A Modest Time Signal
The UCLA trial’s primary result is a difference, but not a sweep. Among physicians assigned access to ambient AI scribes, Nabla showed a measurable reduction in the trial’s time-in-note outcome compared with usual care, while DAX Copilot did not show a statistically distinguishable reduction. Reading that contrast correctly requires three things: the amount of actual use, the absolute change in seconds, and the uncertainty around each tool’s comparison with control.
Those details support a useful judgment. Ambient AI can improve a defined note-writing measure in a real outpatient workflow, but the result is modest, uneven between the two tools, and narrower than a claim about total documentation time. It also does not establish that one vendor is generally better than the other.
The tools were used on only part of the available work
DAX was used in 8,271 of 24,696 patient visits during the study, or 33.5%. Nabla was used in 6,981 of 23,653 visits, or 29.5%. The usual-care group had 24,020 encounters during the same period. These denominators describe visit-level use, not the percentage of physicians who adopted a tool. A physician could use the application on some encounters and decline it on others.
The trial also reports that approximately 15% of physicians in the treatment groups never used their assigned scribe. That fact is not a side note about motivation; it is part of what the assignment-based result means. Access to a tool does not guarantee that the tool fits every visit, schedule, specialty, or documentation style. A clinician may decide that starting the system takes too long for a short encounter, that the draft requires too much correction, or that the workflow is not worth the effort on a particular day. The trial’s average includes those decisions.
This is why it would be misleading to take the best experience of a frequent user and present it as the effect of assigning the tool to an entire practice. A highly consistent user may receive a different benefit from a physician who rarely starts the application. The randomized estimate answers a system-level question: what changed when physicians were given the option under the trial’s training, support, integration, and time limits? It includes both the potential value of the tool and the friction of using it.
The usage percentages also should not be read as a direct measure of effectiveness. A tool might be used less often because it is reserved for encounters where it is most helpful, or because the startup process is inconvenient. The study did not establish which explanation accounts for every unused visit. The numbers show that the workflow was optional and uneven; they do not by themselves show why each physician made each choice.
Three different denominators are in play. The physician denominator tells us how many people received an assignment. The visit denominator tells us how often the software was actually used among encounters counted for the intervention. The outcome denominator is a provider-level average of note-writing time. Those denominators cannot be swapped. Saying that Nabla was used in 29.5% of visits does not mean that 29.5% of physicians adopted it, and saying that the average time fell does not mean that every participating physician or every note became faster.
That separation is particularly important when a tool is optional. Imagine two physicians assigned the same application. One uses it for most long visits and finds that the draft gives them a workable starting point. The other uses it rarely because short visits do not justify the setup or because their specialty’s notes need extensive restructuring. An analysis restricted to the first physician would answer a user-experience question. An intention-to-treat analysis of both physicians answers the broader adoption question: what is the effect of offering the workflow in practice, including the fact that fit will vary?
The trial does not tell us which use pattern is optimal. It does show that the headline estimate should be understood as an average across actual behavior, not as the guaranteed output of a fully adopted product. That makes the result more relevant to a health system deciding whether to deploy a tool, while also making it less appropriate for a sales claim based on the most favorable individual use case.
The seconds are small; the comparator decides their meaning
In the model comparing the second intervention month with baseline, time-in-note declined by an estimated 18 seconds in the usual-care group, 23 seconds in the DAX group, and 41 seconds in the Nabla group. The within-group figures provide scale. The trial was not observing a wholesale disappearance of note writing. It was observing changes measured in seconds per note against a baseline measured in minutes.
The important comparison is not simply “41 is lower than 23.” Each intervention group must be compared with the contemporaneous control, because documentation time can change for reasons unrelated to the scribe. The usual-care group also moved downward during the study period. When Nabla’s decline was compared with that control trend, the estimated relative difference was a 9.5% reduction, with a 95% confidence interval from 17.2% below control to 1.8% below control and a P value of 0.02. In the trial’s primary analysis, that interval stayed below zero, supporting a modest Nabla-versus-control difference under the study conditions.
DAX showed a different pattern. Its estimated difference from control was a 1.7% reduction, but the 95% confidence interval ran from a 9.4% reduction to a 5.9% increase, and the P value was 0.66. Because the interval includes no difference, the study cannot distinguish the DAX estimate from ordinary variation around the control trend. The result is not evidence that DAX increased documentation time. It is evidence that this trial did not establish a DAX reduction relative to usual care on the primary measure.
“Statistically distinguishable” is not the same as “large enough to matter” in every setting. Nabla’s estimate is useful evidence that the groups differed on the study’s chosen measure, but a health system would still need to translate that contrast into its own volume, note mix, staffing, and review costs. Conversely, a result that is not statistically distinguishable from zero is not proof of no benefit. It may reflect a genuinely small effect, wide uncertainty, limited exposure, uneven use, or a metric that misses the relevant work. The confidence interval is valuable because it shows which of those magnitudes remain compatible with the data, even though it cannot decide the cause.
The absolute numbers and relative comparison answer different questions. The estimated seconds per note make the scale tangible: the trial observed tens of seconds, not the complete removal of the note-writing task. The percentage contrast adjusts that change against what happened in usual care, which is why it is the more appropriate figure for the primary comparison. Neither figure should be presented alone. Seconds without a comparator can exaggerate an ordinary time trend; a percentage without the underlying minutes can make a modest change sound transformational.
