Insurance operations teams spend thousands of hours every month trapped in a cycle of manual document sorting. From policy endorsements and regulatory statements to claim forms and legal declarations, the sheer volume of paperwork is staggering. For a human reviewer, the challenge is not just the volume, but the deceptive similarity of the documents. A policy endorsement and a regulatory statement often share the same legal terminology, the same font, and the same corporate branding, yet they trigger entirely different compliance workflows. When these documents are misclassified, the result is not just a clerical error; it is a potential regulatory violation or a critical delay in a customer's claim processing. This is the high-stakes environment where traditional automation has historically failed.
The Orchestration of Specialized Intelligence
To solve the classification bottleneck, developers implemented a multi-agent orchestration layer powered by the Strands Agents SDK. Rather than relying on a single, monolithic prompt to handle every document type, the system decomposes the problem into three distinct specialized roles: text reasoning, visual pattern recognition, and quality assurance. This architecture utilizes the agents as tools pattern, where a central orchestrator manages the flow of information and invokes specific agents based on the needs of the document.
The Validation Agent serves as the system's brain and primary orchestrator. Using the Strands Agents SDK, it coordinates the calls to the other specialized agents and synthesizes their findings. When the text-based analysis and the visual analysis produce conflicting results, the Validation Agent does not simply guess. It analyzes the discrepancy, weighs the confidence scores of each agent, and generates a final classification value. If the resulting confidence score falls below a specific threshold, the system flags the document for human review, ensuring that the automation never guesses on high-risk edge cases.
Supporting the orchestrator is the Document Analysis Agent, which leverages Claude Haiku 4.5 via Amazon Bedrock. This agent is tasked with the heavy lifting of legal language interpretation. It analyzes the semantic content, metadata, and linguistic nuances of the document to form a classification hypothesis. By utilizing structured output, the Document Analysis Agent ensures that its findings are returned in a consistent, machine-readable format that the Validation Agent can process without further parsing errors.
Complementing the linguistic analysis is the Vector Similarity Search Agent. This agent ignores the meaning of the words and focuses entirely on the geometry of the page. It uses Amazon Titan Multimodal Embeddings G1 to convert documents into high-dimensional vectors, which are then processed using the FAISS (Facebook AI Similarity Search) library. Specifically, the agent employs a `perform_vector_classification` function to analyze layout patterns, table structures, and formatting conventions. This allows the system to distinguish between two documents that use identical legal jargon but have different visual layouts, effectively solving the problem of linguistic overlap.
Breaking the Single-Model Accuracy Ceiling
The shift to a multi-agent architecture was driven by the failure of single-model approaches to handle the nuance of insurance paperwork. When tested against a rigorous benchmark of regulatory documents, the traditional pipeline combining Amazon Textract for OCR and Amazon Comprehend for entity recognition performed poorly, achieving only 25% accuracy. The primary failure point was the inability to distinguish between documents that shared a common vocabulary but served different legal purposes.
Even more advanced single-model solutions showed significant gaps. Amazon Bedrock Data Automation (BDA) improved the results considerably, reaching 70% accuracy. While BDA could handle a large portion of the statements and forms, it still misclassified roughly one-third of the total document volume. The failure was most prominent in the overlap between policy documents and regulatory statements, where the model struggled to decide if the document's identity was defined by its words or its structure.
The multi-agent system eliminated these errors entirely, achieving 100% accuracy across all document classes. The breakthrough came from the explicit separation of text reasoning and visual structure analysis. By treating the layout as a first-class data point via Titan Multimodal Embeddings and the text as a separate reasoning task via Claude Haiku 4.5, the system stopped trying to find a single answer and instead started cross-referencing two different types of evidence. The Validation Agent then acted as the final judge, ensuring that the final output was the result of a consensus between linguistic and visual intelligence.
From an operational standpoint, the choice of Claude Haiku 4.5 provided a critical balance between precision and performance. The model recorded an average processing time of 19.3 seconds per document with a 93% confidence level, making it viable for high-volume production environments. To further optimize for latency and availability, the implementation utilizes cross-region inference profiles. By adding geographic prefixes such as `us.`, `eu.`, or `ap.` to the model IDs, the system can route requests to the most available region, reducing the risk of throttling during peak loads.
For developers looking to implement similar systems, the performance delta provides a clear decision matrix. If the workload involves documents where the text is similar but the layout differs, the progression from Amazon Textract/Comprehend (25%) to BDA (70%) and finally to a multi-agent structure (100%) suggests that a hybrid approach is the only way to reach production-grade reliability. The full implementation details and logic are available in the project's GitHub repository.
The transition from single-model prompts to multi-agent orchestration transforms document classification from a probabilistic guess into a deterministic process.
