Key takeaways
- Referral faxes persist because sending and receiving systems differ, not because practices lack EHRs.
- AI tools classify each fax, pull out key fields and suggest a route. Staff still confirm identity, insurance and urgency.
- Set review rules before go-live: what is always checked by a person, and what can pass on high confidence.
- A referral packet is full of health information, so any vendor that reads it needs a business associate agreement.
Why referrals still arrive by fax
By 2024, 91% of office-based physicians used a certified EHR and another 4% used a non-certified one[1]. Yet referral packets still pour into fax inboxes. The reason is the mix of systems. Only 32% of physicians in solo practice used an EHR from one of the five market-leading developers, compared with 90% in practices of more than 50 physicians[2]. When a referring office and a specialist run different systems, fax is the one format both can send and receive without a project.
Source: Office of the National Coordinator for Health IT (ASTP/ONC), HHS, 2024[2]
Show the data
| Item | Value |
|---|---|
| Solo practice | 32% |
| More than 50 physicians | 90% |
That mix also shapes what AI can do for you. A large group on a leading EHR may get referral data through a built-in exchange. A smaller practice will more often be turning images of paper into structured data, which is where AI extraction tools sit.
What AI referral processing does
Capture
Incoming faxes land in a digital inbox instead of a tray. Each one gets a timestamp the moment it arrives, which becomes the start of your turnaround clock.
Classify
The tool sorts each document: new referral, records for an existing patient, prior authorization reply, lab result, or junk. Misfiled results are a safety issue, so this step needs its own check.
Extract
It pulls fields from the cover sheet and referral form: patient name, date of birth, phone, insurance, referring provider, reason for referral and any urgency marks.
Check
It compares extracted data with your rules: is the plan one you accept, is the referral complete, is the patient already in your system?
Route
It places the referral in a work queue by type and urgency, with the original image attached for staff.
Close the loop
Staff schedule or decline, and an acknowledgement goes back to the referring office.
Fields to extract and when a person checks
| Field | Where it usually appears | Always send to human review when |
|---|---|---|
| Patient name and date of birth | Cover sheet, demographics page | Two pages disagree, or the name matches an existing chart with a different birth date |
| Callback number | Demographics page | Missing or matches the referring office's number |
| Insurance and member ID | Face sheet or card copy | Plan is not on your accepted list, or the card image is unreadable |
| Referring provider and NPI | Referral form, letterhead | Provider is new to you or the NPI does not match the name |
| Reason for referral | Referral form, clinic note | Any urgency word such as urgent, stat, rule out or within one week |
| Authorization number | Referral form or payer letter | Payer requires one and none is found |
| Attachments | Remaining pages | Imaging or labs are referenced but not included |
Handwritten forms, faxes with skewed or faint pages, and packets over a set length should go to a person by default. So should anything the tool classifies with low confidence. Tune those thresholds after a month of data, not before.
Health data and the vendor
AI now plays some role in patient visits at 83% of medical groups[3], and referral processing is one of the places it handles the most sensitive data. A referral packet usually holds diagnoses, medications, notes and insurance details. A practice may let a vendor handle that information only with satisfactory assurance in a business associate agreement[4]. There is no version of fax AI that avoids health data, so there is no version without that agreement.
Ask the same questions you would ask any AI vendor: whether documents are used to train models, how long images and extracted text are kept, which subprocessors see them, and how deletion is proven at the end of the contract. The terms in our guide to what vendors do with recordings carry over almost word for word to fax images.
Routing and turnaround
Referrals are a large share of new-patient demand for many specialists. New patients made up 16.8% of physician office visits in 2019, about 173.9 million visits[5]. Slow handling costs those patients time, and the effect is not even. A meta-analysis of 34 secret shopper studies found Medicaid coverage was linked to a 3.3-fold lower likelihood of getting a specialty appointment than private insurance[6]. A referral that sits unread for a week only widens that gap.
- Tier 1, urgent marks or clinical red flags: a clinician reviews the same business day.
- Tier 2, complete and in-network: scheduling contacts the patient within two business days.
- Tier 3, incomplete: staff request missing items from the referring office the same day and set a follow-up date.
- Tier 4, out of network or out of scope: staff tell the referring office and the patient, with alternatives where you have them.
Send the referring office a short acknowledgement for every referral: received, being scheduled, or missing information. Referring offices notice who closes the loop.
Put the tier on the referral itself, not in a separate spreadsheet, so anyone opening the queue sees what is due today. Review the oldest open item in each tier at a daily huddle; a two-minute look catches the referral waiting on a missing insurance card before it turns into a complaint from the referring office.
