“I Use ChatGPT Between Patients” Is Not a Workload Plan

An outpatient psychiatrist finishes the last morning slot. Between patients they open a consumer chat tab, paste a portal message and a letter request, and paste the output back. The afternoon looks the same. The chart does not show a new FTE, a new inbox owner, or a new clock. The work moved. It did not leave.

This fictional composite is not a patient case. It is a job without a plan.

A tab can produce text. It cannot allocate the work.

Sources reviewed August 18, 2026: one National Academy of Medicine consensus report, two ambulatory time-motion and EHR-log studies, and one quality-improvement study of generative-AI drafts inside an EHR inbox. None of those four is a psychiatry sample. None is a trial of a consumer chat tab. This is not a burnout-rate article, not a tool recommendation, and not a time-saved claim.

This article is not about turning on an AI scribe or mapping audio, transcripts, and vendor logs. That is Issue 2. It is not about whether using a model is a documented clinical act. That is Issue 8. The composite above has a portal message pasted into a consumer tab. Whether that paste is a permitted disclosure is not decided here, and Issue 8 does not decide it either; Issue 8 separates the vendor-under-a-BAA case from the consumer tab without classifying either one. The question here is narrower: once a clinician opens a chat tab between patients, what work has actually been planned?

The National Academies’ 2019 consensus report, Taking Action Against Clinician Burnout, treats workload as a work-system problem. Recommendation 1C tells health care organizations to accurately assess total workload and the complexity of the work expected of clinicians, including continuing professional education, maintenance of certification, required institutional learning modules, and work performed outside of scheduled hours. The same recommendation tells organizations to obtain clinician and patient input to identify and eliminate documentation and other administrative burdens that are not mandatory and contribute little or no value to patient care.

Chapter 4 of that report, on time pressure and encroachment on personal time, states that an honest accounting of the collective amount of work being done by health care professionals both on and off the clock, along with a recalibration of a sustainable cumulative work week, is long overdue.

That report is not a paper about large language models. It is a paper about whether the work is counted.

Two ambulatory studies measured the work that sits between visits.

Arndt and colleagues, in Annals of Family Medicine in 2017, used three years of Epic event logs from 142 family medicine physicians in one southern Wisconsin system, validated with direct observation. Per weekday per 1.0 clinical full-time equivalent, those physicians spent 355 minutes (5.9 hours) of an 11.4-hour workday in the EHR: 269 minutes (4.5 hours) during clinic hours, defined as 8:00 a.m. to 6:00 p.m. Monday through Friday, and 86 minutes (1.4 hours) after clinic hours. The after-hours figure includes 51 minutes of weekend EHR time allocated across five weekdays. Inbox management accounted for 84 minutes (23.7 percent) of daily EHR time. Clerical and administrative tasks accounted for 157 minutes (44.2 percent). The authors wrote that most primary care physicians have not allocated time for this additional work, and that much of the non–face-to-face work occurs on top of already full patient care sessions.

That sample is family medicine in one system. It is not psychiatry.

Sinsky and colleagues, in Annals of Internal Medicine in 2016, directly observed 57 U.S. physicians in family medicine, internal medicine, cardiology, and orthopedics for 430 hours. During the office day, physicians spent 27.0 percent of total time on direct clinical face time with patients and 49.2 percent on EHR and desk work. While in the examination room, they spent 52.9 percent of the time on direct clinical face time and 37.0 percent on EHR and desk work. Twenty-one of the 57 physicians completed after-hours diaries. The authors’ conclusion: for every hour physicians provide direct clinical face time to patients, nearly 2 additional hours is spent on EHR and desk work within the clinic day; outside office hours, physicians spend another 1 to 2 hours of personal time each night doing additional computer and other clerical work. Among the 21 diaries, mean after-hours work was 1.5 hours per day, 59 percent of it on an EHR; on-call nights, 2.2 hours, 69 percent on an EHR.

That sample is not psychiatry. The after-hours numbers are a diary from 21 of 57 physicians in self-selected, mostly high-performing practices.

The two studies do not give a psychiatry pajama-time rate. They do show that inbox, portal, letter, and after-hours EHR work already have a clock. Opening a second window does not create one.

NAM asks for an honest accounting of work on and off the clock. A consumer chat tab, by itself, creates no scheduled slot, no FTE, and no inbox owner.

Arndt’s family-medicine logs put inbox work at 84 minutes a day, including portal messages, refills and results, telephone encounters, and letter generation. That time was already in the EHR. It was not leftover minutes waiting for a tab.

