Document AI that separates inbound bundles, works out what each document is, extracts what matters, matches it to the right patient, and files it where your teams already look.
Healthcare still runs on paper that stopped being paper. Referral packets, orders, results, payer correspondence, authorization determinations, records requests and remittance arrive as faxes, scans and images, and a person opens every one of them. CaliberFocus builds document processing that handles the layouts it has never seen before, scores its own confidence field by field, and escalates anything it should not decide alone.
The measure is not how many documents were processed. It is how many were filed correctly, and how quickly
A referral packet lands in the fax queue at four in the afternoon. It contains a cover sheet, a referral form, three progress notes, an insurance card image and a lab report, all in one transmission. Someone has to separate them, identify each one, work out which patient they belong to, extract the detail that triggers the next step, file each document to the right place in the chart, and start the referral. That is roughly ten minutes of work, and it happens hundreds or thousands of times a week.
Every referring office, every payer, every lab and every imaging center uses a different layout, and they change them without telling you. Documents arrive skewed, stamped, handwritten, annotated in the margin, photocopied for the fourth time, or as a phone photograph of a printed form. Template based capture worked on the fixed forms it was built for and broke on everything else, which is why most organizations configured a handful of templates, gave up, and went back to keying.Â
A document filed to the wrong chart is worse than one that was never filed, because now it is wrong in two places and nobody is looking for it.
Manual classification
Someone decides what kind of document it is.
Manual extraction
Names, dates, identifiers, diagnoses, procedures and payer details are re-keyed.
Manual matching and validation
Values are compared by eye against the record, encounter, order or referral.
Manual routing
Documents move based on staff knowledge rather than consistent rules.
Manual exception handling
Poor scans, handwriting, missing pages and conflicts take the longest.
Misfiling is a patient safety event, not an operations metric.
A result filed to the wrong chart, or never filed at all, is a clinical risk with a documented history of harm. It should be governed with the seriousness that implies, and measured against a target of zero rather than reported as an accuracy percentage.
A result filed to the wrong chart, or never filed at all, is a clinical risk with a documented history of harm. It should be governed with the seriousness that implies, and measured against a target of zero rather than reported as an accuracy percentage.
Document AI reads the way a person does. It looks at the page, works out what kind of document it is, finds the information that matters wherever it happens to sit, and checks its own answer against what your systems already know. Where it is not sure, it says so, field by field, and hands that field to a person.
Step 1
Ingest
fax, secure email, SFTP, scanner, portal upload, health information exchange or API, into a single processing pipeline.
Step 2
Split
Multi document bundles are separated into discrete documents at the correct page boundaries.
Step 3
Classify
Each document is identified by type. Referral, order, result, authorization determination, records request, remittance, consent, correspondence.
Step 4
Extract
Relevant fields are located and read wherever they appear, including handwriting, checkboxes, tables and stamps, with a confidence score on each value.
Step 5
Match
The document is resolved to the correct patient, encounter, order, provider or claim against your systems.
Step 6
Validate
Extracted values are checked against source systems and business rules. Cross field consistency, expected ranges, required fields and known sender patterns.
Step 7
Route and file
The document is filed to the correct chart location and note type, discrete data is written where permitted, and the downstream workflow is triggered.
Step 8
Learn
Reviewer corrections feed accuracy monitoring, threshold tuning and model improvement, tracked by document type and by sender.
| Capability | Manual indexing | Template OCR | Document AI |
|---|---|---|---|
| Handles a layout never seen before | Yes | No | Yes |
| Splits multi-document bundles | Yes | Rarely | Yes |
| Reads handwriting and checkboxes | Yes | Poorly | Yes |
| Confidence score on every field | No | No | Yes |
| Validates against source systems | Sometimes | No | Yes |
| Adding a new document type | Training a person | A new template build | Configuration and examples |
| Cost and speed at volume | High, slow | Low, brittle | Low, resilient |
Confidence belongs to the field, not to the document.
