Patient Portal Message Triage: Take Back the Clinician Inbox

Your last patient leaves at 5 p.m. and the portal inbox is still full. That inbox is a second workday nobody scheduled, and it keeps growing. Patient-written portal messages climbed from 0.99 to 2.5 per patient per year between 2020 and 2025, a 153% increase, according to a JAMA analysis of Epic records from 2,067 hospitals and 47,100 clinics.

Advice on this problem comes from two places, and neither one is written for you. Academic health systems publish research on AI-drafted replies piloted across 162 clinicians. EHR and messaging vendors sell inbox features. A 3-provider independent practice gets nothing it can act on next Monday.

This post gives you the triage model instead: classify every message by what it is asking, route it to the queue that owns that category, let software draft only the routine replies, and keep a licensed clinician in front of anything clinical. Most of that inbox never needs to reach a provider at all.

Why the portal inbox became a second, unpaid workday

The portal inbox grew without a staffing plan attached to it. Message volume rose 153% between 2020 and 2025 while office visits rose 17%, so messaging added work instead of replacing visits. Nobody budgeted hours for it, nobody assigned an owner, and the overflow landed on whoever opened the chart last.

The volume is not a local complaint. In an MGMA poll of 223 medical groups, 70% reported portal message volume increasing during 2024. Active portal users across that same JAMA dataset grew from 94.3 million to 140.5 million, and telephone encounters fell only about 6%, so the messages are new work rather than migrated work.

The cost of that work is measurable. A study in the Journal of the American Medical Informatics Association found that primary care clinicians in the top quartile of message volume, above 307 messages per clinical full-time equivalent per week, had 6.17 greater odds of high exhaustion than clinicians in the lowest quartile. That study ran on 2020 data, so read it as a direction rather than a current benchmark.

Here is the reframe that matters for a small practice. This is a routing and ownership problem long before it is a technology problem. Your inbox is not full because you lack a tool. It is full because every message, regardless of what it asks, currently arrives in the same place with the same owner.

The triage model that keeps most messages off the clinician's screen

Triage means every message gets classified before anyone reads it in full, then routed to the queue that owns that category. Scheduling goes to the front desk, refills follow a clinical staff protocol, billing goes to billing, and only genuinely clinical questions reach a provider. Software drafts the routine replies, and a person still sends them.

Classification and routing come first. Drafting is the fourth step, not the first.

The delegable share is large. The American Medical Association puts roughly 75% of portal messages in the category office staff can answer, from a question about an upcoming appointment to a request for a fax number. Treat 75% as an expert estimate, not a measured constant, then check it against a sample of 50 of your own messages.

Publish those windows to patients inside the portal. A large share of duplicate messages exists because the patient has no idea when an answer is coming, so they send a second one. A stated window costs you nothing and removes a chunk of volume.

What triage actually changes

Triage moves the sorting work off the clinician and onto a rule set. The strongest published version of this is unglamorous: an academic internal medicine clinic wrote message standards and a routing guide, and portal volume fell 16% more than at control clinics over the following 4 months. No model, no vendor, just written ownership.

Triage moves sorting off the clinician and gives every category an owner.

The numbers at that site are worth repeating. Volume dropped from 1,342 to 954 messages per physician clinical full-time equivalent per month, roughly 388 fewer messages. Carbon-copy messages, the ones staff send to several people at once, fell 65%, from 4.4% of volume to 1.5%. A meaningful share of inbox noise is generated inside the practice, not by patients.

Triage also gives you something you cannot get from an untriaged inbox: numbers. Once every message carries a category and an owner, you can see which category is drowning you, how long each one takes to answer, and how much of it happens after the last patient leaves. That is the same visibility problem the wider front-office automation stack solves for phones and intake.

Where AI drafting belongs, and where it must stop

Software can classify a message, sort it by priority, pull the right schedule or account detail, and write a first draft. It must not decide whether a symptom is serious, and it must not send a clinical reply on its own. A licensed clinician reads and approves every word that answers a medical question.

Patients hold the same line. Researchers interviewed 40 patients at a large academic health system for a 2026 JAMA Network Open study and found that acceptance of AI-drafted replies was conditional: patients welcomed faster answers, but only if a clinician read every word and stayed accountable for it. Disclosure and review are the price of using the tool at all.

Priority sorting is the safest early win, because a ranking model changes reading order rather than clinical content. At NYU Langone, a model built to flag high-acuity messages reached a 97% C-statistic, and flagged messages were read about 21 minutes sooner outside business hours, a median of 42 minutes instead of 63. Precision was 67% and sensitivity 63%, so it sorts a queue, it does not make a clinical judgment.

If you have not evaluated this category of tool before, start with what an AI receptionist does and where it stops, because the same boundary applies on the phone and in the portal. The documentation side of the same burden runs through ambient AI scribes.

What the published studies actually show

The evidence on AI-drafted portal replies is real but modest, and it comes from large academic systems rather than small independent practices. Clinicians used the drafts roughly 20% of the time. Reply times barely moved. What did move was how heavy the work felt, which matters if your problem is keeping the people you have.

Published figures, each with its source. Reasons to measure your own inbox, not your result.

Stanford ran a 5-week pilot across 162 clinicians in 2023. Draft uptake was about 20%, and read, write, and reply times showed no significant change. Reported physician task load fell 13.87 points and work exhaustion fell 0.33 points, both statistically significant. The tool did not make anyone faster. It made the work feel less punishing.

NYU Langone measured the same thing at larger scale, across more than 55,000 messages between October 2023 and August 2024. Clinicians chose to start from a shown draft in 19.4% of cases, and median reply time was 331 seconds with a draft against 355 seconds without, about 7% faster. Review and editing ate much of the saving.

Why billing for portal messages is not the fix

Billing looks like the obvious answer to unpaid inbox work, and it mostly is not. The online digital evaluation and management codes cover cumulative clinician time on a patient-initiated medical question over a 7 day period, for established patients only. Most of your inbox does not qualify, and the national numbers show it.

Those codes are tiered at 5 to 10, 11 to 20, and 21 or more minutes, and what they exclude is exactly the volume that hurts: scheduling, billing questions, and nonevaluative result notifications are not separately reportable. The scale confirms it. A claims analysis covering more than 270 million individuals found asynchronous e-visits account for roughly 0.1% of all office visits, and MGMA guidance describes billable messages as often under 1% of total message volume.

Billing does one useful thing, and it is not revenue. It signals to patients that clinician time has a cost, which reduces volume. Mayo Clinic reported an 8.8% decrease in portal message threads across 6 months of billing compared with the same period a year earlier, and UCSF Health saw volume drop after letting clinicians decide case by case what qualified. Treat that as demand management, not as a revenue line.

Route one message type this week

Pick the single highest-volume category in your inbox, usually scheduling or refills, and give it an owner, a written rule, and a response window. Run it for 30 days and count how many of those messages still reach a provider. One category handled properly beats a full rollout that nobody maintains.

Route by what the message asks. Every message ends with an owner and a response window.

Walk the flow once. Anything clinical or symptom related goes to a clinician, where software may draft but never sends. Scheduling and administrative messages go to the front desk queue, where automation can answer inside your rules. Billing questions go to billing. Everything left over goes to a human for manual triage, and that leftover pile is your next rule.

Measure three things and nothing else: the share of messages that reach a provider, the median time to a first response by category, and the minutes of inbox work happening after the last patient leaves. If those three numbers do not improve in 30 days, the rule is wrong, not the idea.

If you would rather see where the time and the revenue are leaking before you change anything, the free Growth Leak Audit sizes it from your own numbers, no tooling conversation required.