How Physicians Use AI for Clinical Notes Safely, Without Risking Patient Privacy
A plain look at AI clinical documentation, two real health-system studies, and the de-identify-draft-verify workflow that actually saves charting time while keeping you the author of record.
Key Takeaways
- AI clinical documentation means using a tool to help draft and organize the written record of a visit. A general assistant like Claude can shape a rough set of notes into a clean draft, but the physician reads, corrects, and signs.
- Real systems are getting time back. The Permanente Medical Group reported more than 15,000 documentation hours saved across about 2.5 million encounters in a year.
- A randomized trial at UCLA Health is the honest counterweight. One scribe tool cut time in the note by roughly 10 percent, the other showed no clear time gain. The result you get depends on the discipline around the tool, not the tool by itself.
- The hard rule for a general AI tool is simple. No patient identifiers ever go in. You de-identify first, draft with the visit content stripped of names and numbers, then type the finished note back into the EMR yourself.
- No general chatbot is a certified HIPAA-compliant medical record. The clinician stays responsible for every clinical fact, and your organization's policy and any signed agreement come first.
Source: The Leveraged Years Briefing. Permalink
Charting is the part of the day that follows you home. The visit ends, the patient leaves, and the note is still open. For a lot of physicians the question about AI is not whether it sounds impressive. It is whether it can give back an hour of the evening without creating a privacy problem or a wrong note you have to defend later.
This briefing is about that practical middle. Not the version where AI replaces the clinician, and not the version where it is a gimmick. The honest version is that a general AI assistant can take the rough material of a visit and turn it into a clean first draft of a note, fast. The catch is that the safety lives entirely in how you feed it and how you check it. Get the workflow right and you save real time. Get it wrong and you have handed protected health information to a tool that was never meant to hold it.
What is AI clinical documentation?
AI clinical documentation is the use of an AI tool to help produce the written record of a clinical encounter. In its purest form, a purpose-built ambient scribe listens to the visit and generates a draft note. A more flexible version, and the one most individual physicians can actually try today, is using a general assistant like Claude to do the writing work after the fact. You give it the substance of the encounter, in a de-identified form, and it returns a structured draft: a clean history of present illness, an organized assessment, a readable plan.
The thing to hold onto is what the tool is doing and what it is not. It is drafting and organizing language. It is not practicing medicine. It does not decide the diagnosis, it does not pick the plan, and it does not get to be the final author of the record. You do. The model is a fast writer with no clinical accountability, which is exactly why a physician reads every line before it becomes part of the chart.
Does AI actually save charting time?
Yes, with a real and useful asterisk. The best large-scale evidence and the best skeptical evidence point the same direction: the time savings are real, and they are earned, not automatic.
The case that it works at scale
The most cited example is The Permanente Medical Group, the physician arm tied to Kaiser Permanente in Northern California. In a study published in NEJM Catalyst, more than 7,000 physicians used an ambient AI scribe across roughly 2.5 million patient encounters over about a year. The reported result was an estimated 15,700 hours of documentation time saved for the physicians who used it, a figure the American Medical Association rounded to about 15,000 hours in its own write-up. Just as important for anyone weighing the human side, a large share of surveyed physicians said the tool improved their ability to connect with patients during the visit and raised their job satisfaction. The system was clear about the mechanism. The AI drafts the note. The physician edits it before it enters the record.
What the Kaiser result actually shows
A draft you correct is faster than a blank page you fill. At scale, that gap added up to thousands of hours and a calmer visit. The tool did the typing. The physician kept the pen.
The case for staying realistic
Now the counterweight, because it is the more useful number for setting expectations. UCLA Health ran a randomized clinical trial of two commercial ambient scribes across 238 physicians, 14 specialties, and around 72,000 encounters. One tool reduced the time physicians spent in the note by roughly 10 percent. The other showed no clear time benefit over usual care. Both tools were associated with modest improvements in burnout and workload, which matters, but the headline is the split. Two tools, same study, very different time results.
The lesson is not that AI fails to save time. It is that the savings depend on the tool, the specialty, and the discipline of the person using it. A scribe that produces a sloppy draft you have to rewrite saves nothing. A clean draft you trust enough to lightly edit saves a lot. The difference is the workflow and the review habit, which is the part you control.
Is it safe to use AI for clinical notes?
It can be, and it can also be the fastest way to create a compliance incident. Safety here is not vague. It comes down to two questions. What did you put into the tool, and did you check what came out.
The most important fact for an individual physician using a general assistant is this: no mainstream consumer AI chatbot is, by itself, a certified HIPAA-compliant medical record system. Some vendors offer business agreements and enterprise configurations that change the picture, and large health systems run scribes under formal contracts and review programs. That is a different situation from a doctor opening a public chatbot on a personal account. Until your organization has confirmed otherwise in writing, you should assume a general tool is not a safe place for protected health information.
What should never go into an AI tool?
Keep a short list in your head, and treat it as absolute when you are using a general assistant on your own. The following should never be pasted into a general AI tool: the patient's name, date of birth, medical record number, address, phone number, or any direct identifier. Also out: full dates tied to the patient, photographs, and anything that would let a reader work out who the person is. If a colleague could identify the patient from what you typed, it does not belong in the tool.
The good news is that almost all of the writing value survives de-identification. The model does not need a name to organize a history of present illness. It does not need a medical record number to structure an assessment and plan. You can describe a 60-year-old patient with a two-week cough and the relevant findings, get an excellent draft, and never expose a single identifier.
