Built Technologies launches AI document processor on AWS for real estate finance agents

Built Technologies, a real estate finance software provider that processes over $500B in real estate projects, has deployed an AI-powered document processing engine designed to automate a historically manual, document-heavy workflow. The company partnered with AWS Generative AI Innovation Center (GenAIIC) and AND Digital to build the solution using Amazon Bedrock for generative AI capabilities and the AWS Intelligent Document Processing Accelerator as the foundation.

Real estate finance transactions involve hundreds or thousands of pages of documentation produced by different parties in different formats at different stages—draw packages (collections of invoices, lien waivers, insurance certificates), loan agreements, appraisals, offering memorandums, and compliance forms. These documents are often inconsistent: some are standardized (like ACORD 25 insurance certificates or government forms), while others are highly variable and may contain nested tables, scanned pages, embedded images, handwritten annotations, custom layouts, and borrower-specific terminology. Built's previous system used OCR and traditional machine learning across 26 established processors for extraction, splitting, and classification, which worked for narrower use cases with predictable layouts and explicit fields. But to support more than 250 document types, handle millions of documents, and power agents that reason over documents rather than simply extracting text, the company needed a fundamentally different approach.

The new system works as a multi-stage pipeline orchestrated by AWS Step Functions. When a document is uploaded to an Amazon S3 bucket, an Amazon EventBridge event triggers a Queue Sender Lambda function, which records the event in DynamoDB and places a message on an Amazon SQS queue. A Queue Processor Lambda manages concurrency and, when capacity is available, starts a Step Functions execution. The state machine then runs processing stages in order: OCR, classification and splitting, extraction, assessment, and a final results step. Extraction runs inside a Step Functions Map state, which enables parallel processing of classified sections—so a 150-page draw package split into invoices, lien waivers, and insurance certificates processes each section concurrently rather than sequentially. Results are written to an S3 output bucket, and AWS AppSync delivers real-time updates to the user interface through GraphQL subscriptions. This parallel approach is why workflows that previously took days now finish in minutes.

Built required over 95 percent confidence in classification and extraction to support production use in financial and compliance-sensitive processes. The new system can classify, split, extract, evaluate, and reason over complex real estate documents, providing supporting evidence and routing ambiguous results to human reviewers. Built intentionally designed the solution as a horizontal capability rather than a single-purpose tool, so the same document intelligence infrastructure supports multiple agentic products: draw review agents that identify missing documents and flag exceptions, loan agreement agents that identify covenants and financial thresholds, insurance agents that validate coverage and endorsements, underwriting agents that summarize financial models, asset management agents that monitor reporting packages and surface portfolio risks, and compliance agents that inspect permits and regulatory documentation.

Real estate finance has traditionally relied on manual document review because the documents are inherently complex: long agreements with nested tables, scanned pages, handwritten annotations, and domain-specific language. Built's previous approach used optical character recognition (OCR) and traditional machine learning to extract data from 26 specific document types, which worked well for structured documents with predictable layouts and explicit fields. However, as the company expanded its AI roadmap across the full real estate lifecycle, this narrow approach hit its limits—the company needed to process more than 250 document types, handle millions of documents, and support agents that could reason about context and meaning, not just pull text from labeled fields.

The shift from extraction to understanding is exemplified by the covenant problem in loan agreements. A keyword search for "covenant" might miss obligations expressed in legal language across multiple sections; traditional extraction models struggle with implicit context and distributed information. An agentic workflow, by contrast, can identify relevant sections, infer which clauses represent covenants even when unlabeled, extract structured outputs with supporting evidence, and route low-confidence results to experts for human review. Built designed its new system as a horizontal capability—a shared foundation for multiple agents—rather than a single-purpose tool, so that draw review agents, loan agreement parsers, insurance validators, and compliance checkers all draw on the same core document intelligence infrastructure.

The technical foundation rests on Amazon Bedrock (for generative AI-powered reasoning) and the AWS Intelligent Document Processing Accelerator, orchestrated through AWS Step Functions. By splitting documents into classified sections and processing them in parallel via AWS Lambda, the system can complete in minutes what previously took days. This architectural choice—parallel extraction within a multi-stage pipeline—is a key reason the speed improvement is so dramatic.