Who gets the time back
Most hospitalists already use AI, and Epic is bringing more of it to the wards. The minutes it saves are few, and they belong to patients and doctors, not to a higher census.
Listen to this special topic
read by an AI voiceWhen Sutter Health put AI scribes into 20 of its California hospitals, the doctors in a study of the rollout, half of them hospitalists, used the tool for only about 3% of their notes. The time they spent on notes did not change. Before they started, more than half thought it would let them see one more patient a shift. Afterward, fewer than half did.
The study, published in September, is a fair guide to what is coming. Epic's records cover most of America's hospital beds; more than 85% of its customers already use its AI, which can draft hospital courses and summarize charts, and its listening tool, which drafts notes and queues orders, is offered for inpatient care. Tools for hospitals to build their own AI agents are due to become widely available in 2027, as is a model that forecasts how long patients will stay. The time these tools save is real but small. It belongs to patients and to doctors' evenings, not to a higher census, and hospitalists must claim a say before the tools go live.
The evidence on time is consistent. In one of the largest studies, of some 8,500 clinicians in academic clinics, AI scribes saved about 16 minutes of documentation for every eight hours with patients, about one extra visit every two weeks, and after-hours time in the record did not change significantly. A small pilot among hospitalists found no measurable change in time spent documenting or at the bedside. And drafts need checking: AI-written discharge narratives had more errors than physicians' own, though the potential for harm was low for both, and Epic's hospital-course drafts, though more complete, made up more details.
The push to turn those minutes into patients has begun: as AI spreads, one physician in five already reports higher expectations for productivity. Hospitalists have little slack. In one survey, those outside academia averaged nearly 17 encounters a shift but called about 15 reasonable, and in one study, length of stay rose on busy days once a hospitalist's census passed about 17. The Society of Hospital Medicine told the federal health department that time saved by AI belongs at the bedside, and that success should not be measured in "higher encounter quotas."
The strongest objection comes from those who run the beds: hospitals are full and short of money, so if software saves time, some of it should come back while patients wait in the emergency department. But the time is not there to take. A quarter of an hour a shift does not make room for an admission; beds and discharges do, and in one hospital, AI's predicted discharge dates were right to within a day less than half as often as case managers' the day before discharge. And the tools can pay for themselves without more patients: one health system says higher billing more than covered its scribes.
Hospitalists have more standing than they may think. The American Medical Association says the medical staff should help choose and implement AI "at the outset," and guidance from the Joint Commission and the Coalition for Health AI asks hospitals to govern these tools, test them locally and keep monitoring them. Hospitalists should take seats on their hospital's AI committee, or start one, and insist on a pilot on their own service before go-live, judged by local measures: errors in drafts, time after hours, length of stay. Local testing matters: even Epic's newer sepsis model varied widely between health systems and raised many false alarms. And they should demand training; only a third of hospitalists who use AI say they were adequately trained.
Replacement is not the near-term threat. In the latest national survey, most hospital medicine groups expected to grow, and when one physician group cut 177 jobs in a move described as largely driven by AI, the cuts fell on its billing office, not its doctors. The slower risks are census creep and coverage moved off the ward; remote hospitalists let some veterans' hospitals cut on-site clinicians. Hospitalists should write workload into their contracts, with a census limit and a rule that AI's time savings are not turned into quotas without their agreement, and claim the work software cannot do: the sickest admissions, family meetings, procedures, and checking the tools. In September, doctors at two Allina hospital campuses in Minnesota, hospitalists among them, reached a tentative first contract after a four-day strike in which AI was among their worries.
The doctors in Sutter's study wrote few of their notes with the scribe, and fewer of them came away thinking it would buy them another patient a shift. Hospitals should run the same experiment on their own wards before they write that extra patient into anyone's targets.
- A study in the Sept. 14, 2026, Journal of Hospital Medicine evaluated a March 2025 rollout of ambient AI documentation (Abridge) in 20 Sutter Health hospitals in northern and central California. Just 59 physicians took part, half of them hospitalists, and 20 completed both the before and after surveys. Ambient AI generated about 3% of their inpatient notes (1,198 of 46,161), including 12% of history and physical notes and 6% of consult notes, and record data showed time in notes per day unchanged (41.9 minutes before, 41.6 after). The share who thought ambient AI could help them see one additional patient per shift fell from 55% to 45%, and 53% reported a "high degree of success" with AI documentation. The tool did not offer a copy-forward function for progress notes. Reported by Today's Hospitalist on Sept. 23, 2026.
