Architecting Agentic Document Intelligence: Built Technologies' Transition from OCR to LLM-Driven Workflows on AWS

The transition from rigid optical character recognition (OCR) pipelines to agentic, context-aware document processing represents a critical architectural shift for enterprise AI. According to a recent technical case study published on the AWS Machine Learning Blog, Built Technologies has successfully deployed a highly parallelized document intelligence engine on Amazon Bedrock. This implementation provides a blueprint for replacing siloed extraction models with a horizontal, reusable AI capability capable of reasoning over complex financial obligations.

The Boundary of Traditional Extraction in Financial Workflows

Built Technologies, a software provider processing over $500 billion in real estate projects, initially relied on 26 distinct processors utilizing OCR and traditional machine learning. This legacy architecture was effective for narrow use cases with explicit fields and predictable layouts, such as extracting invoice numbers or policy expiration dates. However, real estate finance is inherently fragmented and context-heavy. A single transaction can involve thousands of pages across offering memorandums, loan agreements, and nested financial models.

The primary failure mode of traditional extraction emerges when dealing with implicit or distributed logic. For example, loan covenants are rarely confined to a neatly labeled table. They are obligations buried within dense legal text, defined through cross-references, and expressed as financial thresholds or default triggers. A standard template-matching model might locate the keyword "covenant" but fail to comprehend the actual borrower obligation. To achieve the required 95 percent confidence threshold for production use in compliance-sensitive processes, Built required a system that could interpret documents in context, distinguish between requirements and exceptions, and infer relationships across multiple sections.

Architecting a Highly Parallelized Agentic Pipeline

To solve the dual challenges of document complexity and processing latency, Built partnered with the AWS Generative AI Innovation Center to deploy a multi-stage pipeline based on the AWS Intelligent Document Processing (IDP) Accelerator and Amazon Bedrock. The resulting architecture shifts the paradigm from sequential document reading to highly parallelized, agentic orchestration.

The ingestion pipeline begins when a document is uploaded to an Amazon S3 bucket, triggering an Amazon EventBridge event. To manage state and prevent throttling, a Queue Sender AWS Lambda function logs the event in an Amazon DynamoDB tracking table and routes a message to an Amazon SQS queue. Concurrency is strictly managed via DynamoDB atomic counters. Once capacity is confirmed, a Queue Processor Lambda initiates an AWS Step Functions state machine.

The critical architectural advantage lies in the Step Functions Map state. When a massive file-such as a 150-page commercial construction draw package containing a disorganized mix of invoices, lien waivers, and insurance certificates-enters the system, it is first classified and split. The Map state then invokes concurrent extraction processes for each individual section. By parallelizing the extraction and assessment stages across multiple Lambda functions, the total processing time is bounded by the longest individual section rather than the aggregate length of the document. This specific orchestration pattern reduces workflows that previously took days into operations completed in minutes, with real-time status updates pushed to a React-based frontend via AWS AppSync GraphQL subscriptions.

Strategic Implications: The Shift to Horizontal AI Capabilities

The most significant takeaway from Built's deployment is the strategic decision to design document intelligence as a horizontal capability rather than a vertical, single-use application. In enterprise environments, AI initiatives often suffer from technical debt when teams build bespoke pipelines for every new feature.

By centralizing ingestion, classification, schema management, and LLM extraction into a unified engine, Built created a foundational layer that serves multiple downstream agents. The exact same infrastructure used to review construction draw packages is reused by loan agreement agents identifying reporting obligations, underwriting agents summarizing appraisals, and compliance agents inspecting regulatory documentation. This modular approach allows product teams to launch new agentic workflows rapidly; they only need to define new schemas and prompts rather than engineering new OCR and routing pipelines from scratch. It demonstrates how enterprises can scale generative AI across an organization by treating document understanding as a core microservice.

Limitations and Unresolved Architectural Questions

While the architecture presents a robust model for document orchestration, the technical brief leaves several critical implementation details unresolved, limiting the ability to fully evaluate the system's operational overhead.

First, the specific foundation models utilized within Amazon Bedrock are not disclosed. The cost and latency profiles of processing 150-page documents vary wildly depending on whether the system relies on smaller, faster models (like Claude 3 Haiku or Llama 3) for classification, versus highly capable models (like Claude 3.5 Sonnet) for complex reasoning over legal covenants. Without model specifics, the unit economics of this pipeline remain opaque.

Second, the exact mechanism of the "assessment and rule validation" stage is ambiguous. Integrating deterministic business rules with probabilistic LLM outputs is a known friction point in agentic workflows. It is unclear if this validation occurs via traditional code executing against structured JSON outputs, or if it relies on a secondary LLM evaluation pass.

Finally, while the system features a human-in-the-loop (HITL) review interface to route low-confidence results to subject matter experts, the technical integration of this feedback loop is not detailed. Capturing corrections is straightforward; programmatically feeding those corrections back into prompt optimization, schema refinement, or model fine-tuning without manual engineering intervention is a complex challenge that the case study does not address.

Synthesis

Built Technologies' implementation on AWS illustrates a necessary evolution in enterprise automation: moving beyond the extraction of text to the contextual understanding of business obligations. By combining the reasoning capabilities of foundation models with the robust concurrency controls of AWS Step Functions and SQS, the architecture successfully handles the high-volume, high-variance demands of real estate finance. For technical leaders in document-heavy industries, this case study reinforces that successful generative AI deployments rely just as heavily on scalable orchestration and parallel processing as they do on the underlying language models.

Key Takeaways

• Built Technologies replaced 26 legacy OCR pipelines with a unified, agentic document intelligence engine on AWS Bedrock.
• The architecture utilizes AWS Step Functions Map state to process highly complex, multi-page financial documents in parallel, reducing processing time from days to minutes.
• The system achieves over 95% confidence in classification and extraction by shifting from rigid template matching to contextual LLM reasoning.
• Designing AI as a horizontal capability allows the same infrastructure to power diverse workflows, including underwriting, compliance, and loan agreement analysis.
• State tracking and concurrency are strictly managed using Amazon DynamoDB atomic counters and Amazon SQS to ensure reliable orchestration at scale.