Example: a fictional specialty practice
Example: a five-physician gastroenterology practice receives about 60 faxes a day, roughly a third of them new referrals. Before AI, one coordinator sorted the stack by hand and keyed each referral, and the backlog peaked at four business days after holidays. With AI classification and extraction, the coordinator now reviews pre-filled entries side by side with the image. In the first month the tool sent 38% of referrals to full human review, mostly for handwritten forms and missing insurance. The practice kept that share for two months before loosening any thresholds, and it measured turnaround from fax timestamp to first patient contact every week.
Where fax automation breaks
Most problems in the first months are not about reading accuracy. They come from the edges of the workflow, where a document does not fit the categories the tool was set up with. Watch these closely:
- Multi-patient faxes. One transmission holding referrals for two or three patients must be split before extraction, or one patient's data ends up in another's record.
- Duplicate sends. Referring offices often fax again when they hear nothing. Match on name, birth date and referring provider so the second copy updates the first instead of creating a new referral.
- Results mixed in with referrals. A lab report misclassified as a new referral can sit in the scheduling queue instead of reaching a clinician. Review every document classified as a result or a report.
- Cover sheets with no patient data. Some offices send the clinical note first and the demographics last. Extraction rules should read the whole packet, not just page one.
- Changes in the referring office's forms. When a large referrer redesigns its form, accuracy can drop overnight. Track accuracy by sender, not only overall.
Setting up a fair pilot
Sign the agreements
Put the BAA and data terms in place before any real document reaches the tool.
Collect a baseline
For two weeks, log each referral's arrival time, first patient contact time and any keying errors found later.
Run in shadow mode
Let the tool process the same faxes while your coordinator keeps working the old way. Compare its fields with what was keyed.
Switch on review mode
Have the coordinator work from the tool's pre-filled entries, checking each against the image, for at least a month.
Decide on thresholds
Only then decide which fields and senders can pass without a full check, and write that rule down.
Is it worth it for your practice?
- Volume: if referrals take less than an hour a day to handle, a better inbox and checklist may be enough.
- Error cost: count how often a referral was misfiled, keyed wrong or lost last quarter. That is your baseline.
- Integration: check whether the tool writes into your EHR or practice management system, or only produces a worklist staff retype.
- Review load: ask the vendor for a pilot on your own faxes and measure the share that needs full review.
- Agreements: a signed BAA and clear data terms before any real packet is processed.
Run the pilot on a copy of real traffic only after the agreement is signed, compare its output with what your coordinator keyed, and decide on the numbers. For the wider picture of where intake automation fits, see AI patient intake automation.
Frequently asked questions
Can AI read handwritten referral faxes?
Many tools can read some handwriting, but accuracy drops sharply. Send handwritten forms to human review by default and loosen that rule only after you have measured the tool on your own faxes.
Does a fax AI vendor need a BAA?
Yes. Referral packets contain protected health information, so any vendor that receives, stores or processes them on your behalf needs a business associate agreement first.
What turnaround should we promise referring offices?
Set targets by tier: same business day review for urgent referrals, contact within two business days for complete in-network ones, and same-day requests for missing items. Measure from the fax timestamp.
Will AI extraction replace our referral coordinator?
It changes the job from keying data to checking it and chasing missing items. Practices still need a person who owns the queue and the relationship with referring offices.
References
Office of the National Coordinator for Health IT (ASTP/ONC), HHS (2024). Office-based Physician Electronic Health Record Adoption, 2008-2024 (ONC Data Brief) (opens in a new tab) โ
Office of the National Coordinator for Health IT (ASTP/ONC), HHS (2024). Office-based Physician Electronic Health Record Adoption, 2008-2024 (ONC Data Brief) (opens in a new tab) โ
Medical Group Management Association (MGMA) (2026). As AI use expands, medical practice leaders shift attention to measurement (opens in a new tab) โ
Legal Information Institute, Cornell Law School (Electronic Code of Federal Regulations text) (current). 45 CFR ยง 164.502 - Uses and disclosures of protected health information: General rules (opens in a new tab) โ
CDC National Center for Health Statistics (2019). National Ambulatory Medical Care Survey: 2019 National Summary Tables (opens in a new tab) โ
Inquiry: The Journal of Health Care Organization, Provision, and Financing (via PubMed Central) (2019). Medicaid Patients Have Greater Difficulty Scheduling Health Care Appointments Compared With Private Insurance Patients: A Meta-Analysis (Inquiry, 2019) (opens in a new tab) โ