Our AI Scribe went into open beta yesterday. There is a walkthrough inside the tool if you want the how. This is the why, because with a scribe the design is the argument, and if you have already tried one and been disappointed by it, the design is the only part worth your time.
Speed was a real problem and it was worth solving. So the scribes arrived, the typing went away, and that part genuinely worked.
Something else went away with it. The note used to be where the thinking happened, the place where you decided whether the sleep diagnosis still fit now that the night shift had ended, or whether the sadness under the anger was a new diagnosis or the same one seen more clearly. When a tool hands you a finished paragraph, that thinking does not get faster. It just never has to surface. What replaces it is reading someone else's version and deciding whether you object. Those are different activities. One builds judgment and the other spends it.
There is a cost on the clock too, and almost nobody measures it. Your work did not disappear, it moved. You read every line looking for what the tool assumed, softened, invented, or quietly left out, and auditing someone else's reasoning is slower than writing your own, because you have to reconstruct the thinking before you can check it. The tools get sold on how fast the draft appears. The draft was never the hard part.
Here is the part that matters. Any scribe that writes an assessment you did not dictate is making those calls for you. Every one of those questions gets answered by something, and the answers are clinical judgments with real consequences for the chart and for the claim. A black box answers them silently, and you find out what it decided by reading the paragraph it already wrote.
So the flip is simple. Surface the calls, let the clinician make them, then write.
You finish the visit and hit Draft, and it writes the HPI. I will be straight about that part. The HPI is the easy one. It is rearrangement. Whether you typed as you went or the room was recorded, the content already exists and the work is putting it in order. Getting it into your template and your voice is the harder half, which is why setup asks for a blank template and one completed note you already like, and it is still rearrangement. That is exactly the kind of work worth handing to a machine. You read it, you edit any line, and your edits are what get used. This is what most AI scribes do, and what they can do reasonably well.
After our HPI has been drafted, you tell it in one or two lines what you did this visit and why: a medication change with your reasoning; a continuation, or just no changes today; labs or monitoring you ordered; a referral you made, or a follow-up interval you moved up.
Then, before it writes a word of the assessment, it stops.
That stop is the whole design. The assessment is the first thing in the note that requires reasoning, context, and history. Some of which only you may know. It is the point where every scribe, this one too, stops recording the visit and starts interpreting it. And once an interpretation is written down, it reads like the only one there ever was. The chart shows the diagnosis you ended up with. It does not show that there was a choice, what the other options were, or whether anybody weighed them. The fork disappears into the paragraph. And a model is very good at finding the pattern in what it has been handed. It has nothing to work with when the thing that matters, your reasoning and your read of the person, was never handed to it, so it guesses and it assumes. So I put the stop exactly on that seam. Everything before it is largely organization, and the tool does it for you. After it, the tool can still write the paragraph. It should not be the one solely deciding what the paragraph says.
Before I describe what it asks you, it is worth saying why that section is worth protecting in the first place.
It is the one section of the note built to hold your reasoning. It is where you tell the next person who opens the chart why you did what you did. Why this diagnosis and not the one next to it. Why a 10 mg Adderall booster in the afternoon for someone already taking 50 mg of Vyvanse in the morning. Why you left the dose alone when the patient is still symptomatic. A covering clinician, the prescriber who inherits the patient in two years, an auditor, all of them are reading for the same thing, and none of them can get it from a diagnosis list and a medication list.
Most of the time, though, the next person opening that chart is you. It is the assessment that reminds you, three months from now with eleven patients ahead of this one, that he lost his father in a car accident in March, or that she was neglected as a child and that is the frame everything else sits inside. None of that lives in a problem list. You either wrote it down once, in the only section built to hold it, or you did not, and you walk in without it.
