Referral Letters, Faxes and Waiting Lists: AI for Specialist Practices

Every specialist practice has a pile. In a small cardiology, dermatology, or orthopaedic practice it might be a fax tray, a shared inbox, a portal, or all three. Each item in it is a referral: a letter from a GP, a scanned form, sometimes a lab report with a sticky note that says "please see". Somebody at the front desk reads each one, finds or creates the patient, checks what is missing, puts it in front of a clinician for urgency, and eventually books an appointment.

On a quiet day that works. On a normal day the pile grows faster than one person can read it, and the scary part is not the delay itself. It's that the referral for the patient with a suspicious lesion or a worrying ECG looks exactly like the other thirty until somebody reads it.

This is a good place for AI in a specialist practice, with one firm rule that I'll repeat throughout: software can read, match, and surface. It does not decide who is urgent.

A process that loses patients in transit

The referral process has been studied for a long time, and the results are sobering.

The first number is the one I keep coming back to. In a large, well-resourced health system, only about a third of referrals made it all the way through to a completed specialist visit with a report back to the referring doctor. Some of the rest are patients who decide not to go, which no software fixes. But many are lost in handoffs: a fax that nobody matched, a referral missing information that nobody chased, a patient who was never called.

The third number describes the other half of the problem. The referring doctor believes the clinical question was sent. The specialist believes it never arrived. Both can be right, because the information got lost in a scan, a cover sheet, or an attachment nobody opened.

What the front desk actually does with each referral

When I map a process like this, I start by writing down every piece of information someone extracts by hand, and who decides what happens with it.

Most of the rows are information work that a language model is well suited for: reading unstructured letters, including poor scans, and turning them into fields. The last row is different. It is clinical judgment, and it stays with a clinician.

The intake pipeline I would build

Here's the shape of it. The practice management system or EHR stays the system of record. The pipeline sits in front of it and prepares work for people.

A few design rules make the difference between a helpful pipeline and a risky one.

The model may raise a flag, never lower one. If the model spots a phrase from the warning list, the referral moves up for human review. If it sees nothing, the referral keeps its normal place in line. It is never pushed down because the model "thought" it was routine. That asymmetry is the whole safety design in one sentence.

The original document is always one click away. The clinician reviews a summary, but the source letter sits next to it. Summaries are for speed, not a replacement for reading when it matters.

Near-matches go to a person. Two patients with the same name and a similar date of birth are exactly where automated matching causes harm. The pipeline proposes, a person confirms.

Completeness, specialty by specialty

The completeness check only works if "usable referral" is defined by the people who use it. That definition is different for every specialty, and writing it down is often the most valuable hour of the whole project. Here is the kind of table I would draft with the clinicians, as a starting point for them to correct.

Two things happen once this table exists. The front desk stops guessing what a clinician will want, and referrers get specific, answerable requests instead of a generic "please send more information". Over time, the practices that refer to you most often start sending complete referrals because they know what you will ask for.

Closing the loop costs one message

The cheapest improvement in the whole pipeline is the acknowledgement at the end. A short message to the referring practice, "we received your referral for this patient, they will be contacted within five working days", removes the "did you get our fax?" calls. A similar message to the patient cuts the "I was told to expect a call" calls.

It sounds trivial. In most practices I would expect it to remove a noticeable share of inbound phone traffic, and it directly addresses the problem in the stats above: referrals that disappear without anyone noticing.

Where the regulatory line sits

Referrals are health data, so the usual rules apply: a Business Associate Agreement for US practices under HIPAA, a data processing agreement and an agreed data location for EU practices under the GDPR, where health data is a special category.

There is a second line that is easy to miss.

That line also happens to be good design. Clinicians trust a tool that shows them the letter faster. They don't trust one that tells them what to think.

What changes for the team

A pilot that runs alongside the pile

The numbers to compare at the end are time from arrival to clinician review, time from review to booking, and the share of referrals that needed a second round with the referrer. If the first number doesn't drop, the pipeline is not worth keeping, whatever the demo looked like.

Questions practices ask me

Can it read faxed referrals?

Yes, within limits. Current text recognition handles typical fax quality well, and a language model is good at pulling fields out of letters that all look different. Poor scans are flagged as uncertain rather than guessed.

Does it work with our EHR?

Most EHRs and practice management systems offer some interface, often HL7 or FHIR, sometimes only a document import. If there is no usable interface, the pipeline can still run as a work queue next to the EHR, with staff copying confirmed data across. That is less elegant but still much faster than reading from scratch.

Could the AI decide urgency on its own later?

I wouldn't build that for a small practice. It moves the software into medical device territory and puts a model's judgment between the referral and the clinician. Highlighting agreed warning signs gets most of the benefit with a fraction of the risk.

What about referrals with no clinical question at all?

Those are exactly the ones the completeness check catches. The pipeline drafts a short request to the referrer for the question and any missing results, and a person sends it.

The one-sentence version

Let software read the pile and let clinicians decide what matters: the model may raise a flag, never lower one.

If your referral inbox has become a daily worry, describe how referrals reach you today and I'll tell you what a small first step would look like. The same pattern of "read, structure, let a person decide" shows up in other practices too, from dental practices' phone lines to home care scheduling.