Use caseDocument AIWorkflow automation
Referral faxes and letters turned into patient records, with missing details chased automatically
Referrals from fax, email or KIM read, checked against each exam's requirements, linked to the right patient and chased for gaps. No clinical decisions.
A blueprint, not a client story. The business described is illustrative; the architecture, integrations and trade-offs are real, and this is how I would build it. By Ergini, .
The short version
A referral intake workflow for multi-site clinics and imaging groups. It reads referrals that arrive by fax, email, KIM or patient upload, extracts the patient, insurance, referrer, diagnosis codes and requested exam, checks completeness against each exam's requirements, links the referral to the right patient record and asks the referring practice for anything missing. Complete referrals move to scheduling. It makes no clinical decisions: urgency goes to clinicians, uncertain patient matches go to staff, and the data stays in EU hosting.
- Best for
- Multi-site imaging, physiotherapy or specialist groups where referrals still arrive by fax and staff re-key them into the practice system.
- Connects to
- Fax-to-email and practice inboxes, KIM (Kommunikation im Medizinwesen), RIS or practice management system, Scheduling, EU-hosted extraction and review service
- The AI does
- Reads faxed, scanned and handwritten referrals, extracts what booking needs, and drafts specific requests to referrers for anything missing.
- People do
- Confirm uncertain patient matches, review every referral marked urgent, decide anything clinical, and book the cases that need a conversation.
- Built as
- AI Workflow Automation, usually $15K - $30K
The fax machine is still the front door
Picture a radiology group with five outpatient sites in Germany offering MRI, CT, X-ray and ultrasound, and around 300 referrals a week. Some arrive by fax from GP and orthopedic practices and land as PDFs in a shared mailbox three people work from. Some come as PDFs by email, a growing share over KIM from practices on the telematics infrastructure, and some as phone photos that patients upload through the website before they call to book.
At the front desk, someone reads each one, finds the patient in the practice system, types in the insurance details, the referring doctor, the diagnosis and the exam requested, and decides whether it can be booked. Often it cannot. The clinical question is missing, 'MRI knee' does not say which knee, or a contrast CT needs a recent kidney function result that the group's protocol asks for. Each gap is a phone call to a practice that is just as busy, and a referral waiting in a pile until someone calls back.
Referrals that stall are lost quietly. The patient books elsewhere, the referring practice sends its next patients elsewhere too, and nobody notices, because a referral that never became an appointment leaves no trace in the schedule. In the US, Tennr raised $101M at a $605M valuation in June 2025 to automate exactly this intake. European groups have the same problem with different systems, different insurers and stricter rules on health data.
What happens to a referral today
In most groups the sequence looks like this, and every step is done by hand.
- A fax lands as a PDF in the shared mailbox, a KIM message appears in the practice system, or a photo arrives in the website's upload folder.
- Someone opens it, works out which site and exam it is for, and deciphers the handwriting.
- They search for the patient by name, then by date of birth, and decide whether the 'H. Müller, 03/58' on the fax is the Hans-Peter Müller already in the system.
- They type insurance status, referring doctor, diagnosis codes and the requested exam into the practice system.
- They check what the exam needs: clinical question, region and side, implant questions for MRI, kidney function for contrast.
- If anything is missing, they phone or fax the referring practice and put the referral on a pile.
- When the answer comes back, someone has to remember which pile it went on.
Nothing in this sequence is clinical. It is reading, matching, typing and chasing, which is the part that can be automated without going near a clinical decision.
From an incoming fax to a referral ready to book
A model appears twice: reading the document and drafting the request to the referrer. Matching, completeness and routing are rules, and anything clinical goes to a clinician, the pattern described in human-in-the-loop AI.
01 Trigger · Microsoft Graph, KIM client
A referral arrives on any channel
Fax-to-email PDFs, KIM messages, practice emails and patient uploads all enter one queue, stamped with the time they arrived. Nothing is processed outside the EU-hosted environment.