There is money in it too. A 99214 or a 99215 rests on the complexity of the problems you addressed, the data you considered, and the risk carried by your decisions, with the level set by two of those three, and a diagnosis list and a medication list rarely make any of that visible. Two visits can be identical in what happened and differ by a full level in what the note supports. The payer does not have your note when the claim is first adjudicated. It becomes evidence later, when the level is questioned or you decide to appeal, and that is when the reasoning either exists or it does not. Sometimes it does not. Often enough it is there and you are not sure it holds, so you do not spend the hour on an appeal you might lose. That second one is solvable, and it is most of what the Chart Audit and Coder is for.
Most of us stopped writing the reasoning down, and typing is why. It is seven at night, the notes are still open, and the difference between finishing now and finishing in another hour is three or four sentences of reasoning in each one. Everything else in the note can be clicked, templated, or carried forward. The reasoning is the only part you have to build from nothing every single time. So it goes first, and the psychotherapy section goes with it for the same reason. Nobody decides to do that. It happens one note at a time, each shortcut defensible on its own, until the assessment is a diagnosis list and the words continue current medications, and nothing in the chart says why any of it happened. That was a fair trade when the alternative was another hour after clinic. It is not a fair trade now, because that hour is the thing AI actually gave back, and the more of the note a machine writes, the more the reasoning matters, because it is the only part of the chart that shows a person was thinking. Take it out and what is left is a record of what was done, by someone, for reasons nobody wrote down.
So back to that stop. It puts a handful of cards in front of you, one for each place your note could honestly go more than one way and the direction changes the chart or the claim. You have been carrying a shift work sleep disorder and the patient just left the night shift, so does that code carry forward, get revised to plain insomnia, or come off. Your HPI describes grief underneath the anger about a business he lost, so is that an adjustment disorder, or is it reactive and already covered by the anxiety diagnosis you have, or does it stay a description with no label attached. Appetite dropped ten days after a GLP-1 start in someone long established on a stimulant, so which one owns it, and if the honest answer is both, you say both, because the cards let you pick more than one. It also asks whether psychotherapy happened and which add-on the documented time supports, and how much of your reasoning you want on the page, because note length is a clinician decision and not a model habit.
It shows you its best-supported reading of each and asks you to confirm or change it. It writes from your answers, and the cards stay on screen, so you can change one and regenerate without starting over. How many you get depends on what is actually unsettled in the visit. A clean continuation with nothing ambiguous in it asks you nothing and just writes.
And sometimes it puts up a question you had not gotten to yet. That is the part I did not anticipate when I built it. The visit went fine, you are moving fast, and a card asks whether the sleep diagnosis still fits now that the night shift is gone, and the honest answer is that you had not thought about it. Nothing is being decided for you there. You are being handed the question at the moment it costs you fifteen seconds, rather than finding it a year from now when someone reviews the chart and asks why that code kept renewing.
This is the part people push back on, and the pushback is fair. It looks like extra work compared to a tool that just hands you a paragraph. The honest answer is that the other tool answered every one of those questions too. It just did not ask you. And because you made the calls yourself, you rarely touch the final note, which is where the time actually comes back.
There is a second version of that pushback worth answering, which is that the models will keep getting better and this will stop mattering. They will get better, and at some things they already are. It changes nothing here. Nothing at any capability level can retrieve what was never recorded anywhere, so the only way your reasoning gets into the note is if you put it there. That is not a limitation of this generation of the technology. It is the shape of the problem.
A tool that meets the visit only at the end may know everything that was said in it. What it cannot do is prepare you before the patient walks in, or put an unresolved question in front of you while there is still time to ask it. A question raised after the visit is a note. Raised before it, it is care. So the rest of the design follows from that first decision.
We have all learned to work inside our EHR, and that habit is the thing to set down first. This works differently. You work the visit here, and your EHR goes back to doing what it is actually good at, which is storing and filing the finished note.
That is a bigger shift than it sounds like, because once the visit happens inside one window, everything you normally leave the note to go find can live in there with you. Screeners drop into their own section ready to score, and the interaction checker, safety plan, LAI guidance, letters, and your own pinned links open beside the note rather than in place of it.
