300 Faxed Referrals a Week, All Typed in by Hand

The bottom of the callback tray

On Friday at 17:10, after the last patient has gone, Aylin clears the tray. It is a grey plastic letter tray at the front desk of the group's site in Herne, and years ago someone stuck a label on it that says Rückruf: callback. Every referral that cannot be booked yet goes into it, with a yellow sticky note saying what is missing.

At the bottom, under eighteen days of other paper, is a referral for a CT of the abdomen with contrast. The sticky note, in Aylin's own handwriting, says Kreatinin fehlt, Praxis ruft zurück: kidney value missing, the practice will call back.

She rings the patient, Renate, to apologize and offer a slot on Tuesday. Renate is kind about it. She had the scan last week at a practice across town that called her back, and her GP already has the report.

Nobody will ever count this. A referral that never became an appointment leaves no gap in the schedule, because the slot went to someone else. The GP practice will simply send its next few patients wherever they were seen faster, and nobody at the group will know why.

The group is invented: four radiology sites around Bochum, with MRI, CT, X-ray and ultrasound, and about 300 referrals a week. But anyone who runs a front desk that still takes referrals by fax knows that tray, whatever its label says.

Eighteen days earlier, 07:52

The referral arrives on a Monday, as a PDF in the shared mailbox where the fax-to-email service drops everything sent to the group's four fax numbers. It is one of about sixty that morning. Three people work that mailbox between check-ins and phone calls, and Aylin is one of them.

She opens it. A referral slip printed from the GP's software, slightly skewed, with CT Abdomen mit KM at the top and the clinical question handwritten in the box. She finds Renate in the RIS by surname and date of birth, then types in the insurance details, the referring doctor's LANR and BSNR, the ICD-10 code and the exam.

Then she checks what the exam needs. The group's protocol asks for a recent kidney function result before a contrast CT, and none is attached. She calls the practice at 08:15 and gets the engaged tone. At 09:40 a medical assistant promises to find the value and fax it over. Sticky note, tray.

The practice keeps its word. On Wednesday afternoon it faxes a single page: a patient name, a lab value, no mention of a CT. It lands in the same shared mailbox, where a colleague who works for another site opens it, cannot match it to anything in her own pile, and files it to the patient's record as a lab result. The referral in Herne stays in the tray.

Nobody did anything careless. Every step was reading, matching, typing or chasing, and not one of them was clinical. That is the part that breaks when people are busy, and it is also the part that can be automated without going anywhere near a clinical decision.

A number nobody can produce

Birgit, the group's practice manager, gets in touch because the owners have asked her a question she cannot answer: how many referrals never turn into appointments? She can count appointments. She cannot count the referrals that stall, because they live on paper, in four trays, at four sites.

We talk by video, and I ask her to share her screen and walk me through the shared mailbox. I look at the faxes on her screen rather than receive them. Referrals are special-category data under GDPR Article 9, and nothing leaves the group until there is a data processing agreement and a decision about where the system will be hosted.

What surprises me is not the handwriting. It is that the rules for a complete referral are not written down anywhere. Aylin knows that a knee MRI needs a side, that a contrast CT needs a kidney value and the allergy question, that an abdominal ultrasound means the patient has to arrive fasting. She applies all of it from memory, sixty times a morning, and so do her colleagues, each a little differently.

Near the end Birgit asks whether the system could also decide which referrals are urgent, and whether an older kidney value would do. No. Both are clinical calls. The system can make sure a clinician sees them at once, and that nothing is booked without what the protocol asks for. It will not answer either question.

Writing down what Aylin knows

So the first thing I build is not the part that reads faxes. It is a requirement list for each exam, owned by people with names: Jens, the lead radiographer, for anything that touches clinical content, and Birgit for the administrative fields such as insurance details and the referrer's numbers. The check compares each referral with its list and names exactly what is missing. It checks whether information is present, never whether it is clinically sufficient, and that sentence goes into the design document with Jens's name next to it.