02 AI model · EU-hosted vision model
Classify and extract
The model decides whether the document is a referral, a doctor's letter, a prior report or something else, then extracts the patient's name and date of birth, insurance, the referring doctor with LANR and BSNR (the doctor and practice numbers), ICD-10 codes, requested exam, region and side, the clinical question and any urgency the referrer marked. Every field keeps its position on the page.
03 Plain code
Catch duplicates and find the patient
A referral matching an existing one on patient, exam and date is attached to it instead of creating a second. An exact match on several identifiers, such as insurance number, surname and date of birth, links the referral to an existing record; anything weaker produces candidates for staff. The system never merges two existing records.
04 Plain code
Check completeness for the exam
Each exam type has a requirement list, owned by the group's lead radiographer and practice manager. The check compares the extracted fields against it and names exactly what is missing.
05 Decision
Where does it go next?
Urgency is checked before anything else.
- The referrer marked it urgent, or the text contains a term on the clinicians' escalation list then straight to the duty clinician's queue, ahead of every other step
- Complete, and the patient is linked then released to scheduling
- Something is missing then a request to the referring practice is drafted
- Uncertain patient match or unreadable fields then to the intake team's review screen
06 AI model
Draft the request to the referrer
Short and specific: which patient, which exam, what is missing and why booking needs it. It goes back over KIM where the practice has it, otherwise by fax or email. Staff approve the first requests to each practice, and a template only sends on its own once its drafts have gone out unchanged for a few weeks.
07 Person
Staff and clinicians take what the rules route to them
The intake team confirms matches and unreadable fields on a screen that shows the fax region next to each value. Clinicians see urgent referrals and everything clinical, including protocoling the exam.
08 System · RIS or practice system interface
Write to the record and release to booking
The referral, the extracted data and the original document are written to the RIS or practice system, and complete referrals appear in the booking queue with exam, duration and preparation known. Chased referrals are re-checked when the answer arrives. If the practice system is unreachable, referrals wait with their status visible and are written once it returns, never twice.
09 Result
A status and an audit trail for every referral
Who or what read each field, every change, every message to the referrer and how long it took to answer. For the first time, the group can see how long referrals wait at each stage and which ones never became appointments.
What 'complete' means depends on the exam
The requirement lists belong to the group, not the model. This is the kind of table the build starts from, filled in with the group's own protocols.
| Exam | Needed before booking | Who closes a gap |
|---|---|---|
| MRI | Clinical question, region and side, contrast yes or no, implant and pacemaker questions answered | The referrer for the clinical question; the patient for implant questions, on a pre-visit form |
| CT with contrast | Clinical question, a recent kidney function result where the protocol asks for one, known allergies | The referrer for lab results; a clinician decides whether an older value is acceptable |
| X-ray | Clinical question, region and side | The referrer; often bookable the same day once complete |
| Ultrasound | Clinical question, region, preparation such as fasting for abdominal exams | The intake team, which sends preparation instructions with the booking |
| Any exam, statutory insurance | Referral slip details: insurance data, the referring doctor's LANR and BSNR, an ICD-10 code | The referrer for missing slip details; the insurance card is read at the visit |
| Any exam, private or self-pay | Payer details and, where an insurer requires it, an authorization reference | The intake team, with the patient |
A faxed MRI referral with one piece missing
A typical fax: a printed form, a handwritten clinical question, and a practice stamp half off the page.
Fax inbox, site 2, Tuesday 08:14
- classify(file: "fax_0814.pdf")referral slip, 1 page, low-resolution fax, rotated about 2 degrees / handwriting in the clinical question field
- extract(file: "fax_0814.pdf")patient name legible, date of birth ?.03.1958 (day unclear) / statutory insurance, insurance number legible / referrer LANR and BSNR legible / ICD-10 M23.2 / exam 'MRT Knie' / side not stated / question 'V.a. Innenmeniskusläsion' (suspected medial meniscus tear)
- find_patient(insurance_no: "A1...", surname: "M...", dob: "?.03.1958")1 candidate, record P-88412 / insurance number and surname match exactly / date of birth on record 14.03.1958 fits the legible digits / below the automatic link rule: staff to confirm
Intake team · Review screen
Confirmed P-88412, 08:31.
