AI inbox management tools for physicians draft replies to patient portal messages, sort and route messages and summarize threads. Studies show they can reduce clinicians’ sense of burden, but so far they have not reliably saved time. Their benefit will last only if health systems also control message volume through triage, team workflows and payment.
Ask a primary care physician when their working day ends and many will describe a second shift. Clinic finishes, the last note is signed, and then the inbox: patient portal messages, test results, refill requests, forms, messages from colleagues. It is done at the kitchen table, after the children are asleep, or early the next morning before clinic.
The patient portal made this worse. A 2022 study in the Journal of the American Medical Informatics Association, led by A Jay Holmgren, analyzed electronic health record data from ambulatory clinicians in 366 US health systems using Epic. By late 2020, messages from patients had risen to 157% of the pre-pandemic average. Each additional patient message was associated with about 2.3 more minutes in the record per day.
Generative AI arrived promising relief: a model that reads the patient’s message and drafts a reply for the clinician to edit and send. Health systems rolled it out quickly.
This piece argues that AI drafting is a useful tool pointed at the wrong variable. The inbox problem is not mainly that replies are slow to write. It is that the number of messages keeps growing, and much of the work lies in reading, deciding and acting, not typing. A tool that makes each reply easier without controlling volume will be overwhelmed by the next increase in messages, and could even encourage it.
Why has the physician inbox become so heavy?
Several forces drive it.
Patient portals made it easy and free to message a clinician at any time. Many patients reasonably prefer a message to a phone queue. Immediate release of test results, required in the United States under information-blocking rules, means patients often see results before their clinician has reviewed them and ask what they mean. Care that used to happen in a visit, such as medication questions or symptom updates, increasingly happens in messages.
Meanwhile, most of that work is unpaid in fee-for-service systems. Medicare introduced billing codes for online digital evaluation, often called e-visits, but they are rarely used. An analysis in Health Affairs Scholar in 2024 found that e-visits made up about 0.05% of evaluation and management services in traditional Medicare in 2022. When a health system does not get paid for message-based care, it does not staff for it, and the work falls on clinicians’ evenings.
We have made the broader case that the EHR is where everyone else’s demands land. The inbox is the purest example. It collects demands from patients, colleagues, payers and regulators, and delivers them to the clinician without any matching change in staffing or time.
Do AI-drafted replies save physicians time?
The early evidence says: not much, but they may make the work feel lighter.
At Stanford Health Care, a study published in JAMA Network Open in March 2024 by Patricia Garcia and colleagues tested AI-drafted replies with 162 clinicians over five weeks. Clinicians used the drafts for about 20% of eligible messages. There was no significant change in time spent reading or replying to messages. But measures of task load and work exhaustion improved: task load scores fell from about 61 to 47 on the scale used.
At UC San Diego Health, a study published in JAMA Network Open in April 2024, led by Ming Tai-Seale, examined 52 physicians using AI drafts. Time spent reading messages rose by about 22%, reply time did not significantly change, and replies became about 18% longer. Physicians reported that drafts were a helpful starting point but needed editing.
Put together, the studies suggest AI drafts shift the experience of inbox work, making it less draining for some clinicians and producing longer, often more empathetic replies, without reliably reducing time. Reading and checking the draft replaces some of the effort of writing it.
That is not nothing. Reducing the cognitive burden of an unpleasant task matters for burnout. But it is not the time saving that many business cases promised, and it does nothing to the volume.
There is also a safety question. When a clinician approves a plausible draft without fully checking it, errors can pass through. We explored that risk in our piece on AI-drafted patient messages. This piece focuses on the other side of the problem: even perfectly safe drafts cannot fix an inbox that keeps growing.
Why does message volume matter more than drafting speed?
Imagine a physician receiving 100 patient messages a week, each taking three minutes of reading, thinking and replying. That is five hours. Suppose AI drafting saves 20% of that time, a generous assumption given the evidence. The physician gets back one hour.
Now suppose message volume grows 20% over the following year, as patients find the portal responsive and more results are released immediately. The physician is back to five hours. The AI has been absorbed by growth.
This is a familiar pattern. When a service becomes easier to use, people use it more. Faster, friendlier replies may make patients more likely to message again. That is not necessarily bad, since some messages replace visits and provide real value. But it means drafting tools cannot, on their own, bring inbox time down.
The organizations that have made progress treat volume as a design problem: which messages should go to a physician at all, which can be handled by nurses, pharmacists or administrative staff, which can be answered with standard information, and which should become a paid visit.
