Doctors given access to AI scribes spent less time documenting care, yet their work in electronic health records outside working hours did not fall significantly. That combination appeared in a multisite study published in JAMA in 2026. It leaves a question that will matter well beyond medical notes. Who receives the benefit when software takes over part of a doctor's work?

I expect AI to change medicine more deeply than a faster consultation would suggest. Once a task can be performed without drawing on a doctor's attention every time, it becomes possible to move that task, offer it more often or make it available to patients who previously struggled to obtain it. The visible change could be an appointment that becomes unnecessary, or an examination that finally happens.

Following the time makes this prospect concrete. A saved minute can relieve a clinician, increase the number of patients a clinic can see, or make room for a more difficult conversation. Those outcomes require different choices about what happens next.

Saving minutes inside the appointment

An ambient AI scribe turns a clinical conversation into draft documentation. In the JAMA study, access to these tools was associated with 16 fewer documentation minutes per eight scheduled patient hours, alongside 13.4 fewer minutes in the electronic record overall. Weekly visits increased by 0.49. The study covered 8,581 clinicians at five US academic institutions, including 1,809 who received access. These were observational associations, with voluntary uptake at most sites. The study's methods and results describe the comparison.

The time measures overlap. Adding them would count part of the same saving twice. Neither figure measures how much longer a doctor listened to a patient. Establishing where the freed time went would require observing what clinicians and their organizations did with it.

Consider the scheduling problem. If documentation becomes faster while appointments retain their existing length, the spare time may remain inside the consultation. It could absorb a delay or allow a question that would otherwise feel rushed. If the clinic shortens appointment slots, it can attempt to convert that capacity into additional visits. If the next constraint is room availability, nursing support or an unfinished referral, faster documentation alone may leave the schedule largely unchanged.

These possible uses of the saving explain why the same tool could produce different experiences in different clinics. The person doing less typing and the person deciding how appointments are booked may have different priorities. A purchase decision can therefore become a decision about the pace of clinical work.

There is also a difference between saving time in scattered moments and recovering a usable stretch of the day. A little space between appointments can make a session less frantic without creating another appointment slot. The reverse is possible too. A schedule built to absorb every expected saving leaves little room for the visits that take longer than expected. The calendar is part of the technology's effect.

For a patient waiting for an appointment, additional visits may be a valuable result. For a clinician already stretched, the same increase may feel like an acceleration of work. Calling either outcome efficiency leaves out whose time is being improved. The useful question is how the change travels through the practice, including the work that still has to be done after the note exists.

An examination can move to the patient

Autonomous retinal screening shows a more substantial change in who performs a medical task. The FDA's review of IDx-DR describes software that analyzes retinal photographs for a specified level of diabetic eye disease. Its scope is narrow. It does not examine every possible eye condition, and a concerning result requires further care. The FDA decision summary retains roles for clinicians in reviewing results and arranging referrals.

That arrangement separates obtaining a screening result from having a specialist interpret each image. The machine can perform a bounded diagnostic operation while people remain responsible for the surrounding care. Its significance lies in where the examination can now fit.

In the ACCESS randomized trial, young people with diabetes were offered an autonomous eye examination at their diabetes visit or referred for a separate eye-care appointment. Within six months, screening completion was 100% in the AI group and 22% in the referral group. The trial randomized 164 participants. Published in Nature Communications in 2024, it measured completion of screening, not prevention of blindness. It also tested a complete service arrangement, including immediate access, rather than isolating the algorithm's contribution.

The authors reported that the device lacked FDA clearance for the youth age group they studied, placing this use within a research setting. Among participants with an abnormal AI result, 16 of 25 completed specialist follow-up. The ACCESS results and limitations make both boundaries explicit.

The benefit here reaches beyond a doctor's desk. When a suitable examination can happen during a visit already underway, the patient may avoid arranging transport, finding another appointment and returning on another day. Some of medicine's time cost sits in the patient's calendar. Moving the test can change whether that time has to be found at all.