- check_completeness(exam: "MRI knee", referral: "R-40318")missing: side / present: clinical question, ICD-10 code, referrer details / implant questions: pre-visit form goes out with the booking
- urgency_scan(referral: "R-40318")no urgency marked by the referrer / no escalation terms / routine
Intake workflow · KIM
Referral for H. M., born 14.03.1958, MRI knee, received today at 08:14: the side is not stated. Could you reply with left or right, or send a corrected referral? We will book as soon as we have it. Intake team, site 2.
- The practice answered at 11:40 with 'rechts'. The check ran again, found nothing missing, and released the referral to booking.
Failure modes a clinic cannot afford
Two patients, one name
Common surnames and shared birthdays are routine in a group that sees thousands of patients. A referral is linked automatically only on an exact match of several identifiers, anything less is shown to staff as candidates, and two existing records are never merged by the system. A wrong link puts one patient's history in another's file, which is far worse than a minute of staff time.
An urgent referral behind routine ones
Urgency is checked first: the referrer's own marking, and terms on an escalation list the clinicians maintain. The scan is tuned to over-flag, because a false alarm costs a clinician a glance and a missed one can cost a patient. Nothing the referrer marked urgent is ever downgraded by the workflow.
Handwriting and faxes of faxes
A fax of a photocopy of a handwritten form defeats any reader some of the time. Low-confidence fields go to staff with the image region beside them, and validation rejects impossible values, such as an ICD-10 code that does not exist or a date of birth in the future.
Drifting into clinical decisions
It is tempting to let the model decide that a referral is 'fine without the side' or that another exam would be 'more appropriate'. It never does. There is a regulatory line as well: software that ranks or selects patients for diagnosis or treatment can qualify as a medical device under the EU MDR, and an intake tool should stay clearly administrative.
Health data leaving where it should stay
Referrals are special-category data under GDPR Article 9. The build runs in EU hosting with a data processing agreement for every processor, a DPIA before go-live, access by site and role, and no referral content in logs or model training. In Germany, patient secrecy under section 203 of the Criminal Code also requires every service provider to be bound to confidentiality. The GDPR-compliant AI development page covers how that paperwork is handled.
A group that will not use external models at all
Some groups, or their data protection officers, rule out sending health data to any external model provider. Extraction then runs on a self-hosted model on EU infrastructure the group controls. Open models are now good enough for classification and extraction of this kind; the trade-off is a fixed monthly infrastructure cost.
Administrative, never clinical: the dividing line
The AI model
Read the fax, letter or KIM message and extract the fields
Varied forms, handwriting and poor scans are what the model is for.
Draft the request to the referring practice
A specific request gets answered faster than 'referral incomplete'.
Plain code
Catch duplicates and link referrals on exact identifier matches
Identity is a rule with a strict threshold.
Check completeness against the exam's requirement list
The group's rules, applied the same way at every site.
Surface urgency markers
The referrer's marking and a fixed term list, tuned to over-flag.
A person
Confirm uncertain patient matches and unreadable fields
A wrong link is a patient safety issue.
Review urgent referrals and protocol each exam
Clinical judgment, including whether an X-ray or CT exposure is justified, belongs to clinicians.
Book exceptions: sedation, interpreters, patients with complex needs
These need a conversation, not a rule.
Tennr, your practice system's own features, or a build for a European group?
In the US, Tennr is the name to know. It raised $101M in June 2025 to automate referral intake and is strongest with medical equipment suppliers and specialty practices, so a US practice on a supported EHR should evaluate it before commissioning anything. Practice management and RIS vendors are also adding document classification features, so ask yours what is already included in the license you pay for.
For European groups the calculation changes. The inputs are German, French or Dutch referral forms, KIM messages and faxes; the systems are European practice and radiology systems with interfaces of mixed quality; and health data needs EU hosting, a DPIA and processors the data protection officer will sign off. A custom workflow earns its cost when you have several sites, several exam types with different requirements, and a completeness problem that is costing referrals.