A worked example: a primary care practice’s inbox
Consider a hypothetical primary care practice with eight physicians, each receiving about 120 inbox items a week, including patient messages, results and refill requests.
Step one, the practice adds AI drafting only. Physicians use drafts for a minority of messages and report feeling somewhat less drained, but inbox time barely changes. Over the next year, patient messages rise and physicians’ after-hours time returns to where it was.
Step two, the practice redesigns the flow. AI classifies incoming messages by type and urgency. Refill requests go to a pharmacist working under protocol. Scheduling and administrative questions go to front-desk staff with AI-drafted replies. Clinical questions go first to a nurse, who resolves many and escalates the rest. Results come with AI-drafted plain-language explanations that the physician reviews and releases in batches. Messages that need clinical assessment beyond a quick answer are converted to a visit or a billed e-visit, with patients told clearly in advance.
In this version, the physician’s share of inbox items might fall substantially, and the AI’s main value is in routing and drafting for the whole team, not in writing the physician’s replies faster. The cost is staff time for nurses, pharmacists and administrators, which needs funding. That is the real price of a manageable inbox.
What is the strongest counterargument?
The strongest objection is that controlling volume means rationing access. Patient messaging lets people with chronic conditions, disabilities or inflexible jobs reach their clinicians without taking time off for a visit. Charging for messages or diverting them away from physicians could fall hardest on those who benefit most, and could push patients back to phone lines or emergency departments.
This is a serious point. Research on billing for portal messages has examined exactly this. A 2023 study in JAMA of UCSF Health’s move to bill some messages as e-visits found only a small reduction in message volume, suggesting billing does not deter most patients, but a follow-up study examined effects across demographic groups because of equity concerns.
The answer is not to suppress messaging but to route it well and fund it properly. Many messages are best answered by someone other than a physician, quickly. Some genuinely need clinical time and should be recognized as care. The goal is an inbox matched to the team’s capacity, not a smaller door.
What this means for health systems, physicians and payers
For health system leaders, treat the inbox as a service line, not a byproduct. Measure volume by message type, who handles it and how long it takes, and fund the team needed to manage it. Use AI where it helps most: classifying and routing, drafting for non-physician staff, and summarizing long threads. Judge drafting tools by after-hours time and staff wellbeing over months, not by a pilot’s enthusiasm.
For physicians, AI drafts are worth using for routine messages if they reduce strain, but read them properly. The time you save is not the point if the draft is wrong. Push for team-based inbox models, which have more potential than any drafting tool.
For payers and policymakers, message-based care is care. Payment codes that exist but are almost never used, as the Medicare e-visit data show, suggest the design or administrative burden is not working. Paying for asynchronous care in a way that is simple for clinicians and fair for patients would do more for inbox burden than any model.
For vendors, the evidence points toward value in routing, triage and team workflows, not just drafting. As with ambient scribes, time saved will be absorbed unless something deliberately protects it.
Frequently asked questions
Does AI reduce physician inbox time?
Current evidence suggests AI-drafted replies do not reliably reduce time. A 2024 Stanford study found no significant change in reply or reading time, though clinicians reported lower task load and exhaustion. A 2024 UC San Diego study found reading time rose about 22% and replies grew longer. AI’s larger potential may be in routing and triage.
How does AI draft replies to patient messages?
A large language model integrated into the electronic health record reads the patient’s message and relevant chart information, then generates a draft reply. The clinician reviews, edits and sends it, or discards it. The clinician remains responsible for the content, and health systems typically disclose or monitor how drafts are used.
Why do doctors get so many patient portal messages?
Portals make messaging easy and free, immediate test result release prompts questions, and more care has moved from visits to messages since the pandemic. A 2022 JAMIA study found patient messages rose to 157% of pre-pandemic levels by late 2020, with each additional message associated with more daily time in the record.
Can doctors bill for patient portal messages?
In the US, Medicare and many insurers have e-visit codes for certain message-based clinical assessments that require clinician time. They are rarely used: a 2024 analysis found e-visits were about 0.05% of Medicare evaluation and management services in 2022. Patients should check their health system’s policy on when messages may be billed.
What is team-based inbox management?
Team-based inbox management routes messages to the most appropriate person, such as nurses for clinical questions, pharmacists for refills and administrative staff for scheduling, with physicians handling only what needs their judgment. Combined with AI classification and drafting, it can reduce physicians’ share of inbox work more than drafting tools alone.
The inbox is the new night shift, and a faster pen does not end a shift. Only fewer letters, shared among more hands, will do that.
Where does your organization’s inbox volume go every week, and who decided that the physician should be the default destination?
