The Department of Veterans Affairs (VA) faces an unprecedented data and workflow challenge in processing claims, especially following the PACT Act, which dramatically expanded eligibility for toxic exposure benefits. Examining each claim requires reviewing documentation in the Veteran’s electronic folder (“eFolder”), which, on average, contains around 100 documents totaling roughly 1,300 pages of medical evidence, service history, exam results, and correspondence.

To verify connection and complete artifacts such as the Toxic Exposure Risk Activity (TERA) Memorandum, Veteran Service Representatives (VSRs) traditionally had to comb through massive numbers of unstructured files—PDFs, document scans, and handwritten notes. Completing a single TERA Memo could take 60 to 90 minutes, as each data point requires manual validation.

As PACT Act claims surged to thousands per day, the manual process became a bottleneck. The scale, speed, and accuracy required to serve Veterans exceeded what human reviewers could achieve with traditional tools. The VA needed a way to unlock the data trapped in billions of documents, transforming unstructured information into structured, searchable content that could power automation, increase accuracy, and accelerate decisions for Veterans.

Working in partnership with VA, the Booz Allen team introduced an AI-enabled Intelligent Document Processing (IDP) pipeline powered by Amazon Textract and layered mission-specific application logic on top. The pipeline ingests documents from the VA eFolder, now with over two billion Veterans' records, and uses Optical Character Recognition (OCR) and Machine Learning to extract structured, machine-readable text not just from PDFs and typed forms, but also from difficult sources like handwritten notes, clinical annotations, and scanned service records. That text is then indexed at scale into a searchable repository.

On top of that foundation, the team built “Smart Search,” which lets a VSR filter and search across a Veteran’s evidence by document content, and jump directly to the exact passages that matter for the specific task they’re performing instead of reading entire documents front to back. Crucially, this data backbone enables downstream automation for high-impact tasks.

The first flagship use case was TERA Memo Automation: instead of a VSR manually assembling exposure history, service location, relevant time windows, and medical indicators for every toxic exposure claim, the system can now pre-fill large portions of the memo by automatically pulling supporting evidence from the indexed record and aligning it with PACT Act policy around presumptive conditions.

Follow-on use cases are already in motion, including automated annotations of evidence, document summarization, and proactive claim checks that prevent delays by acting like a TurboTax®-style file review to identify common claim development issues before progressing the claim to the rating stage.

The impact for VA has been massive, both operationally and for Veterans. First, Smart Search fundamentally changes the review burden for a VSR. Instead of manually reading the equivalent of “War and Peace” for each claim, roughly 1,300 pages of evidence per Veteran on average, the reviewer can often narrow to about 10 documents and jump directly to the relevant sections inside those documents.

That is an order-of-magnitude reduction in cognitive load, and, in practical terms, means reviewers now get to the right evidence roughly 9x faster. Second, for PACT Act toxic exposure claims, the automated TERA Memo workflow identifies relevant evidence in mere seconds, as opposed to the manual workflow which could take upwards of an hour, significantly reducing TERA Memo completion times. The system is already capable of answering half the required TERA questions automatically and is expanding toward near-complete coverage by incorporating additional service data, including Department of War records.

At PACT Act volumes (5,000 to 7,000 TERA memos per day), those time savings convert directly into tens of thousands of labor hours saved, while preserving reviewer oversight and judgment. Just as importantly, quality and consistency are improving: the system doesn’t “forget” to check a tour location against an exposure policy or miss that a presumptive condition is covered. That means Veterans with serious, often delayed-onset toxic exposure conditions are connected to benefits faster, with fewer avoidable delays, and without forcing them to repeatedly resubmit evidence or wait months for follow-up letters and exams.

Beyond individual claims, the program created something even more strategic: a high-fidelity, production-scale data layer the VA can reuse for AI solutions. By extracting and structuring more than a billion pages of Veteran records, including handwritten clinician notes, the VA now has computable evidence it can use to drive decision support, triage missing documentation earlier in the intake process, surface likely next actions for a claim, and highlight risk areas like “missing required exam” or “missing evidence.” This data enablement step is normally the hardest part of AI transformation in government, and it’s often the part that doesn’t get prioritized because it’s not tech-flashy.

Here, it’s already live, and it’s operating at a national scale: at peak, the pipeline processed roughly three million documents per day, to the point where this VA workload became the largest global consumer of Amazon Textract. The result is a structural shift. Claims development, historically the longest and most labor-intensive phase of the benefits process, is being accelerated without lowering evidentiary standards, and VSRs report that the tools make them both faster and more accurate. That matters to Congress, it matters to oversight, but most of all, it matters to Veterans: faster, more consistent decisions on life-changing benefits, delivered at scale.