TL;DR — An endocrinology practice automates inbound fax triage by routing its existing fax numbers into an AI agent that classifies and patient-matches every document, files the high-confidence majority straight to the chart, and escalates only exceptions to staff. Done well, that removes roughly 70 to 85% of manual touches without changing the fax numbers referring providers, labs, and DME suppliers already use. The sequence that works: baseline your volume for two weeks, automate your two highest-volume document types first, run parallel with human verification, then widen scope.
Start by measuring what's actually landing on your fax line
Before you talk to a single vendor, spend two weeks counting. This is the cheapest part of the project and the part most practices skip.
Track four things: total inbound documents per day, the breakdown by document type, average staff minutes per document, and average turnaround from arrival to filed-in-chart. You can do this with a clipboard and a tally sheet if you have to.
Endocrinology practices are usually surprised by the shape of the distribution. The assumption going in is that referral packets and prior authorization paperwork dominate, because those are the documents people complain about. The count usually shows something different: outside lab results and pharmacy refill requests make up the bulk of the volume, and they're the documents nobody complains about because each one only takes four minutes.
That's the trap. Four minutes times 60 documents a day is four hours. The painful documents are painful; the boring documents are expensive.
Count seasonal variation too, if your two weeks let you see it. Endocrinology volume moves with the calendar in ways that catch practices off guard — January brings a wave of insurance changes and re-authorizations as plan years reset, and DME supplier documentation requests cluster around benefit-year boundaries. Sizing a solution against a quiet March and then meeting January with it is a preventable mistake.
The turnaround number matters for a different reason. It's the number you bring to your physicians when you need buy-in. Nobody gets excited about saving coordinator hours. People get excited when you show them a CGM authorization determination that sat in a shared folder for four days while a patient waited.
Automate your two biggest document types first
Once you have the breakdown, pick the top two categories by volume and automate only those in phase one.
This is the single most common place these projects go sideways. Practices look at their fax problem, identify the most complex and annoying document type — usually multi-page referral packets with attached records — and decide to start there because it hurts most. Six weeks later the model is still being tuned on edge cases, nobody has seen a win, and enthusiasm is gone.
Start where volume is high and structure is consistent. In most endocrinology practices that means:
- Outside lab results — A1c, thyroid panels, metabolic panels, lipid panels from reference labs. High volume, relatively predictable formats, clear destination.
- Pharmacy refill requests — insulin, GLP-1 agonists, thyroid replacement, test strips. Repetitive, well-structured, and a large share of daily count.
Prove the pipeline on those two. Once staff trust it and you have real numbers, extending to DME supplier forms, payer determinations, and referral packets is a configuration change rather than a new project.
How does the deployment actually work?
Three connections, in this order.
Your fax line routes to the AI. Your existing numbers stay live. Inbound traffic forwards to the platform instead of dropping into a shared inbox. No referring provider, lab, or supplier is notified, because nothing changes on their end. If a vendor's implementation plan involves telling hundreds of external senders to use a new number, that's a red flag and a six-month delay.
The platform connects to your EHR. This is the part with real variance. Ask specifically whether documents and extracted data get written back into the chart natively, or whether your staff work in a separate portal and re-enter information. The second pattern hands back most of what you just automated.
You set confidence thresholds. Every document gets a confidence score. Above the threshold, it files automatically. Below it, it lands in a review queue with the AI's classification and patient match pre-filled, so a coordinator confirms in seconds instead of processing from scratch.
Start the threshold conservatively — more documents in review than you think you need. Loosen it as you accumulate evidence. Practices that start aggressive and get burned by two misfiled documents spend months rebuilding staff trust they didn't have to lose.
Platforms like Honey Health's Fax Triage agent are built around this pattern: existing fax lines, native EHR write-back, tunable confidence thresholds, exception queue for the rest.
One configuration decision deserves its own thought: where auto-filed documents land. Filing a lab result directly to the chart without provider review is appropriate for routine results and inappropriate for anything flagged critical. Most practices end up with three destinations rather than two — auto-file to chart, route to provider inbox for review, and hold in the exception queue. Decide which document types go where before go-live, with a physician in the room. That conversation takes twenty minutes and prevents the argument you'd otherwise have in month two.
Run it in parallel before you trust it
For two to four weeks, the AI processes everything and a human checks its work. Both workflows run. Yes, this is temporarily more work, not less.