A document that is ninety five percent accurate sounds excellent until the five percent is the medical record number. We score and threshold every field separately, and set the tolerance on identifier and clinical fields far tighter than on a fax cover date. Document level confidence hides exactly the errors that matter most.
A document that is ninety five percent accurate sounds excellent until the five percent is the medical record number. We score and threshold every field separately, and set the tolerance on identifier and clinical fields far tighter than on a fax cover date. Document level confidence hides exactly the errors that matter most.
Most document AI evaluations test extraction accuracy on clean samples and stop there. In production, the work that determines whether anything is saved sits either side of extraction: separating the bundle correctly at the front, and matching, validating and filing correctly at the back.
Ingest and Understand
Multi-Channel Ingestion
Fax, secure email, SFTP, scanner, portal upload, health information exchange, Direct messaging and API into one pipeline with one queue, one audit trail and one set of rules.
Bundle Splitting
Separates multi document transmissions at the correct boundaries by reading content rather than counting pages or looking for separator sheets. This is the step that silently corrupts everything downstream when it goes wrong.
Document Classification
Identifies document type across the healthcare categories your organization receives, including sender specific variants and layouts that have never been seen before.
Extraction on Real-World Quality
Reads printed and handwritten text, checkboxes, signatures, stamps, tables and marginal annotations on skewed, low contrast and repeatedly photocopied pages, with a confidence score on every extracted value.
Resolve and File
Patient and Encounter Matching
Resolves each document to the correct patient, encounter, order, provider or claim against your systems, with a confidence floor below which nothing is filed automatically.
Validation Against Source Systems
Checks extracted values against the record and against business rules. Cross field consistency, expected ranges, required fields, duplicate detection and known sender behavior.
Cross-Document and Completeness Validation
Compares values across the documents inside one packet, and checks for missing pages, signatures, attachments and required fields before the workflow moves forward rather than after it stalls downstream.
Document Summarization
Structured summaries of long external records so staff can find what matters without reading every page, with citations back to the source pages they came from.
Structured Output and Coded Mapping
Produces discrete data alongside the document image, mapped to your coding standards where the document type supports it, so results and orders contribute to the record rather than sitting as an unsearchable image.
Chart Filing and Workflow Trigger
Files to the correct chart location and note type, writes permitted discrete data, and starts the downstream workflow. Referral created, order acknowledged, denial routed, request queued.
Control and Improve
Per-Field Confidence and Thresholding
Independent confidence on every value with a separate threshold per field, tuned by risk rather than set once for the whole document.
Provenance and Source Traceability
Every extracted value links back to the exact region of the page it came from. A reviewer sees the number and the pixels it was read from, side by side, without hunting through the document.
Exception Queue and Reviewer Console
One place to work everything the system would not decide alone, with the source image, the extracted values and the reason for escalation together on one screen.
Accuracy Monitoring by Type and Sender
Straight through rate, field accuracy, correction patterns and drift tracked per document type and per sender, so a single referring office changing its form shows up as a signal rather than as noise.
What CaliberFocus does, and does not do?
We are not a capture or content management platform vendor, and that is deliberate. We assess your real document mix against your systems, select and configure the right processing stack, build the matching, validation and filing logic that determines whether documents land correctly, establish the confidence thresholds and governance with your health information management leaders, and own accuracy through the first document types.
