From Days to Minutes: How Built Built a Document Intelligence Engine for Real Estate Finance

Built partnered with the AWS Generative AI Innovation Center (GenAIIC) and AWS Partner AND Digital to turn document understanding into a shared capability that now powers agents across the real estate finance lifecycle.

Real estate finance runs on documents. Draw packages, loan agreements, invoices, insurance certificates, and dozens more. Every one of them holds information that a lender or a stakeholder has to review, validate, and act on. The work is long, manual, and slow, and it sits directly in the path of the decisions that move money.

Built has processed more than $500B in real estate projects. To take the manual work out of that flow, we built an AI document processing engine on Amazon Bedrock and the AWS Intelligent Document Processing (IDP) Accelerator. Classification and extraction workflows that used to take 3 to 9 days now finish in minutes per package. That same engine is now the foundation for a new generation of agents launching across the real estate finance lifecycle.

Why Documents Are the Hard Part

A single transaction or asset can involve hundreds or thousands of pages, produced by different parties, in different formats, at different stages. Some documents are standardized, such as ACORD 25 certificates or government forms. Most are not. Offering memorandums, loan agreements, appraisals, Excel-based financial models, and plans and specifications arrive with nested tables, scanned pages, embedded images, inconsistent labels, legal language, handwritten annotations, and borrower- or lender-specific terminology.

Built already automated extraction across many of these document types using optical character recognition (OCR) and traditional machine learning, with 26 processors handling extraction, splitting, and classification. That approach works when fields are explicit and layouts are predictable. As Built expanded its AI roadmap across the lifecycle, the team needed something more capable: a single foundation that could support more than 250 document types, handle millions of documents, and power agents that reason over documents rather than only pull text out of them.

From Extraction to Understanding

Traditional OCR finds text and matches it to expected fields. That is fine for a loan amount, an invoice number, or a policy expiration date, where the value is labeled and sits near predictable text. It breaks down when the task requires judgment.

Take covenants in a loan agreement. They are rarely sitting in a table labeled “Covenants.” They appear across multiple sections, embedded in legal language, expressed as borrower obligations, reporting requirements, financial thresholds, default triggers, or remedies. A keyword search for “covenant” misses the substance.

An agentic workflow approaches the problem the way a person would. It identifies the relevant sections, reasons over definitions and obligations, distinguishes requirements from exceptions, extracts a structured result (the covenant name, requirement, threshold, frequency, responsible party, and consequence of breach), and points back to the source document so a reviewer can check the work. Anything low-confidence routes to a subject matter expert, and their corrections feed back into the schemas, prompts, and evaluations that improve the next document.

That is the shift, from document extraction to document understanding. The same pattern repeats across real estate finance, in questions like whether coverage satisfies a loan requirement, whether a draw package is complete, whether an appraisal supports underwriting, whether an offering memorandum carries risk indicators, and whether a portfolio document hides an exception. Answering any of them takes business context, not extraction alone.

One Capability, Many Agents

Built designed the engine as a horizontal capability, not a single-purpose tool. The first production use case is commercial construction loan draw packages, where borrowers submit large, variable, and time-sensitive collections of documents to request funding. That combination makes draw packages a strong first test for the engine.

The same classification, splitting, extraction, evaluation, and human-review capabilities support agents across the lifecycle, including the following:

• Draw review agents that classify package contents, flag missing documents, and extract invoice and lien waiver data.
• Loan agreement agents that surface covenants, reporting obligations, financial thresholds, and default provisions.
• Insurance agents that validate certificates, coverage limits, endorsements, exclusions, and expiration dates.
• Underwriting agents that summarize offering memorandums, appraisals, rent rolls, budgets, and financial models.
• Asset management agents that monitor reporting packages and surface portfolio-level risk.
• Compliance agents that inspect required forms, permits, inspection reports, and regulatory documentation.

Several of these already run in market on this same foundation: Built’s AI Draw Agent for construction draw review, and AI offering memorandum extraction for underwriting.

Every one of these depends on the same foundation: turning unstructured, inconsistent, high-volume documents into structured, validated, explainable intelligence. Because document processing is now a shared capability, Built can add new agents without rebuilding the pipeline each time.

How It Works, with AWS and AND Digital

Built built the engine with AWS GenAIIC and AND Digital on the AWS IDP Accelerator, using Amazon Bedrock for classification, splitting, schema generation, extraction, assessment, and reasoning. A document lands in Amazon S3, moves through a pipeline orchestrated by AWS Step Functions (OCR, classification and splitting, extraction, assessment, and rule validation), and returns real-time status to a custom review interface AND Digital built on React.

The design choices that carry most of the value include the following:

• Parallel extraction: When a 150-page draw package is split into invoices, lien waivers, insurance certificates, and a cover letter, each section is extracted concurrently. Total time is bounded by the longest section, not the sum of all of them. This is a large part of why days become minutes.
• Schema-driven accuracy: Each document type is defined by a schema, and the description written for each field becomes part of the extraction prompt. Telling the model where a value usually appears and what it is called guides it far better than a field name alone. Teams can generate a first-draft schema from sample documents, then refine it, which is what makes supporting more than 250 document types practical.
• The right model for the job: Straightforward documents such as standard invoices run on smaller, faster models like Amazon Nova Lite. Loan agreements and offering memorandums, which need deeper reasoning, run on larger models such as Anthropic’s Claude in Amazon Bedrock. The model is a configuration choice, not a rebuild.
• Confidence and human review: A separate assessment step scores every field from 0 to 1, explains the score, and marks where the evidence sits on the page. Built requires over 95% confidence in key production workflows. Anything below threshold goes to a reviewer working in a split-pane view, source image on one side, extracted fields on the other, with low-confidence values highlighted. Those corrections improve the system for the next package.
• Reasoning over documents: Extraction answers what a document contains. Many workflows also need to know whether it satisfies a requirement. A rule-validation step gathers the relevant facts, then reasons over that evidence to return a determination, compliant, non-compliant, or insufficient evidence, with citations back to the source. The rules are written as plain questions grouped by policy area, so the lending and compliance experts who own them can read and change them without an engineering ticket.

The Results

• Days to minutes: Classification and extraction workflows that took 3 to 9 days now finish in minutes per package.
• Built for complexity: The engine handles multi-hundred-page packages, nested tables, embedded images, scanned content, and non-standard layouts.
• Scale across the lifecycle: The architecture supports more than 250 document types, with production workloads scaling toward 20 million documents per month, 300,000 per week, and over 50,000 batch runs.
• Reuse across agents: One capability powers agents across construction lending, insurance, underwriting, asset management, compliance, and portfolio intelligence.
• Expert oversight, kept: Low-confidence and ambiguous results route to reviewers, so automation scales without giving up control.

For Built, document intelligence is the horizontal layer that lets agents understand real estate finance workflows, surface exceptions, accelerate decisions, and turn unstructured documents into trusted business actions. It is the foundation beneath the agentic operating paradigm taking shape across real estate lending, and the base the rest of the roadmap runs on.

Thanks to the AWS Generative AI Innovation Center (GenAIIC) and AND Digital for partnering with Built on this work. To learn more about using Amazon Bedrock for document processing, see Document processing with generative AI on AWS.