Priya Nandakumar leads engineering at Scribemarrow. This piece reflects what we've heard consistently in conversations with physicians over the past year - before trials, during onboarding, and after they've been using ambient documentation tools for several months. We ask a lot of questions. The answers below are patterns, not single data points.
They want less editing, not just faster generation
The instinct in product development is to optimize for generation speed - how quickly can we produce a note draft after the visit ends? Speed matters, but it is not the primary driver of physician satisfaction with ambient documentation tools. The primary driver is editing time: how much does the physician need to change before the note accurately reflects what happened?
A note that is generated in 10 seconds and requires 4 minutes of editing is less useful than a note that takes 30 seconds to generate and requires 60 seconds of editing. Physicians understand this intuitively. When we ask what they want from a documentation tool, "accuracy" and "minimal editing" consistently rank higher than "speed."
This has implications for how practices should evaluate tools. Time-to-draft is easy to measure. Editing time is harder to measure but much more predictive of long-term satisfaction.
They want the note to sound like them
Different physicians have distinct documentation voices. Some write dense, highly structured notes with precise clinical language. Others write more conversational notes that read like a narrative. Some use specific abbreviations or format the assessment section in a particular way.
When ambient tools produce notes that don't match a physician's voice, the editing task becomes onerous not because the clinical content is wrong but because every sentence needs to be rewritten for style. This is perhaps the most underappreciated source of physician frustration with early ambient tools.
The better tools have addressed this in two ways: template-based output that preserves the physician's preferred structure, and adaptation over time as the model learns from how the physician edits its output. Neither fully replaces a physician who has been using a tool for six months and whose editing time has dropped to close to nothing, compared to a physician in week two who is still correcting stylistic mismatch on every note.
They want the assessment and plan to actually be useful
The subjective section of a SOAP note is where ambient tools perform most reliably. The model has clear audio of the patient describing their symptoms; this is relatively structured content that transcribes and organizes well.
The assessment and plan is where physicians report the most dissatisfaction. A good assessment and plan section captures the physician's clinical reasoning - the working diagnosis, the differential considerations, the rationale for ordered tests, the plan for medication changes or referrals, and follow-up instructions. This requires the model to understand medical decision-making, not just transcribe speech.
When we ask physicians what they most want improved in documentation tools, the assessment and plan comes up in nearly every conversation. Specifically: they want the plan section to capture what was actually ordered, what was changed in the medication list, and why - not just a generic summary of what the physician said during the visit. This is a harder problem than subjective section generation, and the tools that solve it most reliably have a significant advantage in long-term adoption.
They want it to handle the visit types that are actually hard
Simple visits are not the problem. A straightforward hypertension follow-up - blood pressure checked, medication at therapeutic dose, no new complaints, refill written - takes very little documentation time to begin with. Ambient AI on a visit like this produces a perfectly adequate note, but the value delivered is modest because the baseline burden was modest.
The visits physicians describe as hard are complex multi-problem visits: the patient presenting for a scheduled chronic disease follow-up who also brings up new symptoms, a family concern, and a medication side effect they've been experiencing for three months. These visits are hard to document because they require the physician to track multiple concurrent threads and then structure them coherently in the note.
Ambient tools that handle these visits well - that correctly organize multiple chief complaints, track which plans apply to which problems, and structure the note so a reviewing physician can follow the clinical logic - are tools that earn sustained use. Ambient tools that handle simple visits well but fall apart on complex visits end up getting abandoned or used selectively.
They want patient consent to be handled cleanly
Patient consent to be recorded during a visit is a non-trivial workflow step, and physicians care about how their documentation tool handles it. The worst implementations put the consent conversation entirely on the physician: the physician has to introduce recording, explain it, field questions, and handle opt-outs, all within a 15-minute visit window that is already fully allocated.
Better implementations treat consent as a front-desk step. The patient receives a brief explanation of how visit documentation works when they check in. By the time the physician enters the room, the consent question has been handled. This preserves visit time and removes an awkward conversational task from the physician-patient interaction.
Physicians also want clarity on what happens when a patient declines. The workflow for non-consenting visits should be clearly defined, low-friction, and not require the physician to remember a special mode or procedure mid-visit.
They want it to not add new tasks
Every ambient documentation tool requires some physician interaction to function: starting a recording, reviewing a note, signing off. Physicians are willing to accept these steps if they are minimal and the net result is time saved. They are not willing to accept tools that add multiple new workflow steps in exchange for modest time savings.
When physicians tell us a documentation tool "didn't work," a common underlying reason is that the total physician effort - including the new steps the tool required - was not much less than the effort of writing the note without the tool. The value proposition collapsed because the tool itself was too demanding.
The best ambient documentation tools feel like they disappear after setup. The physician sees patients; the notes appear; the physician reviews and signs. The fewer deliberate actions the physician must take to get from "patient visit" to "signed note in the EHR," the higher the sustained adoption rate.