Do it anyway. The parallel run is where you find out that one reference lab sends results in a format the model struggles with, or that your patient-matching logic breaks on hyphenated last names, or that a DME supplier batches twenty patients into a single transmission. Every one of those is fixable in week three and expensive in month six.
Give the review a structure. Have the person checking output log not just errors, but the category of error: wrong document type, wrong patient, missed extraction, or document that shouldn't have been auto-filed at all. Those four buckets tell the vendor exactly what to tune.
Set an exit criterion before you start. Something like: three consecutive days at or above your target accuracy on the two automated document types, with no patient-matching errors. When you hit it, turn off the parallel process. Without a written exit criterion, the parallel run quietly becomes permanent, and you've bought software while keeping all the labor.
Your coordinators' jobs change, and that needs managing
The staff question is where honest and dishonest versions of this project separate.
What actually happens is not that you need fewer people. It's that the work changes character. Coordinators stop reading and indexing documents and start handling exceptions, working the denial queue, chasing authorization statuses, and doing the follow-up that has been backlogged for two years. That's a job redesign, and it deserves to be treated as one.
This matters more than it used to. MGMA polling has found medical assistants among the hardest roles to recruit, with 47% of practice leaders naming them the single hardest hire. Losing an experienced coordinator to burnout costs far more than the software.
Three things to do deliberately:
- Say plainly, early, that this isn't a headcount reduction. If staff suspect otherwise, you will not get honest error reports during the parallel run, and the parallel run is the whole ballgame.
- Name someone the exception-queue owner. Not "whoever has time." An unstaffed exception queue silently fills until someone declares the system broken.
- Show them the before-and-after numbers. People who spent two years indexing faxes deserve to see the hours they got back, expressed as work they'd rather be doing.
Where these rollouts go wrong
A short list, drawn from the patterns that repeat:
- Starting with the hardest document type. Covered above, and worth repeating because it's the most common failure.
- Dirty patient data. If your EHR has duplicate records and inconsistent demographics, patient-matching accuracy will disappoint no matter whose AI you buy. Clean that first or accept a lower automation rate honestly.
- Never ending the parallel run. No exit criterion means no end date.
- Treating it as an IT project. The people who know whether it's working are the coordinators. If the project lives entirely with IT and a vendor, nobody notices the review queue has been growing for three weeks.
- No baseline. Without the before numbers, you cannot prove the after numbers, and the renewal conversation becomes a matter of opinion.
Automation also arrives with some tailwind. The 2025 CAQH Index reported a 17% increase in administrative cost avoidance from automated transactions and found that more than 25% of provider organizations now use AI in administrative workflows. You're not the first practice to do this, which means reference calls are available. Ask for them.
Frequently Asked Questions
How long does the whole rollout take?
Plan for six to ten weeks end to end for the first two document types: two weeks of baselining, one to two weeks of technical connection and configuration, two to four weeks of parallel running, then cutover. Practices that compress this by skipping the parallel run tend to spend the saved time later, rebuilding staff confidence after an early misfile.
Do we need IT staff to manage it?
Generally no for day-to-day operation. The connection work is handled by the vendor during implementation, and ongoing tuning happens through configuration rather than code. What you do need is an operational owner — someone who watches the exception queue volume and flags when accuracy drifts.
What if our EHR doesn't have a good API?
Ask the vendor directly how they handle it before signing. Some platforms work through established integration paths, others through more manual write-back methods. The answer affects both implementation time and how native the experience feels. A vague answer here is worth treating as a no.
Will this work if our fax volume is low?
Below roughly 20 to 30 inbound documents a day, the labor math gets thin and your EHR's built-in fax module may be sufficient. The threshold isn't really volume, though — it's whether a meaningful fraction of someone's week goes to reading and indexing. If it doesn't, you have a different bottleneck.
Can we automate prior authorization paperwork the same way?
Partly. Fax triage handles the inbound side well: recognizing payer determinations and DME supplier requests, matching them to the open authorization, and routing them. Building and submitting the authorization itself is a separate workflow, and peer-to-peer reviews still need a clinician. Many practices deploy triage first and add PA automation once the fax side is stable.
How do we know it's actually working after go-live?
Track the same four baseline numbers monthly, plus two new ones: auto-file rate and exception-queue aging. A rising exception queue usually means document mix has shifted or a sending partner changed formats. Catching that in week two is routine maintenance; catching it in month four is a backlog.