The right first document type is high volume, arrives through one dominant channel, has a clear downstream action, and is currently keyed by hand. Pull your fax and scan volumes by type before you pick. In most organizations two or three document types account for the majority of inbound pages.
| Workflow | What the system handles | What always routes to a person | What moves |
|---|---|---|---|
| Referral intake | Splits the packet, classifies each document, extracts patient, payer, ordering provider and clinical detail, checks completeness and creates the referral. | Clinical triage, urgency and appropriateness decisions. | Referral leakage, time to appointment |
| Results, reports and clinical correspondence | Classifies, extracts key values, matches to the ordering encounter, files to the correct chart location and notifies the ordering clinician. | Abnormal or critical values, and any ambiguous patient match. | Time from receipt to clinician visibility |
| Orders and requisitions | Reads the order, validates against the record, extracts diagnosis and authorization detail and queues for acknowledgment. | Clinical validation and anything requiring clarification with the ordering provider. | Order entry time, scheduling delay |
| Prior authorization determinations | Classifies approvals, denials, pends and requests for information, extracts reference numbers and criteria, and routes by outcome. | Medical necessity interpretation, peer to peer and appeal strategy. | Determination turnaround, delayed procedures |
| Payer correspondence and remittance | Extracts from paper and non standard remittance and correspondence, reconciles against the claim and routes variances. | Unresolved variances and cash application exceptions | Posting lag, unapplied cash |
| Release of information and records requests | Classifies the request, validates authorization scope and expiry, identifies the records in scope and assembles the response for review. | Every disclosure decision, and any request involving sensitive categories. | Request turnaround, regulatory response times |
| Patient forms and intake documents | Reads completed forms including handwriting and checkboxes, validates against the record and updates registration. | Conflicting information and anything requiring verification. | Registration rework, check in time |
| Credentialing and provider documents | Classifies licenses, certifications, insurance and attestations, extracts expiry dates and updates provider records. | Committee decisions and primary source verification judgment. | Time to credential, revenue lost to lapses |
| Fax and shared inbox triage | Classifies everything arriving in the shared fax and scan inbox, identifies the intended workflow or recipient and routes it without a person opening each item. | Unrecognized documents, uncertain routing and anything flagged urgent. | Inbox backlog, time to route |
| Chart abstraction and audit support | Locates, retrieves and pre-abstracts documents for audits, quality reporting and payer requests. | Clinical validation of every abstracted finding. | Abstraction hours, audit response time |
| Administrative and enterprise documents | Invoices, contracts, vendor documents and HR forms processed on the same pipeline as clinical content. | Commercial decisions and policy exceptions. | Processing cost, cycle time |
Some documents are never auto-filed, regardless of confidence.
Substance use disorder records, behavioral health, HIV and other specially protected results, genetic testing, records relating to minors, and anything arriving with a restrictive authorization are routed to a qualified reviewer before filing. These categories carry disclosure restrictions that a filing rule cannot safely interpret, and a correct extraction filed to the wrong access level is still a disclosure event.
Substance use disorder records, behavioral health, HIV and other specially protected results, genetic testing, records relating to minors, and anything arriving with a restrictive authorization are routed to a qualified reviewer before filing. These categories carry disclosure restrictions that a filing rule cannot safely interpret, and a correct extraction filed to the wrong access level is still a disclosure event.
Plenty of tools will read a document and give you structured data back. The saving only arrives when that data reaches the chart, the queue and the workflow without a person moving it. Filing correctly is a deeper integration problem than extraction, and it is where these programs are won or lost.
Correct chart destination
The right patient, the right encounter, the right document type and the right chart location, resolved automatically and confirmed before filing.
The image and the data together
The source document is filed alongside the discrete data extracted from it, so a clinician can always see the original.
Downstream workflow triggering
Filing is not the end. The referral is created, the ordering clinician is notified, the denial is routed, the request is queued.
Sender feedback loop
Where a sender consistently transmits incomplete or unreadable documents, that pattern is surfaced so it can be fixed at source rather than absorbed forever.
One audit trail
Every document, every extracted value, every correction and every filing action logged in one place, linked to the source image.
Graceful degradation
If a downstream system is unavailable, documents queue safely with their processing intact rather than failing or filing into the wrong place.
