What the Trial Actually Tested
The UCLA randomized trial is the most direct evidence in this series about whether ambient AI scribes changed outpatient documentation. Its useful question is narrower than “Do AI scribes work?”: what happened when physicians were assigned access to Microsoft Dragon Ambient eXperience, or DAX Copilot, Nabla, or usual care under a defined clinical workflow? Understanding that design lets the reader interpret the later time estimates without confusing a physician’s assignment with the number of visits on which a tool was used.
The answer also depends on the clock. The trial measured a specific electronic-health-record activity called time-in-note, after allowing a period for physicians to become familiar with the tools. It did not create a stopwatch for every act of review, correction, or editing in a vendor interface. The design is therefore both a strength and a boundary: it gives a real comparison in routine practice, but it cannot make an incomplete metric stand for the whole documentation task.
The comparison was three groups of physicians, not a contest among every visit
The study enrolled 238 outpatient physicians from 14 specialties. Seventy-nine were assigned to DAX Copilot, 79 to Nabla, and 80 to a contemporaneous usual-care control group. The unit randomized was the physician, not the individual visit. That distinction matters because a physician assigned to an AI-scribe group could choose when to use the tool, and a physician’s overall documentation activity included notes produced on visits where the tool was not used.
The assignment used a 1:1:1 covariate-constrained design. In plain language, the researchers did not simply distribute names at random and hope the groups looked alike. They balanced the assignment around three factors that could influence the result: baseline time-in-note, a baseline burnout score, and the number of clinic days per week. Balancing does not eliminate every difference between groups, but it reduces the chance that one group started with a systematically heavier documentation load or a different level of reported burnout.
Those covariates also show what the researchers thought could distort a comparison. A physician who already spends much longer in notes has more room for a time reduction than one whose baseline is short. A physician beginning with a different level of burnout may report a different response to the same workflow. Clinic days shape exposure to outpatient documentation. Balancing these factors improves the comparison, but it is not a guarantee that the groups are identical or that every relevant influence has been measured. Randomization supports a fairer estimate; it does not turn a short trial into a controlled laboratory.
The choice to randomize physicians rather than visits has another consequence. A physician brings a stable documentation style and a repeated set of preferences to many encounters. If the study assigned one visit to an AI tool and the next to usual care, the same person’s habits could contaminate the comparison, and the tool could be selected for the visits most likely to benefit. Physician-level assignment avoids that particular problem by keeping the workflow available across the physician’s study period. It also means that the result represents an average change in a clinician’s practice, not a guaranteed improvement for every individual note.
The intervention was embedded in outpatient work rather than tested as a detached demonstration. DAX Copilot and Nabla were connected to the electronic health record so that their generated text could populate the native note environment. Physicians received a standardized hour of virtual training before the study, a later presentation about advanced features and customization, and support from internal and vendor channels. Participants who normally used human scribes had to give up that assistance while they were in an intervention group. The comparison was consequently between defined documentation arrangements, not between AI and an idealized blank slate.
The study ran from November 4, 2024, through January 3, 2025. Visits had to be in English because the systems’ translation capabilities had not been internally validated. That restriction is easy to overlook, but it defines who was studied. The results speak most directly to English-language outpatient documentation in one academic health system during that period. They do not establish how the same tools would perform for multilingual encounters, different institutions, or workflows with different staffing and note conventions.
The trial’s pragmatism is important. Physicians had to use the tools in the conditions of actual practice, with the interruptions, preferences, and uneven adoption that come with an optional workflow. At the same time, pragmatism makes the estimate more specific, not less. It tells us what happened under this training, support, integration, period, and participant mix. It does not tell us what would happen after a long optimization program or in a clinic organized differently.
The setup also avoids a misleading comparison with a nonexistent “no-work” condition. Physicians in usual care still had to document encounters; physicians in the intervention groups still had to review their records. The experimental difference was the availability of a particular way to begin the note. Any benefit must therefore be understood as a change in division of labor. The system supplies draft language from the conversation, while the clinician retains responsibility for deciding whether that language belongs in the permanent record.
The clock started before the result and stopped before the whole workload
The prespecified primary outcome was documentation time, expressed through Epic Signal’s time-in-note measure. At the provider level, the metric represents the total number of minutes spent writing notes in a week divided by the total number of notes. Those provider results were averaged by month. It is a useful operational measure because it records a recurring activity across a clinician’s ordinary notes rather than asking them to remember how long documentation took.
The denominator changes the meaning of the result. “Minutes per note” is not the same as total minutes spent documenting in a week. A physician may write fewer notes, more notes, or notes of different lengths while the average time per note moves. The study’s provider-level measure is still relevant to note-writing efficiency, but it should not be silently expanded into a claim about total workload, patient volume, or time available after work. A good measurement discussion keeps the unit visible: the outcome is an average writing time associated with notes, not an all-purpose measure of the physician’s day.
The metric is also a behavioral record rather than a direct observation of quality. It records how long the EHR identifies the physician as writing, but it does not determine whether a shorter note was more concise, whether a missing detail had to be repaired elsewhere, or whether a clinician moved checking to another application. Those questions require other measures. The trial included surveys and safety questions for that reason, but adding several measures does not make any one of them a complete account of the workflow.
The researchers compared each group’s later study-period measure with a six-month pre-study baseline. They allowed the first intervention month to serve as a proficiency period, then used the second month for the main comparison. This choice recognizes that a new tool can impose a temporary learning cost. A physician must learn how to start an encounter, position a device, correct a draft, and decide when the system is worth using. Measuring too early could mistake unfamiliarity for the tool’s settled performance.