A chat tab opened between patients does not appear in those logs. It does not appear on a schedule. It does not replace the 86 after-hours minutes, or the 1 to 2 personal hours in the Sinsky diaries. It is another place the same work can go.

This is editorial, not a statute. The tab is not the NAM assessment. It is not allocated time.

Tai-Seale and colleagues, in JAMA Network Open in 2024, studied generative-AI draft replies inside the EHR inbasket at one academic system. Eligible message types were refills, results, paperwork, and general questions. Fifty-two family medicine and general internal medicine physicians were randomized to immediate or delayed activation; 70 physicians in the same departments who did not join the pilot were contemporary controls. The study examined 10,679 replies.

The a priori hypothesis was that drafts would be associated with less physician time spent reading and replying. The estimated association of the drafts was a 21.8 percent increase in read time (95 percent CI, 5.2 percent to 41.0 percent; P = .008), a −5.9 percent change in reply time (95 percent CI, −16.6 percent to 6.2 percent; P = .33, not statistically significant), and a 17.9 percent increase in reply length (95 percent CI, 10.1 percent to 26.2 percent; P < .001).

The authors’ meaning statement: generative AI was not associated with reduced time on writing a reply, but was associated with longer read time, longer replies, and perceived value in making a more compassionate reply. They note that the uptick in read time may be attributable to the need to read both the patient’s original message and the draft reply.

That study is primary care. It is an EHR-native inbox draft. It is not a consumer ChatGPT study, and it is not psychiatry. Perceived empathy is not treated here as time saved. No vendor is recommended.

If a draft sitting in the inbasket was not associated with reduced reply time, a draft sitting in a personal chat tab is not a workload plan.

Operational questions, not legal conclusions.

A plan would name five things. The companion worksheet records the same five. The tab itself names none of them.

NAM Recommendation 1C tells organizations to accurately assess the work expected of clinicians, including work performed outside of scheduled hours. Work that is assessed is work that is named. A tab opened between patients has no named job. It has whatever was pasted last.

A named job comes with a rule about what goes into it. A refill queue takes refills. A letter template takes letter requests. A blank text box takes anything, and anything is not a rule. This is editorial, not a measurement: work that can be anything cannot be counted. Whether a given paste is a permitted disclosure is a separate question, and this article does not answer it.

Arndt: most primary care physicians have not allocated time for this additional work. Unallocated time does not remove the owner; it leaves the clinician as the owner by default. A plan makes that ownership explicit. A tab does not allocate the work, and it does not name who answers for the output.

Tai-Seale measured drafts that landed inside the inbasket, next to the message each one answered, and reply time still did not go down. A tab’s output lands nowhere in particular. It is pasted onward, into an inbox, a letter, or the chart, and the tab does not record where.

Sinsky’s conclusion language: another 1 to 2 hours of personal time each night on additional computer and other clerical work. The diaries, from 21 of 57 physicians, put the mean at 1.5 hours. That clock exists and is measured. A tab that is supposed to replace a clock has to name one. A tab does not.

NAM’s recommendation is to assess the total work and to eliminate burdens that contribute little or no value. A tab that adds a draft, and then adds the time to read and edit that draft, has not done that assessment.

This article does not prescribe a tool.

Operational failures, not universal legal conclusions.

NAM asks for an accounting of work on and off the clock. A tab is not that accounting.

Tai-Seale: not associated with reduced time on writing a reply. Read time went up.

Tai-Seale measured drafts inside an inbasket, for four message types, in family medicine and general internal medicine. It did not measure a personal chat window.

Arndt is one Wisconsin family-medicine system. Sinsky is four specialties, none of them psychiatry, with after-hours diaries from 21 of 57 physicians. Neither number is a psychiatry rate.

None of the four sources is a trial of a consumer model as a workload intervention. Do not treat the tab as one.

Download the Between-Patients Workload Record (PDF) (371 KB)

Companion worksheet: present / absent for job named, input rule, output owner, landing place, and clock replaced. The worksheet does not authorize a tool.

If any answer is unknown, do not treat the tab as a workload plan.

Educational Disclaimer: The Psychiatric Record provides general educational information for psychiatric and mental-health professionals. Content does not constitute medical, legal, regulatory, compliance, billing, or other professional advice; does not establish a standard of care; and is not a substitute for independent professional judgment. Requirements and appropriate practices may vary by jurisdiction and circumstance. Verify current authoritative sources.

This article does not recommend a tool, authorize a workflow, or determine whether a particular use of a model is permitted.