The AI drafts the note. The physician reads, corrects, and signs every word.

Start with the problem. Why does clinical documentation warrant this kind of attention?

Documentation stopped being a byproduct of care and became a second job. Depending on the specialty, our physicians were spending one to two hours a day in the record for every clinic day, and a meaningful portion of that happened at night after their families had gone to bed. That pattern correlates closely with burnout scores.

Burnout is not only a wellbeing concern. It is a business problem. A physician who is exhausted sees fewer patients, is more likely to cut back to part time, and is far more likely to leave. Replacing a departing physician is slow and expensive, and patients lose continuity in the meantime. There is a quality dimension as well. When a clinician is typing and clicking through templates, they are not fully listening, and patients notice the back of a laptop. So the documentation burden was quietly eroding three things at once: our people, our capacity, and the relationship at the center of what we do.

Dictation and speech recognition have existed for years. What specifically changed to make this viable now?

The older tools transcribed words. You dictated after the visit, and the software turned speech into text that you still had to arrange into a note from memory. It saved some typing, but it left the structuring work to you.

What changed is the combination of reliable speech recognition with language models that can follow a live, multi-speaker conversation and organize it into clinical structure. The system distinguishes the clinician from the patient, separates the history from the exam from the plan, and produces a draft in the format we actually use rather than a wall of text. Reported error rates for modern ambient tools sit meaningfully below what we saw with older dictation, though I am careful not to overstate that. The other shift is integration. These tools can now write a draft note directly into the chart through standard healthcare data interfaces, so the output lands where the physician already works instead of in a separate window.

Walk me through how it actually works in a visit, step by step.

It begins with consent. Before recording starts, the patient is told an AI tool will help draft the note, and they can decline or ask us to pause at any point. In practice very few patients opt out.

Once the visit begins, the physician sets a phone or workstation to listen ambiently and then simply has the conversation. There is no dictating to the room and no reciting findings for a microphone. The system captures the audio, produces a transcript, and identifies who said what. A language model then drafts a structured note, sorting reported symptoms into the subjective section, examination findings into the objective section, and the rest into assessment and plan. That draft is written into the record in an unsigned, draft state. Nothing files automatically. The physician reads it against their own memory of the visit, corrects anything wrong or missing, adds the clinical reasoning the conversation did not capture, and signs. The signature, not the AI, is what commits the note to the legal record.

This involves recording confidential patient conversations. How do you handle privacy, security, and HIPAA obligations?

We treat the scribe as a system that touches protected health information, which means it belongs inside the same governance we apply to the record itself. Before deploying anything, we ran a security risk analysis covering how audio and text move through the tool, where they are stored, and for how long. The vendor signed a business associate agreement, which is not optional under HIPAA when a third party processes patient information on our behalf.

We confirmed encryption in transit and at rest, role-based access so only the treating clinician and authorized staff can see a given note, and audit logging so we can trace who accessed what. Data retention was a specific negotiation. We did not want raw audio held indefinitely, so we set a short retention window for recordings once a note is finalized. Consent is documented per visit with a timestamp. We also placed the tool under our existing clinical AI oversight, with a defined path for reporting errors, because these models are updated over time and that needs monitoring rather than a one-time approval.

How do you manage accuracy, and who is ultimately accountable for what ends up in the chart?

The accountable party has not changed. The physician who signs the note is responsible for every word in it, exactly as they were when they typed it themselves. The AI produces a draft, and a draft is a proposal, not a record. That distinction is the whole safeguard.

We are candid with our clinicians that these systems can hallucinate. They can insert a plausible detail that was never discussed, or drop one that was. One blinded study found fabricated content in a higher share of ambient drafts than of physician-written notes, so this is a real failure mode, not a hypothetical. Our answer is a mandatory review gate. No note reaches the chart without a clinician reading it against their recollection and correcting it. We coach physicians specifically to check medications, doses, laterality, and any number the model reports, because those are where an error does the most harm. The efficiency comes from editing a good draft, not from trusting it.