But the lead-in does not solve every timing problem. A short study can show early behavior under a real contract and a common time window; it cannot show whether clinicians become more efficient after months of customization or whether vigilance becomes more tiring with prolonged use. It also cannot distinguish every change caused by the tool from changes in the clinic, workload, season, or the physician’s growing familiarity with the measurement itself.
The primary analysis used the intention-to-treat principle. That means physicians were analyzed according to the group to which they were assigned, not reclassified every time they used or declined the tool. This is the right question when a health system is considering an offer of access: what is the effect of assigning the workflow to clinicians who may adopt it unevenly? It is not the same as asking how much faster a perfectly compliant user would be on a single visit. An intention-to-treat estimate includes the friction of adoption, nonuse, and fit with individual practice.
This distinction is especially important for optional clinical software. People who choose to use a system often differ from people who do not: they may have heavier documentation loads, more interest in experimentation, more compatible schedules, or more tolerance for correcting drafts. Looking only at users can produce an attractive estimate that describes self-selection as much as the tool. An assignment-based analysis avoids that particular bias, although it can make the average effect look smaller when substantial nonuse is part of real adoption. Both views can be useful, but they answer different questions and should not be presented as interchangeable.
The trial also compared changes against a contemporaneous usual-care group rather than assuming that the pre-study baseline was a fixed backdrop. Documentation time can change because of seasonal workload, staffing, local policy, or simple familiarity with the EHR. If the control group changes too, the intervention estimate asks whether the scribe group changed more than the surrounding practice. That comparator is what turns a before-and-after observation into a more informative test, even though it cannot remove every time-related influence.
The metric’s most important limitation is its boundary. Physicians could edit AI-generated text inside the vendor platform as well as in Epic, but Epic Signal did not capture time spent editing in those vendor environments. A decrease in EHR writing time could therefore represent a genuine reduction in composition work, while still omitting some review time. The trial authors explicitly note that the reported time saving may be an overestimate. This does not invalidate the result; it changes the claim the result can support.
A useful way to state the distinction is:
- Measured: a provider-level change in EHR note-writing time, compared with a prior baseline and a contemporaneous control.
- Partly represented: perceived task load, work environment, stress, work exhaustion, and usability, captured through surveys.
- Not fully captured: every minute spent reading and editing a draft in a vendor interface, the mental effort of deciding what to keep, and the total time required to verify a clinically important omission or attribution.
The distinction is not a technical footnote. If the tool produces a concise, accurate draft, a clinician may spend less time writing and less effort maintaining a running mental list during the encounter. If it produces a verbose or subtly wrong draft, the physician may spend more time inspecting it than the EHR clock records. The same measured reduction can therefore correspond to different real experiences for different clinicians.
Randomization does not mean everyone used the tool in the same way
The trial’s three groups establish a comparison, but they do not impose identical behavior within the groups. A physician assigned to DAX Copilot or Nabla still had to initiate the workflow, obtain the required consent for audio recording, and decide whether the draft was useful for that encounter. The tool was an option inside practice, not a machine that automatically documented every visit. The control group, meanwhile, continued with usual care during the same period, so its physicians were exposed to the same broad clinical environment without the assigned AI-scribe access.
That distinction helps explain why physician-level assignment is the appropriate unit for the primary analysis. Notes are not independent laboratory trials. One physician’s style, schedule, specialty mix, comfort with dictation, and tolerance for editing can affect many visits. Randomizing the physician keeps those repeated decisions within the assigned group and asks whether the overall workflow changed their measured documentation pattern.
It also limits what can be inferred from a favorable result. If an estimate reflects access to a tool under training and support, it may be smaller than the effect among highly motivated users who employ it on nearly every eligible visit. That is not a flaw to be corrected by silently selecting only enthusiastic users. It is information about adoption: a technology that helps only when use is consistent may deliver less benefit at the system level than its best individual demonstrations suggest.
The trial used surveys to capture outcomes that a writing-time metric cannot see. The Mini-Z 2.0 assessed aspects of burnout, work environment, work pace, and electronic-record stress. A four-item physician task-load instrument addressed the cognitive load associated with documentation, and the Professional Fulfillment Index–Work Exhaustion scale addressed exhaustion. Separate questions asked about accuracy, safety, usability, and whether the tools made patient engagement easier. These measures broaden the evidence, but they remain measures of reported experience, not a substitute for observing every action in the workflow.
Each measure has a different relationship to the burden claim. Time-in-note is closest to a recorded behavior, but it is narrow. A task-load score asks how demanding the work feels, but it is subjective and may be influenced by expectations. A burnout or exhaustion score describes a broader state with many possible causes, so a short intervention cannot easily attribute a change to documentation alone. Usability ratings tell us whether physicians can operate the tool and see value in it; they do not certify that every generated sentence is accurate. Keeping these meanings separate is more informative than combining them into one reassuring “impact” number.
The same logic applies to the trial’s safety questions. A physician may notice an occasional error without encountering one during the particular notes represented in the time metric. A time reduction may coexist with a need for careful checking. The absence of a recorded event in a short study would not prove that no risk exists, just as the presence of one event would not establish a population-wide error rate. The design can identify signals that deserve attention, but each signal must retain its denominator, response method, and uncertainty.
The design thus gives us a disciplined reading strategy. First ask what the intervention was: assignment to access to one of two tools, in a particular outpatient environment. Then ask what changed: a defined EHR time measure and several survey outcomes. Finally ask what was not randomized or recorded: the exact amount of vendor-platform review, the long-term learning curve, and the full consequences of errors. The result that follows is meaningful precisely because the comparison is clear, and limited precisely because the comparison is not a complete digital replica of documentation.
The next step is to read the observed trajectories for DAX Copilot, Nabla, and usual care. With the unit of assignment, timing, and metric boundary in view, a difference in time-in-note can be interpreted as a study result rather than mistaken for a universal statement about ambient AI.