; Amazon Quick provides the interface for asking questions about individual contracts and the collection. These are capabilities described in AWS’s architecture walkthrough, rather than independently benchmarked results.
From contract PDF to checked fields
The workflow begins when contract PDFs are stored in an Amazon Simple Storage Service (Amazon S3) bucket, triggering an automated processing pipeline. An extraction agent powered by the Claude Sonnet series reads the PDF and extracts eight key fields, each accompanied by a confidence score.
A separate verification agent, powered by Claude Haiku, independently reads the same PDF and verifies the extraction. That second document read is the architectural distinction: the system gives verification its own agent and access to the source document, instead of treating the initial extraction as the final result.
Confidence scores and a second reading serve different purposes. The scores accompany the extracted fields; verification checks the extraction against the contract. The supplied evidence does not establish how well those scores predict correctness or how much accuracy the second agent adds. Agreement between the agents should therefore be understood as a result of this checking process, rather than an independently demonstrated guarantee of accuracy.
A specific fallback for signature disputes
AWS describes a further step when the two agents disagree on signature detection: Amazon Textract provides what AWS calls a deterministic tiebreaker using computer vision.
The scope matters. This fallback is documented for signature detection. The evidence does not describe an equivalent resolution mechanism for every possible disagreement between extracted contract fields. Nor does it provide error rates for the signature check.
The architectural lesson is to identify a particular disputed task and give it a separate resolution path. Calling that path deterministic describes AWS’s stated tiebreaking mechanism; it does not establish that every resulting signature decision is correct. Architects evaluating the approach still need evidence about its behavior on their documents.
AgentCore runs the agents; Quick exposes the results
The extraction and verification agents use the Strands Agent SDK, which AWS describes as an open-source, model-driven framework for autonomous agents. They are deployed on AgentCore runtime, a capability of Amazon Bedrock AgentCore. According to AWS, AgentCore provides serverless hosting, automatic scaling, and session isolation for the AI components.
The system stores verified results in Amazon Aurora PostgreSQL. Users can see real-time pipeline status over WebSocket and query the data through embedded dashboards and a natural language chat agent. Within Amazon Quick, the Amazon QuickSight capability powers the dashboards, while natural language querying and chat handle plain-language questions.
This separates processing status from analytical access. The status channel shows progress through the pipeline; the dashboards and chat expose the resulting contract data. AWS describes support for both aggregate questions and questions about a single contract, connecting the document-processing workflow to two distinct ways of using its output.
What the architecture leaves open
AWS also describes Policy in Amazon Bedrock AgentCore as a way to define and enforce security controls around agent operations. That establishes a documented control capability, but the supplied evidence does not specify the policies applied to this contract workflow.
The evidence likewise supplies no measured extraction accuracy, end-to-end latency, operating cost, or comparison with a single-agent approach. It does not explain how disagreements outside signature detection are resolved. Those gaps limit what can be concluded about production performance and reliability.
The design nevertheless offers a concrete architecture to assess: extract fields, check them against the original PDF, route signature disputes to a specialized check, store the verified results, and expose them through analytics. Its value as guidance lies in those explicit stages. Whether they deliver sufficient accuracy and acceptable operating costs remains an evaluation question.