EHR and practice management
Patient and encounter matching, chart filing, note type mapping, discrete data write back
Document management and content platforms
Storage, retention, versioning and retrieval within your existing repository
Fax and secure transmission platforms
Digital fax, Direct messaging, secure email and health information exchange intake
Scanning and capture infrastructure
Front desk and back office scanning brought into the same pipeline
Integration engine
Routing and transformation kept on your governed interface layer
Clearinghouse and revenue cycle systems
Remittance, correspondence and authorization document flow
FHIR and vendor APIs
Structured write back of results, orders and observations where supported
Data and analytics platforms
Straight through rate, accuracy and backlog measurement
The EHR and your document repository remain the record. Document AI is a processing layer in front of them, not a parallel store of clinical documents. Retention, legal medical record definition and disclosure governance stay exactly where they already sit.
PHASE 1
Document mix assessment
2 to 3 weeks
Volume and channel analysis by document type, sample review against real inbound documents, baseline turnaround and touch counts captured.
PHASE 2
Build and calibrate
4 to 6 weeks
Pipeline built, classification and extraction configured, matching and validation logic implemented, gold standard test set created with your HIM team.
PHASE 3
Shadow and pilot
3 to 4 weeks
Runs in parallel on live volume. Field level accuracy measured against human decisions. Thresholds set from real data before any automatic filing
PHASE 4
Production and scale
Ongoing
Straight through rate raised in steps against measured accuracy. Monitoring by type and sender, and expansion to the next document type.
Straight through processing is not a setting you switch on at go live. It is a rate that rises as measured accuracy justifies it, per document type and per field. A vendor who quotes you a straight through rate before seeing your documents is quoting you their average, not your outcome.
| Level | What the system does | What a person does | Typical stage |
|---|---|---|---|
| Classify and route | Identifies the document and sends it to the right queue. Extracts nothing that is relied upon. | Keys and files everything | First weeks, and any new document type |
| Extract and present | Presents extracted values alongside the source image for confirmation. | Confirms or corrects every field, then files. | Once classification accuracy is proven |
| File with exception review | Files automatically where every field clears its threshold. Routes documents with any low confidence field. | Works only the flagged fields, not whole documents. | The working state for most document types |
| Straight through | Files and triggers the workflow with sampled audit only. | Audits samples and investigates alerts. | High volume, low variance types with proven accuracy |
| Filing on an unresolved patient match | Not offered. | Not applicable | Never |
Thresholds set per field,
by risk
Identifier and clinical fields carry far tighter tolerances than administrative ones. One document level threshold would either flood the queue or let the wrong errors through.
Patient match has a
hard floor
Below the match confidence floor, nothing is filed automatically, regardless of how clean the extraction was. This is the one rule that is not tunable by workflow.
Source image beside
every value
Reviewers see the extracted value and the region of the page it came from together, which turns a correction from a search into a glance.
Field level review, not document level
A reviewer confirms the two uncertain fields rather than re-reading the whole document. This is where most of the labor saving actually comes from.
Second review on critical fields
Defined high risk fields can require independent confirmation before filing, in line with your existing HIM controls.
Misfile detection and correction
A defined route for clinicians and staff to report a misfiled document, with correction, root cause analysis and threshold adjustment, not just a re-file.
Corrections drive the model
Every override is captured with its reason and feeds accuracy monitoring, threshold tuning and configuration by document type and by sender.
The exception queue is the product working, not the product failing.
A system that files everything and flags nothing is not more accurate. It is less honest. We would rather send you a document with one uncertain field than file it confidently in the wrong chart, and we design the review experience so that handling the exception takes seconds.
A system that files everything and flags nothing is not more accurate. It is less honest. We would rather send you a document with one uncertain field than file it confidently in the wrong chart, and we design the review experience so that handling the exception takes seconds.
Document processing touches protected health information at scale, holds images at rest, and makes filing decisions that carry disclosure consequences. It also regularly receives documents that were never meant for you. Governance here is not a policy appendix. It determines how the pipeline is built.
