Amazon Bedrock's multi-agent system classifies insurance docs with Claude Haiku 4.5

Amazon Bedrock now demonstrates a production-ready approach to high-stakes document classification using a multi-agent architecture built on the Strands Agents SDK. Aimed at insurance companies that process thousands of daily documents—policies, affidavits, endorsements, and regulatory forms—the system tackles a common pain point: documents that look similar but serve different legal purposes. Misclassifying these can lead to compliance violations or claims delays, and traditional single-model classifiers fall short on such edge cases.

To solve this, the solution deploys three specialized agents. The Document Analysis Agent uses Anthropic's Claude Haiku 4.5 on Amazon Bedrock to interpret legal language and extract semantic context. The Vector Similarity Search Agent employs Amazon Titan Multimodal Embeddings G1 to convert documents into high-dimensional vectors, capturing visual and structural cues that text models miss. A Validation Agent orchestrates both, compares their outputs, resolves disagreements, assigns confidence scores, and flags ambiguous cases for human review—all without custom orchestration code. This tools-based pattern delivers modularity, transparency, and flexibility, allowing agents to be improved or added independently. By fusing textual reasoning with visual pattern recognition, the system achieves better accuracy than any single-model baseline, giving insurers a scalable path to compliance-ready document processing.