How did adoption actually go? Physicians are not always eager for new software.

The first lesson was that mandates produce compliance, not use. We made it opt-in and started with a pilot of volunteers across a few specialties rather than a top-down rollout. Our early mistake was treating it as an IT deployment. It is really a workflow change, and it needed to be led by clinicians, not handed to them.

Once we named physician champions in each department and gave them a little protected time to coach peers, uptake improved noticeably. Peer guidance solved problems that formal training missed, small things like where to place the phone or how to phrase a plan out loud so it lands cleanly in the note. Templates mattered too. The generic note structure did not fit every specialty, so we built specialty-specific formats with the physicians who would use them. We were also upfront about expectations. In the first week the drafts need more editing while the clinician learns the tool's habits, and some people nearly gave up before that curve flattened. Naming the curve in advance kept them going.

What measurable difference has it made, and how confident are you in those numbers?

The clearest change is time. Across physicians who use it regularly, we estimate roughly two hours a day returned from documentation, though I would frame that as an average with wide variation rather than a guarantee. A high-volume primary care day saves more than a light specialty clinic.

Just as important, notes are increasingly finished the same day, often before the physician leaves, which is what actually reduces the after-hours work that drives burnout. The third change is harder to quantify but is the one physicians mention first: they are looking at the patient instead of the screen for more of the visit. I want to be careful with these figures. They come from our own experience and from clinician-reported time, not from an independent audit, and the gains depend on the person and the specialty. What I can say with more confidence is the direction. The clinicians who stayed with it are measurably less likely to be doing paperwork at nine at night, and that is the outcome we were trying to buy.

What would you say to a skeptical colleague, and where does the tool still fall short?

I would tell them their skepticism is appropriate and worth keeping. The main risk is not that the tool is bad. It is automation bias, the tendency to skim a fluent draft and sign it because it reads well. A confident, articulate note that is subtly wrong is more dangerous than an obviously bad one, so the discipline of genuine review has to hold even after the novelty wears off.

The tool also has real limits. It performs worse with patients who do not speak English or who have heavy accents, and it depends on a decent device and connection, which is not universal. It does not understand medicine. It reflects patterns in language, so nuance, uncertainty, and the physician's actual reasoning still have to be supplied by hand. I would also caution against expecting week-one results. There is a learning curve. The honest pitch is not that it writes your notes. It is that it gives you a strong first draft, so you spend your evening with your family instead of with the record.

Is there a specific moment that captured why this mattered?

One that stays with me involved a follow-up with an older patient who had received a difficult diagnosis. Ordinarily, part of my attention in a conversation like that is on the record, making sure I capture the history and the plan while the person is still speaking. That day I did not touch the keyboard. I sat with her, we talked through what the diagnosis meant and what came next, and I was fully present for a conversation that deserved it.

Afterward the draft note was waiting, and it had captured the clinical substance accurately enough that my edits were minor. To be clear, I still read it line by line and corrected it before signing, because that is the rule and it is a good one. But the visit itself was different. I had given her the thing that is genuinely scarce in medicine, undivided attention at a moment that mattered. That is the case for this technology in a single encounter. It is not that it documents faster. It is that it lets you be a physician in the room.

Looking ahead, what does this mean for your practice and for the profession?

I expect ambient documentation to become ordinary, the way electronic records themselves did. The strategic question is not whether to use it but how to govern it well, because practices that adopt it carelessly will trade a documentation problem for a quality-and-liability one.

For us, the near-term value is retention and capacity. If experienced physicians are less exhausted, they stay longer and can care for more patients without our simply adding hours to their day. Over a longer horizon, I expect these tools to help with more than the note, drafting orders or referral letters for a clinician to approve, always inside the same review discipline. What I do not expect, and would resist, is any drift toward letting the software practice medicine. The value depends on the human staying accountable and staying present. If we hold that line, the technology strengthens the relationship at the center of care. If we let it erode, we will have automated the wrong thing. My advice to peers is to adopt deliberately, govern seriously, and keep the physician's judgment in charge.