Confidence intervals matter here because a point estimate alone can create false precision. The DAX estimate is close to zero, but its interval allows for a meaningful reduction as well as a small increase. The Nabla interval is also broad enough that the true effect could be smaller or larger than 9.5%, but every value in the reported interval is a reduction relative to control. Neither interval describes what will happen in every clinic or for every clinician. It describes uncertainty around the study’s estimated contrast.
The control trend is not merely a statistical nuisance. It represents the documentation environment that the intervention had to beat. If all groups become faster during a period, perhaps because of changing workload or growing familiarity with the EHR, then an intervention’s raw decline overstates its incremental contribution. The trial’s question is whether the scribe group moved beyond that background change. In Nabla’s case the reported contrast supports that conclusion for the primary metric; in DAX’s case the data do not separate the tool’s estimate from the background movement.
A result can also be real without being sufficient for a purchase decision. A small average reduction might be valuable if it arrives with few corrections and improves attention during visits. The same reduction might be unattractive if clinicians must perform extensive checking in another application or if the tool works only for a narrow slice of encounters. The randomized estimate does not contain those local costs. It should inform a deployment decision, not replace the measurement a deployment would need.
The two results also should not be turned into a universal vendor ranking. Nabla’s estimate was statistically distinguishable from the control trend; DAX’s was not. That is a reason to report the tools separately, not a license to conclude that Nabla is categorically superior. The tools were used in the same trial but could differ in integration, adoption, clinician fit, and unmeasured review behavior. A non-significant DAX result does not prove that DAX never helps, and a significant Nabla result does not prove that Nabla will help in a different setting.
A shorter note-writing measure did not shorten every kind of EHR time
The trial found no reported differences between groups in time spent in the EHR on unscheduled days or in the EHR outside scheduled hours. Those measures are relevant because after-hours work is often what makes documentation feel burdensome. Their lack of a detected difference means the study did not show that either scribe reduced those categories during the measured period. It does not prove that no individual physician experienced an after-hours benefit, nor does it show that every form of work remained unchanged.
The contrast illustrates the importance of outcome scope. A physician can spend less time writing a note during the recorded EHR activity while still spending similar time elsewhere. They may review a draft in the vendor platform, finish an encounter through a different interface, or use the saved time to handle another documentation task. The trial did not provide enough measurement to identify every path. Therefore, the primary result should be described as a change in time-in-note, not as a demonstrated reduction in all EHR or after-hours work.
There is also a practical reason not to dismiss the measured change because it is modest. A reduction of seconds per note can matter when multiplied across many encounters, particularly if it also makes the visit less cognitively fragmented. But that multiplication is a decision calculation, not a result reported by the trial. It would require knowing how many eligible notes a service produces, how often the tool is used, how much review occurs outside the measured clock, and whether any error creates rework or downstream harm. The study supplies one input to that calculation, not its final answer.
The absence of an off-hours difference should be read with the same discipline. It does not show that the scribes failed to help anyone after the clinic day, because a group average can conceal gains for some physicians and losses for others. Nor does it show that work was merely moved outside the EHR. It shows that the study did not detect a between-group difference in the two additional EHR-time measures it collected. The result narrows the claim: the observed benefit was not a demonstrated reduction in every category of recorded EHR activity.
The distinction between detected and not detected is important in both directions. A detected Nabla contrast deserves attention, but it does not identify the mechanism or guarantee durability. A not-detected DAX contrast leaves several possibilities open, including a small benefit that the study could not separate from variation. Evidence is strongest when the wording follows the design: “this comparison showed” is safer than “this technology always does.”
What the result does—and does not—mean
The primary outcome is best read as a three-part statement. First, the randomized comparison provides evidence that Nabla assignment produced a modest improvement in the defined time-in-note measure relative to usual care during the second intervention month. Second, the same comparison did not establish a DAX improvement on that measure. Third, the time-in-note measure omits editing time inside the vendor platforms, so it may overstate the reduction in the complete documentation workflow.
That third point does not cancel the first two. An incomplete measure can still change reliably, just as a partial thermometer can still show that one room is warmer than another. It becomes misleading only when the partial reading is described as the whole system. Here, the EHR clock can tell us about writing activity captured by Epic Signal. It cannot tell us how much time a physician spent inspecting the draft before transferring it, whether a clinician accepted or rewrote the suggested structure, or whether the note required extra checking after it was closed.
The usage pattern makes the same point from another angle. The trial did not test an imaginary workflow in which every physician uses the scribe for every visit and accepts the first draft. It tested access in practice, where some physicians used the tool and some did not. That is valuable evidence for adoption decisions because real systems also face uneven use. It is less useful as a promise of what a perfectly optimized individual user will experience.
For a decision maker, the next evidence question is not simply which point estimate is lower. It is whether the result survives contact with the local workflow. That means asking whether clinicians can start the tool without disrupting the visit, whether the application works for the relevant note mix, whether review time is visible in the same measurement system, and whether any savings persist once early novelty has faded. Those questions are not answered by the trial’s primary outcome, but the outcome tells us why they are necessary.
It is also worth resisting a false choice between believing the result and dismissing it. The Nabla estimate can be a genuine, useful improvement and still be too incomplete to support a claim of net burden reduction. The DAX estimate can fail to reach statistical distinction and still leave room for benefit in another workflow. Careful interpretation preserves both possibilities until better measurements show how drafting, review, and validation interact.
The most defensible conclusion at this stage is therefore conditional: Nabla showed a modest time-in-note advantage over usual care in this trial; DAX did not show a statistically distinguishable advantage; neither result demonstrates that total documentation burden or after-hours work was reduced. The next question is why physicians can report a better work experience even when the clock moves only modestly—and what kind of work is hidden inside the draft that the clock does not see.