Only then does a model come in. Faxes, practice emails, KIM messages and patient uploads enter one queue, stamped with the time they arrived. An EU-hosted vision model decides what each document is, a referral, a doctor's letter, a lab result or a prior report, and extracts what booking needs. Every value keeps its position on the page, so whoever checks it sees the fax itself, not just the model's reading of it.

Here is Renate's Monday fax, taken apart the new way:

The last row is the one that would have saved Renate's scan. Wednesday's fax no longer has to find the right person in the right tray. It finds the referral that asked for it. The whole design, including duplicate handling and the review screen, is in the blueprint for referral intake at clinic groups.

Two patients called Schulz

Before anyone relies on it, the intake runs over a month of old faxes that the front desk already processed by hand, so every referral has a known outcome. Most of it reads the way Aylin would have. The interesting cases are where it does not.

The RIS holds two women called Schulz with the same date of birth. On one fax the insurance number is a grey smear, and a looser link rule would have put one woman's CT referral into the other's file. That is not an admin slip; it puts one patient's history in someone else's record. So the rule stays strict: an exact match on all three identifiers or no automatic link, candidates shown side by side for a person to choose, and no merging of existing records, ever.

On a fax of a photocopy, a 1 in a date of birth reads as a 7. Validation rejects impossible values, a birth date in the future or an ICD-10 code that does not exist, but this one is merely wrong. So any field below a confidence threshold goes to the review screen with the fax region beside it. The extraction side of that is described in AI document extraction.

And one GP has written dringend, urgent, in the margin instead of ticking the box. The urgency scan reads the whole page for the terms on a list the radiologists maintain, and it is tuned to over-flag. A false alarm costs the duty radiologist a glance. A missed one can cost a patient.

The line Jens draws

The split is visible on the review screen from the first day. The model reads documents and drafts requests to referrers. Code links patients, checks the lists and raises urgency markers. People keep everything else: the front desk confirms uncertain matches and unreadable fields, radiologists review every urgent referral and protocol every exam, including whether the radiation of an X-ray or CT is justified, and patients who need sedation or an interpreter get a phone call from a person.

The line is regulatory as well as clinical. Software that ranks or selects patients for diagnosis or treatment can qualify as a medical device under the EU MDR. An intake tool that decides a referral is "fine without the side" has crossed it.

The data side is equally plain: EU hosting, a data processing agreement with every processor, a DPIA before go-live, access by site and role, and no referral content in logs or model training. Patient secrecy under section 203 of the German Criminal Code means every service provider is bound to confidentiality too. If the data protection officer rules out external model providers, extraction runs on a self-hosted model the group controls.

Another Monday, the same mailbox

The referring practices do not change how they work, and nobody asks them to. The shared mailbox still fills up overnight. But by the time Aylin sits down, most faxes are read, linked and checked. Complete referrals wait in the booking queue with exam, duration and preparation known. What reaches her is a short list: a date of birth the model was unsure of, a patient with two candidates, the first request to a new practice, which she approves before it goes. Requests only send on their own to practices whose drafts she has passed unchanged for a few weeks.

A contrast CT referral arrives without a kidney value, and before eight the practice has a short request back: which patient, which exam, what is missing and why booking needs it. When the lab fax comes back on Wednesday, to any of the four numbers, it finds its referral, the check runs again, and somebody calls the patient that afternoon.

The tray is still on the desk, because a front desk always needs somewhere to put paper. But Birgit can answer the owners now. Every referral has a status, so the ones that stall show up by practice and by exam, with how long each waited on the referrer, on the patient and on the group itself.

If your Monday mailbox looks like Herne's

In the US, look at Tennr first; it raised $101M in June 2025 to automate exactly this intake. Ask your RIS vendor, too, what document features your licence already includes. A build earns its place in a European group with several sites, several exam types, German or French referral forms, KIM, and health data that has to stay in the EU.

I would start narrow: fax only, three or four exam types, and the rules the front desk already knows, written down and owned. The first thing I check is the practice system's interface, an API, an HL7 feed or only an import folder, because it moves the effort more than anything else. A first version is usually a multi-step AI workflow automation project. The full referral intake blueprint covers the rest, and once referrals turn into bookings reliably, the same group can fill late cancellations with a waitlist agent.

Frequently asked questions