This is why I take the possibility of task replacement seriously. A profession contains many activities, and each activity has its own requirements. Transferring a defined operation to software can change a service substantially while leaving the name of the profession intact. Waiting for a machine to assume an entire doctor's role would miss that change.

The specialist appointment then has a different place in the sequence. If screening reaches more patients, it may generate more referrals requiring attention. Capacity released at one point can expose a shortage at the next. For someone whose screening result needs follow-up, the relevant saving is incomplete until that next appointment is available. The patient's journey continues after the algorithm finishes.

Clinical reasoning is entering the same discussion

The possibility extends into reasoning about a case. A randomized study in Pakistan, published in Nature Health in 2026, compared physicians using GPT-4o and conventional resources with physicians using conventional resources alone. All had completed AI training. Among 58 physicians completing the study, diagnostic reasoning scores averaged 71.4% with AI access and 42.6% without it. The adjusted difference was 27.5 percentage points; time per case was similar.

An exploratory comparison also found stronger performance from the model alone than from the assisted physicians. The assessments used written clinical scenarios. They did not test autonomous treatment of real patients, and the score was not a percentage of patients successfully treated. The study design supports a finding about reasoning performance under those conditions.

An improvement in the quality of a decision made within the same time allowance deserves attention alongside a faster decision. The economic question changes with the outcome being sought. A clinic trying to serve more patients would need evidence about capacity. A service trying to improve diagnostic reasoning would care about what the same clinical time can accomplish. Both could be worthwhile, but they call for different measurements.

My expectation is that more medical work will be reorganized around this distinction. Some tasks may become autonomous within a defined scope. Others may remain with a clinician whose preparation and reasoning are supported by software. The amount of human work left in either arrangement will depend on what information must be gathered, what needs checking and what action follows the answer.

For example, a model that requires a carefully assembled case history still depends on someone producing that history. If a physician must reconstruct the record before asking the question, part of the promised saving has already been spent. Conversely, a system that can reliably assemble the relevant material could change the value of the reasoning tool attached to it. That is a dependency worth watching as these products develop.

This makes broad replacement forecasts difficult to express as a single percentage. An answer generated, a decision accepted and a course of care completed are different units of work. I would expect substantial change to become visible in those units before it becomes clear in the number of people carrying a particular job title.

The next appointment belongs to someone

Scaling these arrangements requires resources outside the model. The OECD's March 2026 assessment identifies fragmented data, regulatory uncertainty, governance and workforce capacity as barriers. Its account helps explain why a capable tool can coexist with a service that changes slowly. Connecting the tool to usable records, trained staff and a functioning care pathway is part of providing it.

An institution also has to decide what a successful change looks like. Suppose released capacity goes into appointments for people already waiting. The gain may appear as better access while the clinicians' working hours remain similar. Suppose it supports longer consultations for complex cases. Visit counts might stay flat. If it reduces overtime, the gain may be most visible to staff. Each choice would leave a different trace in the service's results.

The commercial arrangement can influence that choice. Where additional completed visits bring additional income, expanding the schedule has an immediate financial attraction. Where a service is funded to care for a population, avoiding unnecessary work may have greater value. Neither arrangement determines every decision, but each gives the institution a different reason to use the same technical capacity.

Patients also have an interest in what fills the available time. More examinations can be useful when they reach people who need them. Additional activity can consume appointments too, including follow-up generated by uncertain findings. A faster first step is therefore only part of the capacity calculation. The work it creates elsewhere belongs in the same account.

I expect the scale of AI's effect to become clearer as these arrangements spread. Medicine can feel familiar at the point of contact while the work behind an appointment changes considerably. A patient may still see a doctor, yet arrive with more of the preliminary work completed, or reach that doctor because a screening task no longer occupies a separate specialist visit.

Who gets the time will depend on the decisions made around that appointment. Software can release capacity; the service decides whether to spend it on another patient, a longer conversation or a clinician leaving on time. For the patient still waiting, the consequential change is when the newly possible appointment becomes theirs.