The fax is still the front door

The channel that would not die

Every few years a federal official announces the end of the healthcare fax. The 2025 hospital survey is the most recent data point against that announcement. Among 2,351 non-federal acute-care hospitals, 40% said they often send a summary of care by mail or fax when a patient leaves — up from 30% in 2023, and identical to 2018. Another 34% do it sometimes. On the receiving side the picture is worse: 81% of hospitals still take summaries by mail or fax at least some of the time (35% often, 46% sometimes). (ASTP/ONC, AHA IT Supplement 2025)

This is not a story about hospitals failing to modernise. In the same years, the share that often send through a national network (Carequality, CommonWell, eHealth Exchange) rose from 20% to 60%. EHR-vendor networks went from 30% to 55%. Direct Secure Messaging is in wide use. Fax did not lose; it kept the cases the other pipes do not cover — the independent specialist, the imaging centre, the skilled nursing facility, the payer portal that still wants a cover sheet.

For an independent pain-management practice the inbound mix is even more one-sided. Referring primary-care offices, hospital discharge planners and occupational-medicine clinics are not on the same EHR. They dial a number. The document that comes off the line is the new patient.

What the queue actually costs

The cleanest published measurement is from a health system that built the reader itself. At the University of Pennsylvania Health System, 8,000 to 9,000 faxes arrive every day. Staff were spending about two minutes filing each one into the chart. A homegrown platform (coordn8) that classifies the document, identifies the patient and indexes it into the EHR brought that to about 40 seconds. Staff satisfaction with the intake process moved from 35% to 60% in the first two weeks of the pilot. The NEJM Catalyst paper reports more than 370,000 faxes processed and more than 8,500 staff hours returned. (Daniel et al., NEJM Catalyst, 2025; Penn Medicine, 2025)

Two minutes is the filing step at a well-staffed academic system with a single EHR. At a specialty office the inbound page is often a referral: demographics, insurance, the reason for consult, sometimes an MRI report, sometimes a medication list, sometimes three of those stapled to a cover sheet with the wrong patient on page four. Vendor operations teams that sit in that queue put the manual handling of a complex referral at 8–15 minutes. That is not a peer-reviewed figure, and we will not treat it as one. It is consistent with what the front desk describes.

The cost is not only labour. It is the referral that sits in the pile until Thursday, by which time the patient has booked somewhere else — or not booked at all.

Half of them never become a visit

The Institute for Healthcare Improvement, drawing on Weiner and colleagues, put the national picture simply: more than 100 million specialist referrals a year in the ambulatory setting, and only about half completed. (IHI/NPSF, 2017)

The underlying measurements are older than the fax debate, which is itself a finding: the conversion problem predates every current EHR. In a large urban ambulatory network, 71% of specialty-consultation orders produced an appointment and 70% of those appointments were kept — a 50% completion rate from the original order. Scheduling itself ranged from 12% to 90% by specialty. (Weiner et al., 2010) A companion study of 40,487 referrals found that replacing faxed paper with a shared web queue and automated notification moved the share that got scheduled from 54% to 83%, and cut median time-to-appointment from 168 days to 78. (Weiner et al., JGIM, 2009)

Closing the loop is stricter than booking the slot. In a 2018 analysis of 103,737 primary-care referral attempts, only 34.8% resulted in a documented complete appointment — defined as the specialist sending a report back to the referring clinician. (Patel et al., JGIM, 2018)

Three failure points, in order:

1. The page is not read in time. A faxed referral that waits is a patient who calls the office that sent them, hears nothing, and stops.
2. The chart is not ready when they arrive. The demographics were retyped; the MRI was in the pile labelled “records”; the insurance card was a photograph of a photograph. The first visit is spent reconstructing the referral.
3. Nothing goes back. The consult note never returns to the referring clinician, so the next referral from that office goes somewhere that faxes a thank-you.

AI does not create demand. It attacks (1) and (2) directly, and it makes (3) a background job instead of a memory.

What a reader is for — and what it is not

The useful shape of the work is not “an AI that triages referrals.” It is a pipeline that turns an unstructured transmission into a reviewable draft, with a person still on the last step that writes to a chart or a schedule.

Read, classify, match

The inbound document is stored, rasterized, and read. Classification is a constrained label set — referral, lab result, imaging report, insurance correspondence, clinical note, advertisement, other — not a free-text guess. Patient matching uses the identifiers the page actually contains (name, date of birth, phone, member ID) against the panel, and returns a confidence with candidates rather than a silent attach. Low confidence stays in a queue. A new patient is a draft until staff or the patient confirm identity.