- Epic said on March 10, 2026, that more than 85% of its customers use Epic AI and that clinicians at multiple organizations complete discharge summaries 20% to 30% faster with its Draft Hospital Course Notes. Its built-in ambient tool, Chart with Art, launched in February 2026 in outpatient specialties; Epic's product page lists note drafting and order queuing for inpatient and emergency clinicians. At its users group meeting in August 2026, Epic said Chart with Art was live in more than 70 specialties, that its Agent Factory for building AI agents would be widely available in 2027, and that it plans to integrate Curiosity, a model trained on its Cosmos database that simulates patient trajectories such as length of stay, into its record software in March 2027. Epic holds 43.7% of the U.S. acute-care record market and 56.9% of hospital beds, according to KLAS Research.
- In JAMA (April 1, 2026), a study of 8,581 clinicians at five academic medical centers found AI scribe adoption was associated with 16.0 fewer minutes of documentation and 13.4 fewer minutes of total record time per eight hours of scheduled patient time and 0.49 more visits a week; record time outside work hours did not change significantly. In a three-arm randomized trial of 238 outpatient physicians at UCLA (NEJM AI, Nov. 2025), one scribe (Nabla) cut time in notes by 9.5% and the other (Microsoft DAX) made no significant difference. A stepped-wedge pilot with nine hospitalists at four hospitals (Society of Hospital Medicine abstract, 2025) found no significant change in documentation time and no change in direct patient care time.
- In a blinded comparison of 100 hospital medicine encounters at UCSF (JAMA Internal Medicine, 2025), AI-written discharge narratives were of comparable overall quality but averaged 2.91 errors per summary against 1.82 for physicians' own; individual errors were no more harmful, and overall potential for harm was low for both, though higher for the AI narratives (0.84 vs 0.36 on a scale of 0 to 7). At NYU Langone (JAMA Network Open, Aug. 13, 2025), residents changed on average 31.5% of the text of Epic's embedded AI hospital-course drafts against 44.8% of physicians' drafts; attendings rated the AI drafts more complete but found more confabulations.
- In the 2023 Today's Hospitalist survey, nonacademic hospitalists reported an average of 16.7 patient encounters per shift and said 14.7 is reasonable. In a two-hospital study of about 20,000 hospitalizations (JAMA Internal Medicine, 2014), length of stay rose significantly at a census of about 17 when occupancy exceeded 85%, and each one-patient increase in census was associated with $205 more in cost, after adjustment for length of stay; mortality, readmissions and patient satisfaction showed no significant association. In Doximity's 2026 physician compensation report, 20% of physicians said they had already faced higher expectations for productivity because of AI.
- In a Feb. 23, 2026, letter to the Department of Health and Human Services, the Society of Hospital Medicine said time saved by AI should prioritize face-to-face and bedside time, not financial cuts, and that success should not be measured by "pure economic efficiency aims such as higher encounter quotas or reduced reimbursement." SHM's 2026 workforce report found 64% of hospitalists use AI and only a third of users reported adequate training. Most participants in the Peterson Health Technology Institute's AI Taskforce (March 2025), which includes health system leaders, said increasing patient throughput was not a priority and that reporting on it "may reverse the positive impact on burnout."
- In 22,349 encounters at Houston Methodist (JAMA Network Open, Sept. 3, 2026), a commercial AI tool's predicted discharge dates were about as accurate as case managers' at admission, but 24 hours before discharge 79.5% of case managers' estimates fell within a day against 37.9% of the AI's. Riverside Health in Virginia said its AI scribes had not led to more patient visits; it reported an 11% rise in physician relative value units, and a Riverside executive said a shift toward higher-level visit billing more than paid for the tool (Healthcare Brew, Aug. 5, 2026).Sources: JAMA Network Open, Healthcare Brew
- AMA policy (June 2024) recognizes "that organized medical staff should be an integral part at the outset of choosing, developing and implementing augmented intelligence and digital health tools in hospital care." The Joint Commission and the Coalition for Health AI's guidance (Sept. 2025) calls for formal AI governance, local validation, ongoing monitoring and training. In a study of 227,091 inpatient encounters at four health systems (JAMA Network Open, Feb. 27, 2026), version 2 of Epic's sepsis model had an area under the curve of 0.82 to 0.92 but high variation between institutions, low positive predictive value and a high alert burden.
- In the Society of Hospital Medicine's 2025 State of Hospital Medicine report, 64% of groups anticipated growth in positions in the coming year. Medical Economics (March 5, 2026) noted that few AI-driven layoffs in health care had been publicly reported; an employment lawyer it quoted called Revere Health's cut of almost 200 jobs "a largely AI-driven layoff." Revere, a Utah physician group, cut 177 positions in its central business office, which handles billing, claims and collections, after partnering with IKS Health on automated claims processing (HealthExec, Sept. 9, 2025). The Veterans Health Administration's tele-hospitalist cross-coverage, at 11 rural hospitals, allowed an average reduction of two full-time on-site clinicians per site. On Sept. 23, 2026, Allina Health and a Doctors Council SEIU unit of more than 150 physicians, including hospitalists, at its Mercy and Unity hospitals announced a tentative three-year first contract, reached after a four-day strike and pending a ratification vote.