Guardoc Health Cuts Documentation Errors 46% Using Amazon Nova Models on Bedrock

Guardoc Health, a healthcare documentation company serving skilled nursing facilities and assisted living centers, has deployed Amazon Nova foundation models through Amazon Bedrock to process clinical documents at volumes exceeding 1 million per day, according to a joint post published on the AWS Machine Learning blog.

The company reports a 46 percent reduction in documentation errors, a 70 percent decrease in audit fines, and more than $400,000 in annual return on investment for a single facility. These figures are Guardoc's own claims and have not been independently verified.

The problem: unstructured medical records

Medical documents arrive in inconsistent formats — multi-page PDFs with handwritten annotations, prior authorization forms with checkbox fields, medication lists embedded in tables or scanned images, and intake forms mixing typed text, handwriting, and rubber stamps. Research cited in the post, published in BMJ Quality and Safety, estimates that diagnostic errors affect approximately 12 million U.S. outpatient adults annually, with information-handling failures cited as a contributing factor.

At Guardoc's processing scale, even a 1 percent error rate in condition detection generates thousands of incorrect records daily.

The architecture

Guardoc's pipeline uses a cost-tiered approach across multiple AWS services:

• Amazon Textract performs initial text and structural extraction from incoming PDFs at low per-page cost.
• Extracted content is chunked along clinically meaningful boundaries (medication lists, diagnosis sections, physician notes) rather than arbitrary byte ranges.
• Chunks are embedded with Amazon Titan Text Embeddings V2 and stored in Amazon DynamoDB, partitioned per patient.
• A custom pre-filter narrows candidates by document type, recency, and patient context before an in-memory k-nearest neighbor (k-NN) search retrieves the most relevant chunks.
• Amazon Nova 2 Lite performs a lightweight text-based review to eliminate non-matches before the more expensive multimodal stage.
• Amazon Nova Pro handles final classification on surviving pages, reasoning over raw PDF bytes including layout, handwriting, signatures, and stamps.

Every classification output resolves back to a specific source page in the original document, which Guardoc describes as a non-negotiable traceability requirement given the patient safety implications of clinical decisions.

According to AWS, Amazon Nova Pro showed "meaningful improvement" in multimodal PDF handling during Guardoc's internal evaluation, specifically around layout interpretation, checkbox state detection, and handwritten annotation recognition. No specific benchmark scores were disclosed for this comparison.

What this means

This is a customer case study, not a model release, and the performance claims — 46 percent error reduction, 70 percent fewer audit fines, $400K ROI — come from Guardoc, not from independent benchmarking. AWS has an obvious incentive to promote Bedrock adoption stories, so these figures should be read as a vendor testimonial rather than validated research.

The more interesting signal is architectural: Guardoc's cost-tiered pipeline — cheap embeddings and Nova Lite for filtering, expensive Nova Pro reserved only for final multimodal reasoning — reflects a now-common pattern in production LLM deployments where teams route the bulk of volume through cheaper models and escalate only the hardest cases. This mirrors similar tiering strategies used across the industry to control inference costs at scale, and it's a more replicable takeaway for other builders than the ROI figures themselves.

For healthcare specifically, the traceability requirement — every output resolving to a source page — is a meaningful design constraint that other regulated-industry applications of multimodal LLMs will likely need to adopt as a baseline, not an optional feature.