I would start narrow: fax only, three or four exam types, and the completeness rules the front desk already applies from memory. The workflow sits beside the practice system and writes into it rather than replacing anything, and grows exam by exam as the intake team trusts it. Once referrals turn into bookings reliably, the same group can fill late cancellations with a waitlist agent.
How you would know it is working
A blueprint has no results to report, so here is what I would measure from the first week instead, on your own data.
- Referral-to-booking time
- From arrival to booked appointment, split into waiting on the referrer, waiting on the patient and internal handling.
- Referrals that never became appointments
- Visible for the first time, because every referral now has a status. Broken down by referring practice and exam.
- Complete on first receipt, by referring practice
- Shows which practices send incomplete referrals, and whether specific requests back to them change the pattern.
- Staff touches per referral
- How many referrals needed a person to confirm a match or a field, and which fields cause it, so extraction and the review screen improve where it counts.
- Time to clinician review for urgent referrals
- Minutes from arrival to a clinician opening each urgent referral, with every urgent case audited weekly in the first months.
What a build like this costs
This is built as AI Workflow Automation, which runs $3.5K - $60K overall. A build like this one usually lands in the multi-step workflow tier: $15K - $30K, 3-5 weeks. The first working version runs on your real data well before the end of that window.
What it costs to run
Reading a one- to three-page referral and drafting the occasional request costs cents per referral on hosted models. Self-hosting turns that into a fixed monthly server or GPU cost, which makes sense at higher volumes or when policy rules out external providers. Fax-to-email and KIM services are costs most groups already carry.
What moves the price
- The practice system or RIS interface: a documented API or HL7 feed, a file import, or screen automation as a last resort
- How many channels at launch: fax only, or fax, KIM, email and patient uploads
- The number of exam types and how detailed each requirement list is
- Hosting: EU cloud under a data processing agreement, or self-hosted models on the group's infrastructure
- Handwriting share and fax quality, which decide how much staff review the extraction needs
Who this is for
- Multi-site radiology and imaging groups taking referrals by fax, email and KIM
- Physiotherapy and rehabilitation groups checking prescriptions and referral letters before the first appointment
- Private specialist clinics in the UK where referrals need an insurer's authorization reference before booking
- Clinic groups that suspect referrals get lost between the fax and the booking screen but cannot measure it
Questions people ask about this
Can AI read faxed referrals, including handwriting?
Mostly, yes. Current vision models read typed forms reliably and handwriting of reasonable quality well, even from low-resolution faxes. They still misread some fields, so each value carries a confidence and its position on the page, validation rejects impossible values such as a nonexistent ICD-10 code, and anything uncertain goes to staff with the image beside it.
How do I automate referral intake without making clinical decisions?
Draw the line in the design. The workflow reads, matches, checks completeness against lists the clinicians own, and chases missing information. It never judges whether a referral is clinically appropriate or how urgent it is: urgency markers go straight to a clinician, protocoling stays with radiologists, and every rule touching clinical content has a named clinician as its owner.
Is referral intake automation GDPR compliant?
It can be, and in the EU it has to be designed that way from the start. Referrals are health data under GDPR Article 9, so the build uses EU hosting, data processing agreements with every processor, a DPIA before go-live, access by site and role, and no health data in logs. Where external models are ruled out, extraction runs on self-hosted models.
Does it work with KIM and German practice systems?
KIM runs on email standards within the telematics infrastructure, so the workflow can receive referrals and doctors' letters through the group's KIM client and reply the same way. The connection to the practice or radiology system depends on its interfaces: some offer HL7 or an API, others only an import folder. That is the first thing I check, because it moves the price more than anything else.
What happens when two patients have similar names?
The workflow does not decide. It links a referral automatically only when several identifiers match exactly, such as insurance number, surname and date of birth. Anything less produces candidates for the intake team to confirm on screen, and the system never merges existing records. A wrong link is a patient safety problem, so a person makes that call.
What does referral intake automation cost?
A first version covering fax and email for a few exam types, with the review screen and the practice system connection, usually fits the multi-step tier of AI workflow automation shown on this page. Self-hosted models, a difficult practice system interface or many sites push it higher. On hosted models, running costs are cents per referral.
Sources