TL;DR: To automate referral intake from inbound faxes, route every fax line into an AI ingestion layer that classifies each document, extracts the referral fields, matches the patient in your EHR, and files a ready-to-schedule referral — so coordinators only touch the exceptions. The practical rollout is five steps: consolidate your fax lines, pick an extraction approach, map fields to your EHR's referral record, set a human-review threshold, and measure leakage and turnaround before and after. Done right, you automate referral ingestion from fax without replacing your EHR or your team.
Start by admitting where the time actually goes
Before you automate anything, look at what a referral costs you today. A faxed referral arrives as an image. Someone opens it, identifies the patient, reads the insurance, finds the referring provider, and types all of it into your EHR before scheduling can begin. That's five to fifteen minutes of staff time per referral, every referral, and it happens after the fax has already waited in a queue.
The queue is the real problem. When intake falls behind, referrals age, patients book elsewhere, and revenue leaks. Research on closing the referral loop has repeatedly found that a large share of referrals never complete — and a fax sitting unprocessed is one of the quiet ways that happens. Automating intake is really about killing that queue.
Step 1: Consolidate and digitize your fax lines
You can't automate what you can't reach. The first move is to get every inbound fax into a digital channel software can read. That usually means a cloud fax number or a fax-to-email bridge for each line — front desk, records, per-location numbers, all of it.
This step matters more than it looks. Groups that grew by adding locations often have a dozen fax numbers landing in a dozen inboxes with no shared visibility. Pulling them into one digital stream is what makes a single automated queue possible. Physical fax machines and paper trays have to go; a referral trapped on a printer can't be ingested.
Step 2: Choose your extraction approach
Once faxes are digital, decide how the data comes out of them. There are two broad options, and the difference is bigger than it sounds.
- OCR-only. Optical character recognition turns the fax image into searchable text. It's cheap and it's a start, but text isn't structure — a human still has to read it and decide which string is the birthdate and which is the member ID.
- AI agent extraction. An AI model reads the document, understands the fields, and outputs structured data mapped to referral fields: patient, DOB, insurance, referring provider and NPI, reason for referral. This is what actually removes the typing.
For referral intake, OCR-only rarely pays off because the goal is structured data your EHR can consume. AI extraction is the approach that gets you to a schedulable referral. This is the layer Honey Health's referral intake agent provides — reading the fax, extracting the fields, and preparing the referral for write-back.
Step 3: Map the fields to your EHR's referral record
Extraction is only useful if the data lands in the right place. Map each extracted field to the corresponding field in your EHR's referral or patient record, and confirm how the write-back happens — an HL7 or FHIR interface, or a direct integration.
Get specific here. Decide what happens when the referring provider isn't in your directory, how a new-patient referral creates a chart, and where the original fax attaches. The mapping is a one-time setup that determines whether every future referral files cleanly or generates rework. For multi-EHR groups, this is also where you normalize different systems into one consistent referral shape.
Step 4: Set a human-in-the-loop confidence threshold
Automation should handle the clean majority and escalate the ambiguous minority — and you decide where that line sits. Good ingestion systems score each extracted field for confidence and only route the low-confidence ones to staff.
Set the threshold deliberately. Too loose and errors slip into charts; too tight and your team re-reviews everything, which defeats the point. Most groups start conservative, watch what gets escalated, and loosen as they trust the extraction. The target state: a coordinator reviews the handful of referrals the AI flagged, each already populated with extracted data, instead of typing every one from scratch.
Step 5: Measure leakage and turnaround before and after
If you don't baseline, you can't prove the win. Before you flip anything on, capture two numbers: how long a referral takes from fax-received to ready-to-schedule, and how many referrals never convert to an appointment. Then measure the same numbers after.
Turnaround should drop from minutes of manual entry to roughly two minutes of automated processing per referral. Leakage should fall as fewer referrals age out in a queue. MGMA's February 2025 polling of medical groups flagged limited referral tracking as a top operational challenge — so the measurement itself often surfaces problems you didn't know you had. Those two metrics are also how you justify the spend to a partner or board.
How automation handles the messy faxes
The objection every operator raises is "our faxes are a mess." They are — and that's exactly where the design matters. Multi-page faxes get split into discrete referrals. Missing insurance triggers an exception flag rather than a bad record. A possible duplicate patient routes to review instead of creating a second chart. The AI does the reading; your team makes the judgment calls on the edge cases, with the extracted data already in front of them.
The result isn't a system that never needs a human. It's a system where humans only handle what actually needs a human — which is the whole point of automating referral intake from inbound faxes.
Frequently asked questions
Do we have to replace our EHR to automate referral intake?
No. The ingestion layer sits in front of your EHR and writes referrals back through standard interfaces like HL7 or FHIR, or a direct integration. Your EHR stays exactly as it is; automation just replaces the manual reading and typing that used to feed it.
How long does it take to roll out?
Most practices reach meaningful automation within a few weeks. The gating work is consolidating fax lines and mapping fields to your EHR — setup tasks, not ongoing effort. Once mapping and confidence thresholds are set, new referrals flow through automatically.
What happens to referrals the AI can't read confidently?
They're routed to staff as exceptions, with whatever the AI did extract already filled in. Instead of typing the whole referral, a coordinator corrects or completes the uncertain fields. You set the confidence threshold that decides what gets escalated.
Will this work across multiple locations and EHRs?
Yes, and it's often most valuable there. A single ingestion layer can pull every site's fax lines into one queue and normalize referrals across different EHRs, giving operations shared visibility instead of a dozen separate fax inboxes.
How do we prove it's working?
Track two metrics before and after: time from fax-received to ready-to-schedule, and the share of referrals that never become appointments. Faster turnaround and lower leakage are the direct, defensible signals that automated intake is paying off.