Draft the chart; do not file it

Problems, medications, allergies, the referring provider, the reason for consult can be extracted into draft sections sitting next to the original page. The clinician or the front desk accepts, edits or rejects. Writing into an existing chart without that step is how a mis-matched fax becomes a wrong-patient event. Scheduling from a fax without that step is how a wrong number gets a reminder about someone else's injection.

The patient can confirm the easy parts

Demographics and insurance pulled from a referral are things the patient already knows. A portal prompt that asks them to confirm or correct those fields is a strong matching signal and a reduction in staff re-keying. It is not a substitute for clinical review of the referral itself.

Notice what is missing from that list: diagnosis, treatment, “this patient should be seen urgently.” Those are clinical judgements. Software that makes them, even helpfully, is in a different regulatory category — see what “human in the loop” actually means.

The other door: the order that leaves the room

Intake is half the fax problem. The other half starts when the clinician signs an imaging study, a procedure or a medication that the patient's plan will not pay without a fight.

Only 35% of medical prior authorizations ran fully electronically on the X12 278 in the 2024 CAQH Index — twenty-plus years after that transaction was the HIPAA standard for this exact job. The rest are portals, phone and fax. CAQH's 2025 figures put the fully-electronic share at 38%, with a 15-minute saving per request against the manual path. (CAQH CORE, citing the 2024 Index; X12, citing the 2025 Index)

The AMA's 2025 physician survey still finds 40 prior authorizations per physician per week and 13 hours of physician and staff time completing them, most commonly by telephone for medical services. That is a separate essay (Thirteen hours a week). The design implication here is narrower: the authorisation should not be a project someone starts after the patient has left. It should be a background check that fires when the order is signed. The clinical half of that argument — the note has to contain the requirement while the patient is still here — is The denial is prevented in the room.

A workable order loop, for a specialty practice, looks like this:

1. Coverage first. When the order is created, ask whether this patient's plan requires prior authorization for this CPT or HCPCS. If the payer will not answer, fall back to a practice-owned list (advanced imaging, interventional pain procedures, the medications this office actually fights over). Unknown is unknown — it is not a green light.
2. Assemble from the chart, do not invent. The packet is diagnoses, the procedure, the coverage on file, and the clinical facts the note already contains. It reports what the chart documents. It does not conclude that medical necessity is met. That sentence is how a drafting tool becomes a False Claims problem.
3. Submit the administrative request; keep a human on cancel. Where a 278 (or, later, a FHIR prior-authorization API) is available, sending the coverage request can be automatic. Staff can still stop it. Therapy and DME stay tracking-only: the rendering therapist and the supplier are the legal submitters, and a referring pain practice cannot produce the documents those packets require.
4. Pick the destination from a roster, not a guess. A 271 eligibility response says whether this office is in-network for an E/M visit. It does not say which MRI shop the member can use. Commercial, Medicare Advantage and TennCare directories are not a BAA-clean feed a small practice can call. The office already knows the centres it sends people to; writing that roster down, marked in- or out-of-network per payer, is what lets the requisition go to a place the plan will pay. Original Medicare is the exception — enrolled facilities participate as a matter of the program. Absence of both is unknown, never guessed in-network. The clinician can still send out of network.
5. Send the requisition on the same channel the destination uses. For most imaging centres that is still a fax. The difference is that the cover sheet is assembled from the signed order, the chosen destination and the patient's coverage, rather than typed from a sticky note.
6. Close when the result lands — and when it does not. Matching a report to the order that produced it is deterministic and should not wait for a person to notice. An order that is still open past a deadline is a case, once, not a recurring inbox item.

The loop back to the person who sent them

A referral that converts is not finished. The IHI nine-step closed loop ends when the referring clinician has the consult findings and the patient has a plan. That is the step independent specialists lose most often, because the referring office is on another EHR and the natural transport is — again — a fax of the signed note to the number on the original cover sheet.

If the inbound read already captured the referring provider's name, NPI and fax, sending the consult back is a background job with a human checkpoint, not a separate project for the medical assistant. The practices that do this reliably are the ones the next referral goes to. The 34.8% “report returned” figure is the measurement of how rarely it happens today.

Where this is actually AI, and where it is plumbing

Worth being exact, because vendors will not be. Classification, extraction and matching against a panel are model-shaped. Routing a labelled document to a queue, assembling a 278 from structured fields, picking a roster row that is in-network and has a fax number, and marking an order complete when a report cites it are not. They are the unglamorous work of making two records agree. Calling the second group “AI” does not make it better or worse. It does make procurement conversations dumber.

The FDA line is the same one we have drawn elsewhere. Drafting a chart section from a fax is support. Auto-filing it is a clinical action. Recommending a destination from a participation roster is not a treatment suggestion; auto-picking “the right MRI” on clinical grounds would be. Prior-authorization submission of a coverage request is administrative; a generated paragraph that asserts the patient meets criteria is not.