Guardoc Health says it processes more than one million clinical documents each day using Amazon Nova models. The system handles PDFs, handwriting, checkboxes, stamps and medication tables for long-term care providers, where a small extraction error can affect treatment, reimbursement or an audit.
Its architecture is notable because it reserves expensive multimodal reasoning for the pages that cheaper stages cannot resolve. The reported outcomes are substantial, though their baselines and methodology remain unpublished.
The pipeline narrows evidence before multimodal review
Amazon Textract first extracts printed text and document structure. Guardoc then divides the output along clinical boundaries, keeping sections such as medication lists intact instead of cutting by an arbitrary character limit.
Titan Text Embeddings V2 represents each chunk, which is stored in DynamoDB and partitioned by patient. That separation prevents retrieval from crossing patient boundaries—an essential control when records are processed at scale.
A custom filter uses document type and recency before similarity search selects material relevant to a classification question. Only page references move at this stage, reducing data transfer. Nova 2 Lite performs a text-based pass to remove obvious non-matches.
The remaining pages reach Nova Pro as raw PDF bytes. The model can consider layout, handwriting, signatures and stamps before returning a classification. Cheap tools therefore handle volume, while the more costly model is used where visual context affects the decision.
This tiering is a practical pattern for document AI. It can reduce cost and latency without forcing a small model to resolve ambiguous evidence. It also creates several failure points: faulty OCR, incorrect chunking or retrieval can prevent the decisive page from ever reaching Nova Pro.
Handwriting and layout contain the highest-risk details
Guardoc says earlier versions most often missed physician attestations on prior-authorisation forms and patient-reported symptoms. A handwritten note may override a printed checkbox, while a symptom may appear nowhere else in the chart.
Medication records are equally varied. Names, doses, routes and frequency can appear in tables, prose, handwritten additions or degraded fax scans. Guardoc combines Textract output with the original PDF so Nova Pro can resolve wrapped columns and non-standard formats.
The system uses retrieval-augmented generation to ground condition classification in a patient’s own documents. Every output should retain the page, region and source file supporting it. A reviewer needs to distinguish “not found” from “not present” because missing retrieval is not clinical absence.
High-risk fields should have confidence thresholds and mandatory review. Staff corrections need to flow into evaluation sets, with special attention to handwriting quality, document type and facility-specific forms. Patient partitions and access controls should be tested rather than assumed from architecture diagrams.
Automation can focus staff on complex cases, but it should not make documents less visible. The original record remains the clinical authority when model output and source evidence conflict.
Reported operational gains need reproducible evidence
Guardoc reports a 46% reduction in documentation errors, 70% fewer audit fines and more than $400,000 in annual return for one facility. It has not published the baseline period or calculation method.
Across two facilities and 200 patients, the company says the system prompted 847 corrections, flagged 86 reimbursement issues and coincided with a 74% reduction in hospital transfers per 100 admissions. Another case across seven facilities and 1,618 residents identified 10,612 issues.
These figures do not establish that document processing caused the reduction in transfers. Staffing, care pathways and patient mix may also have changed. Independent evaluation should compare facilities over time, define each error category and report false positives as well as corrected omissions.
At one million documents daily, even a 1% failure rate creates thousands of problematic records. Aggregate accuracy can therefore look strong while producing unacceptable harm in a rare but critical category.
Guardoc’s strongest contribution is its cost-aware escalation design: use simple extraction first, preserve patient boundaries and invoke multimodal reasoning only for difficult pages. The clinical standard requires more. Every decision must remain traceable to the document, uncertain cases must reach people, and impressive operational claims must be tested against transparent baselines. Scale magnifies both the savings and the mistakes.
