The scale of the documentation challenge
Guardoc Health says it now processes more than one million clinical documents every single day using Amazon Nova models through Amazon Bedrock. The company builds documentation software specifically for long-term care providers — nursing homes, skilled nursing facilities, and similar settings where paperwork volume is enormous and the margin for error is razor-thin.
The stakes are brutally concrete. Get a diagnosis code wrong, and Medicare claims get denied under the Patient-Driven Payment Model. Audit fines pile up. Worse, a missed condition can alter a patient’s entire treatment trajectory. AI in healthcare documentation is not some abstract efficiency play; it’s a risk management problem where the cost of failure is measured in both dollars and patient outcomes.
Guardoc’s published figures suggest they’re tilting the odds in the right direction. A 46 percent reduction in documentation errors. A 70 percent drop in audit fines. More than $400,000 in annual ROI for a single facility. The company hasn’t disclosed the baseline period or methodology behind those numbers, but the direction of travel is clear.
Why clinical documents break most AI pipelines
Clinical documents arrive in formats that would make most document processing systems cry. Multi-page PDFs with handwritten physician annotations layered over printed text. Prior authorization forms where a single checkbox state determines a coverage decision — but that checkbox might have a handwritten note next to it that overrides everything. Medication lists that show up as clean tables in one chart and free text buried inside a doctor’s note in the next. Patient intake forms mixing typed fields, rubber stamps, and handwriting on the same page.
Research published in BMJ Quality and Safety estimates that around 12 million US outpatients are affected by diagnostic error each year, with information-handling failures cited as a contributing factor. At Guardoc’s volume — one million documents daily — even a one percent error rate in condition detection would generate thousands of incorrect records every day. Each one carries its own patient safety or compliance consequence.
In one quarterly deployment covering two facilities and 200 patients, Guardoc says its system drove 847 documentation corrections, flagged 86 issues tied to PDPM reimbursement accuracy, and was associated with a 74 percent reduction in hospital transfers per 100 admissions. A separate case study across seven facilities and 1,618 residents identified 10,612 issues.
A retrieval pipeline built around cost as much as accuracy
Guardoc’s architecture is a multi-stage retrieval augmented generation (RAG) pipeline designed with a clear cost-tiering logic: cheap components handle the high-volume grunt work, and expensive multimodal reasoning only fires when it’s actually needed.
It starts with Amazon Textract, which extracts text and structural metadata from each incoming page. Guardoc treats this as the lowest per-page cost point in the pipeline. That output gets chunked along clinical boundaries — so a medication list or a diagnosis section stays intact rather than getting split by arbitrary character count.
Each chunk is embedded using Amazon Titan Text Embeddings V2 and stored in Amazon DynamoDB, partitioned by patient so retrieval never crosses patient boundaries. A custom pre-filter narrows the candidate set by document type and recency before a k-nearest neighbour search retrieves the chunks most relevant to a given classification query. At this stage, only page references are returned — keeping data transfer light.
Nova 2 Lite and Nova Pro handle the heavy lifting
Amazon Nova 2 Lite runs a text-based pass to remove obvious non-matches. Only the pages that survive every prior filter reach Amazon Nova Pro, which receives the raw PDF bytes and reasons over layout, handwriting, signatures, and stamps to produce the final classification. This is where the computationally intensive multimodal reasoning happens — and only for the subset of documents that actually need it.
Two document types account for most of what earlier pipeline versions missed, according to Guardoc. The first is physician attestation fields on prior authorization forms, where a handwritten note can override a printed checkbox. The second is patient-reported symptom sections, where handwriting often carries information that doesn’t appear anywhere else in the record.
Medication extraction presents a related problem. Drug names, dosages, routes, and frequencies show up in structured tables, in prose buried inside physician notes, in handwritten additions to printed lists, and in scans that have been faxed through multiple hands. Guardoc’s hybrid pipeline runs Amazon Textract first for clean printed tables, then passes both the original PDF and the Textract output to Amazon Nova Pro to resolve wrapped table columns, handwritten additions, and non-standard formats that OCR alone can’t parse correctly.
What this means for long-term care
“With the Nova family, we’re making it easier for healthcare organisations to detect high-risk cases earlier and act before issues become costly,” said Assaf Amiaz, Director of Product at Guardoc Health. “By automating workflows that once required manual oversight, the Nova family helps teams reduce compliance gaps, prevent errors, and focus more of their time on improving patient outcomes.”
The real takeaway here is that AI clinical documentation processing isn’t about replacing human judgment — it’s about catching the things humans miss when they’re buried under paper. A single nurse or coder might process hundreds of documents per shift. The error rate that seems small per document becomes massive at scale. Guardoc’s approach, with its cost-aware RAG pipeline and multimodal reasoning on the hard cases, is a template for how to deploy AI in regulated healthcare settings without blowing the budget on unnecessary compute.
The question now is whether other vendors follow the same playbook — or whether they’ll keep throwing expensive models at every document, regardless of whether the problem actually warrants it.
