Forty scanned pages in. The six fields that actually get used.
Referrals, prior auths, claims files, safety reports and outside records arrive as images, and someone must read them before anything moves. Exemplary AI reads them inside your own network, returns structured values with the page each came from, and never sends PHI anywhere, yours or your customers'.
A scanned referral packet is read, sorted into documents, and six fields are extracted. Each field carries a leader line to the page it came from. A person signs the schedule before anything is filed.
Same walls, whether you deliver the care, pay for it or supply it.
A health system, a payer's review team, a sponsor's safety desk and a device maker's complaint file all hit the same three walls, and so does every software company serving them. None of it is a technology gap.
It still arrives as an image.
A referral packet, a claims file, an adverse event report: scanned, faxed or photographed. Your system of record stores it. So does the product you sell. Neither can read it.
Someone reads every page.
Behind abstraction, coding, claims review, safety triage and registry submission is a trained reviewer retyping a document into a form. A department, or an outsourced contract. Always a cost line.
The same record, in six systems.
The discharge summary here, the imaging report there, the correspondence somewhere else. Answering one question means opening all of them and holding the thread together in someone's head.
The same walls, from your side of them.
This page serves everyone who holds or handles clinical records. Two audiences get a page of their own — same platform, their desks named.
The packet cannot tell you whether this is new.
Read alone, a referral looks like the beginning of the story. That is the edge of what one source covers, which is a different thing from the edge of what happened.
A question about time
Is this a new problem, or a continuation? It is the first thing a clinician needs, and answering it means knowing whether anything came before — so the drawing is a course, and the axis is months.
Read the document. Then do something with it.
Extraction on its own just produces another file. These are the jobs that become possible once the page is machine-readable, run by your own team or shipped inside what you sell.
Clinical document intelligence
Extraction from clinical notes, lab reports, prescriptions, claims files and insurance paperwork, with medical entity linking: a drug name becomes a drug, a date an event on a timeline. 99.2% accuracy on the document types it is set up for. Every value carries its source page. Checking, not searching.
Prior authorisation, from both sides
One side assembles the payer's checklist from the chart: diagnosis, prior therapies tried, contraindications, each line pointing at the page behind it, not someone's recollection. The other side reads the packet that arrives against the criteria that apply. Same document, two desks, one job.
Medical record review at scale
Claims, appeals and risk adjustment start with the same forty pages. The agent pulls the documented evidence behind each line and points at the page it came from.
Safety and complaint intake
Adverse event reports and device complaints arrive as faxes, portal exports and emailed forms. The agent classifies the report, pulls the reportable fields, and flags what a person must confirm.
Correspondence in the reader's language
Discharge instructions, appointment letters and medical information replies, drafted in the language the reader speaks, and a person still signs it before it goes out.
A forty-page referral packet arrives at 7:40am.
The whole thing — no step skipped, and no step automated that shouldn't be. By the time anyone opens the intake queue, five of the six have already happened.
Draft summary
- Referred for a specialist opinion. The referring practice and the reason are on the covering letter. p.1
- Two prior therapies are documented in the record, the more recent one inside the last year. p.6
- The most recent imaging predates the referral and is already on file. p.9
Marked for a person
- New problem, or a continuation?The packet reads as new. The indexed record does not.
- Outcome of the second therapyDocumented as tried. No page states how it ended.
Every line carries the page it was read from. Nothing is filed until this row is signed.
It lands in the filestore.
Faxed in, or dropped into a watched folder. OCR runs, the document is indexed, and it is visible only to the group that owns it.
The agent reads all forty pages.
Not a keyword scan. It identifies what is actually in the packet — a referral letter, two outside progress notes, an imaging report, and an insurance card — and treats each as its own document type.
It pulls the fields that actually get used.
Reason for referral, referring provider, primary diagnosis, prior therapies tried, and the date of the most recent imaging. Nothing else — a summary of everything is a summary of nothing.
It checks them against the chart.
Grounded retrieval over the patient's indexed history answers the questions the packet cannot: is this a new problem or a continuation, and has the prior therapy already failed here?
It drafts the intake summary — with citations.
One page. Every extracted value carries the page number behind it, so whoever reviews it verifies in seconds instead of re-reading forty — a coder, an intake team, or a reviewer inside your own product.
A person signs it off.
Nothing is filed automatically. The agent proposes, your team confirms, and the audit trail records who approved what and when.
The answer stops depending on which document turned up.
The packet, the chart, the imaging and the correspondence stop being four separate readings and become one record, with the stretches nobody observed marked as unobserved. That changes the question you can ask. Not what does this file say, but what does everything we hold say, and which part of it nobody has looked at.
A claims review, a safety signal and a registry submission all start from indexed records rather than a records pull.
A gap in the record shows up as a gap, instead of being read as nothing having happened.
The corpus built for intake is the corpus the next question reads, so the second use costs a configuration, not a project.
PHI does not leave your network. Not even for inference.
Most AI vendors ask you to accept that patient data crosses a boundary and comes back. We took the other route — the model goes to the data. If you sell into health systems, that is also the difference between a procurement conversation and a security review.
The model runs where the data lives
Bare-metal, private cloud, or fully air-gapped. Inference happens inside your network — no page, no extracted field, and no embedding is sent to a third party.
Encrypted, scoped and logged
AES-256 at rest, TLS 1.3 in transit, a BAA-ready architecture, and role-based access so a group only ever sees the records it owns.
One fewer sub-processor
Every cloud dependency you add becomes a line in your customer's vendor review. Running inside their perimeter removes the line instead of defending it.
Deployed inside your perimeter — or your customer's — under controls that already exist. Model updates arrive by secure transfer, so an air-gapped site never needs a connection.
The same platform, pointed at clinical documents.
Nothing here is a separate healthcare product. It is the platform, configured — which is why the same deployment also covers the rest of what your organisation does, and why it can ship inside what you build.
- Document intake
- Filestore folders, watched sync, upload or API
- Reading
- OCR, indexing, medical entity recognition and linking
- Retrieval
- Grounded retrieval across indexed records
- Structure
- Knowledge Graph entities and relationships
- Access
- Groups and IAM — least privilege, deny overrides allow
- Systems
- EHR / FHIR connectors and 170+ integrations via MCP
- Oversight
- Human-in-the-loop approval and a full audit trail
- Delivery
- Branded workspace, embedded in your product, or by API
- Deployment
- On-premise, private cloud or air-gapped
What healthcare teams ask first.
Does this replace our EHR, or sit alongside it?
How do we check that an extracted field is actually right?
Is patient data ever sent to a third-party model?
We are a payer, a sponsor or a manufacturer, not a provider. Is this the same deployment?
We build software for healthcare rather than run a hospital. Does this still fit?
Our archive is scanned, handwritten in places, and not all in English.
See Exemplary AI in action
Book a demo and we'll show you purpose-built agents grounded in your own knowledge — deployed on your infrastructure.