Data
protection
- BAA executed before any access to protected health information
- Encryption in transit and at rest for images, transcripts, extracted data and logs, with key management under your control where required
- Minimum necessary access. Each pipeline stage receives only the data its function requires
- Retention and deletion defined separately for source images, extracted data and processing logs, aligned to your legal medical record definition
- Data residency and processing location defined and contractually fixed
Sensitive & misdirected documents
- Specially protected categories, including substance use disorder, behavioral health, HIV, genetic and minor related records, route to a qualified reviewer before filing
- Restrictive authorizations and disclosure limitations are read and honored rather than ignored at filing
- A defined procedure for documents received in error, including quarantine, non-filing and notification to the sender
- Minimum necessary collection, processing and access for the approved documentation workflow
- Role based access to recordings, transcripts, drafts, configuration and administrative functions
- Data residency and processing location defined and contractually fixed
Model
governance
- No client documents, images or PHI used to train foundation models, enforced contractually and technically
- Model version and configuration documented per document type, with change control on anything affecting extraction or classification
- Accuracy evaluated against a gold standard set built from your real documents, including degraded and handwritten samples rather than clean ones
- Performance reported by document type and by sender, never as a single aggregate accuracy figure
- Accuracy requirements set by the consequence of getting the field wrong, not one automation threshold applied to every field and document
- Drift alerts and defined rollback when a sender format change degrades accuracy
Accountability &
audit
- Every extracted value traceable to its source region on the page, and every filing action attributable and time stamped
- Complete audit trail linking document, extraction, review, correction and final filing destination
- Misfile rate reported as a safety metric with a zero tolerance target, not folded into an accuracy percentage
- A named accountable owner for the pipeline, with health information management leadership sign off on thresholds and sensitive category handling
A single aggregate accuracy number is the least useful figure in document processing. It averages a clean payer remittance with a handwritten referral and tells you nothing about either. We baseline by document type before go live, build a gold standard set from your real documents, and report at that level.
| Category | What we measure | Why it matters |
|---|---|---|
| Safety | Misfile rate and time to detection, by document type | Reported as a safety metric with a zero tolerance target |
| Accuracy | Field level accuracy by document type and by sender, with identifier fields reported separately | Aggregate accuracy hides the errors that matter |
| Coverage | Straight through rate by document type, and the categorized reason every other document needed a person | The second half is what drives the fix backlog |
| Speed | Receipt to filed, receipt to clinician visibility, receipt to downstream action, backlog age | Clinical information nobody can see is not filed |
| Capacity | Documents per FTE, staff hours returned, touches per document, overtime and outsourcing spend | The operating business case |
| Downstream | Referral leakage, time to appointment, authorization turnaround, posting lag, audit response time | Where document processing actually shows up in the P and L |
| Workflow completion | Whether the complete downstream workflow finishes faster, not only whether the document was processed faster | The test of whether automation reached the work or stopped at the document |
| Quality of intake | Incomplete and unreadable documents by sender | Some of the biggest wins are fixing the sender, not the system |
Not every document type will reach high straight-through rates.
Automate the right types well and leave the messy ones in review rather than forcing a misfile problem.
Automate the right types well and leave the messy ones in review rather than forcing a misfile problem.
Not clean samples. The actual inbound pages, including the skewed ones, the handwritten ones and the fourth generation photocopies. We will run them, report classification and field level accuracy on your own documents, show you what would have filed automatically and what would have gone to review, and tell you whether this document type is worth automating. If it is not, we will say so and tell you which one is.
Bring us a specialty or clinical setting where clinicians are spending too much time creating notes. We will assess where ambient documentation fits, what must remain clinician controlled, how it should integrate with your EHR, and how to measure whether it is actually reducing burden.
One conversation with people who have run these deployments, and a written readiness view you can use with or without us.
- AI Agents and Workflow Automation
- Voice and Conversational AI
- Document AI and Intelligent Processing
- Generative AI and Enterprise Copilots
- AI Strategy and Governance
- HCC and Risk Adjustment Analytics
Security & Compliance