Stop before you paste
De-identify first, every time. Strip names, dates of birth, record numbers, and contact details before any visit content goes near a general AI tool. A de-identified input is a safe input. This one habit prevents the privacy problems most physicians are right to worry about.
What does a safe workflow look like?
The safe pattern is small and repeatable. Think of it as four steps that keep the identifiers out and keep you as the author of record.
The two-window habit
In practice this works as a two-window routine. One window is the AI tool, where everything is de-identified. The other is the EMR, where the real patient record lives. The two never touch directly. You draft in the de-identified window, read the draft critically, and then write the finished, accurate note into the EMR by hand. Nothing is copied from your chart into the chatbot, and nothing is pasted from the chatbot into the chart without you reading and owning every word. That gap between the two windows is where the patient's privacy stays protected.
Verify every clinical fact
AI writes fluently even when it is wrong. It can smooth over a detail, invent a finding you never mentioned, or state a dose with false confidence. So the verification step is not optional polish, it is the clinical safeguard. Read the draft as the clinician of record. Confirm the medications, the doses, the allergies, the assessment, and the plan against what actually happened in the room. If a line is not true, it does not go in. You are accountable for the note, so you read it like you wrote it, because you are about to.
The rule that keeps it safe and useful
The AI drafts and organizes. You de-identify the input, verify every clinical fact, and remain the author of record. Confirm your organization's policy and any signed agreement before a tool touches anything close to real patient data.
Where does AI not belong in the note?
There is a clear line. AI should not be the source of a clinical fact. It does not decide the diagnosis, it does not choose the medication or the dose, and it does not get to assert something happened in the visit that you did not confirm. It is a drafting aid for language and structure, working from material you supplied. The clinical content is yours, the judgment is yours, and the signature is yours. A tool that drafts is an asset. A tool that you let invent clinical facts is a liability, and the difference is entirely in how you use it.
A realistic first week
If you want a safe place to start, do not rebuild your whole documentation routine at once. Pick one note type that you write often and that is straightforward, and try the four-step workflow on it for a few patients at the end of clinic. De-identify the content, let the tool draft, read it hard, and type the final note into the EMR yourself. Notice how much editing it actually needed. That edit time is your honest measure of whether the tool is saving you anything, which is the same lesson the UCLA trial taught at scale. Build from the note types where it clearly helps, and skip the ones where it does not.
Done this way, AI charting is a series of small, reviewed wins that give you evening time back without putting a patient's privacy at risk. If you want the full method laid out step by step, with the de-identification habits, the prompt patterns, and the review checklist built in, that is exactly what the course Cut Charting Time with AI: Practical Notes for Physicians is built to give you. For the broader pattern of drafting with AI while a professional stays accountable, the companion briefing on how HR teams use AI safely shows the same draft-and-review discipline in another field, and the full set lives in The Briefings.
Frequently Asked Questions
Can I put patient information into a general AI tool like Claude?
No. Treat any identifier as off limits in a general AI tool, including the name, date of birth, medical record number, address, phone number, and full dates tied to the patient. De-identify the visit content first. You can describe the clinical picture in general terms, get a strong draft, and never expose who the patient is. No mainstream consumer chatbot is, by itself, a certified HIPAA-compliant medical record, so confirm your organization's policy and any signed agreement before going further.
Does AI really save physicians documentation time?
It can, and the savings are real but earned. The Permanente Medical Group reported an estimated 15,700 documentation hours saved across roughly 2.5 million encounters in a NEJM Catalyst study. A UCLA Health randomized trial found one ambient scribe cut time in the note by about 10 percent while a second tool showed no clear time benefit. The takeaway is that results depend on the tool and the review discipline, not on the tool alone.
What is a safe AI charting workflow?
Four steps. De-identify the visit content so no identifiers go into the tool. Let the AI draft a clean, structured note. Verify every clinical fact, medication, dose, and plan against what actually happened. Then type the finished note back into the EMR yourself. A two-window habit keeps the tool and the real record separate, so the patient's privacy stays protected and you remain the author of record.
Who is responsible if the AI gets a clinical fact wrong?
The physician. AI can write a confident sentence that is simply wrong, so the verification step is a clinical safeguard, not a formality. Read every draft as the clinician of record, correct anything that is not true, and sign only what you have confirmed. The tool drafts the language. You own the note.
Sources and notes. Kaiser Permanente / The Permanente Medical Group figures: "Ambient Artificial Intelligence Scribes: Learnings after 1 Year and over 2.5 Million Uses," NEJM Catalyst (catalyst.nejm.org/doi/abs/10.1056/CAT.25.0040), reporting more than 7,000 physicians, about 2.5 million encounters, and an estimated 15,700 documentation hours saved over roughly a year, with most surveyed physicians reporting better patient connection and job satisfaction; summarized by the American Medical Association as approximately 15,000 hours ("AI scribes save 15,000 hours and restore the human side of medicine," ama-assn.org). UCLA Health figures: "A Randomized-Clinical Trial of Two Ambient Artificial Intelligence Scribes," published in NEJM AI (PubMed 40672471; UCLA Health news release, uclahealth.org), covering 238 physicians across 14 specialties and about 72,000 encounters, with one scribe reducing time in the note by roughly 10 percent and the second showing no clear time benefit. Figures are reported by the cited sources and summarized here for general information. This briefing is general information for physicians and is not legal, compliance, or clinical advice. Confirm your own organization's policy and any vendor agreement before using any AI tool with patient information.