It starts before the patient walks in. You hand it last visit's note, the HPI, the assessment with the diagnosis list, and the current medication list. It does not reach into your EHR, and it does not need to. From that it gives you a short snapshot of where you left off, and then it turns last visit into today's questions.
That second part is the piece I use most. Last visit you learned he was about to put in two weeks' notice at the overnight job. You learned Mounjaro was pending a PCP appointment. You learned the trazodone was weekends only and he had never tried it on a weeknight. Those are the open threads, and in real practice they are exactly what gets dropped when you walk in cold and the patient opens with something else. The Scribe carries them forward as a checklist you answer inline, either by typing under each line or, if you are recording, by simply asking them out loud and letting the transcript fill them in. Continuity of care is a phrase we all write on forms. This is the operational version of it.
Which is the other reason to write a real assessment. The one you wrote last time is what gets read back to you now. Put the context in and your prep is worth reading. Write continue current medications and there is nothing to hand back.
The HPI can only reorganize what you documented, and that is enforced in the drafting pass rather than promised in the marketing copy. It will not add a symptom, a denial, a finding, a timeframe, or a risk statement that is not in your source, and it will not fill in normal or negative defaults to make the note look complete. Thin notes stay short, because padding is the failure mode that gets people in trouble on review. Patient quotes stay verbatim, spelling gets fixed, and words do not get swapped for tidier words. Documented risk is never softened, never omitted, and never escalated past what you wrote.
The assessment works differently, and the difference is deliberate. It does reason across the visit, the prior assessment, and the medication decisions, because that is what a formulation is. It just does not get to do it in private, which is what the stop is for.
Once you put checklists and therapy prompts into the working note, you have created an obvious hazard. The model can see all of that text. What stops it from reading a suggested technique and writing it up as though you performed it?
One rule, and it is the sharpest thing in the build. A checklist question or a therapy prompt becomes chart content only through what you actually typed underneath it. The question itself, the technique name, the rationale, the sample script, none of it is ever evidence that anything was asked, done, or found. An unanswered question contributes nothing. It does not get answered for you, it does not get written up as deferred, and it does not appear in the note at all. That rule is enforced in the drafting pass and checked again in the review pass, rather than left to a vague instruction to be careful.
That single rule is what makes the working window possible. Everything else in the visit can be in front of you because none of it can walk into the note on its own.
Plenty of us do real therapeutic work inside a medication visit and then document it thinly, or not at all. Some of that is the same time pressure that emptied the assessment. The rest is a second barrier the assessment does not have. You can always describe your own reasoning, even tired, because it is yours. Naming a modality is different. Active listening, validation, and reinforcing what a patient is already doing well is supportive therapy. Reflecting something back so a patient hears their own ambivalence is motivational interviewing. Helping someone connect what is happening now to a pattern that has happened before is insight oriented work. None of it comes with a worksheet, which is exactly why it has gone invisible to us, and all of it has a name with decades of evidence behind it. Most of us are doing several of these every day without calling them anything, and a psychotherapy note that cannot show what happened is hard to defend on review no matter how good the session was.
None of which makes an empathic medication visit automatically billable. It becomes psychotherapy when you use a technique deliberately toward a therapeutic target, for time you document separately from the E and M, with the intervention and the patient's response on the page. The naming problem is real. Naming something you did not do is a different problem, and this tool is not built to help with that one.
There are two ways in. The psychotherapy guide is a per-visit toggle that gives you a suggested technique, a short how-to, and sample prompts folded into your note. The Therapy Coach is the live version, a set of brief-therapy moves matched to what this patient's note actually shows, ordered by relevance, each one giving you a focus, a few things you could genuinely say, and a technique you could teach. You pick one, add it to your note, and it lands in its own section with a line for what you did and how the patient responded.
The order matters here, and it is deliberate. You are handed a move you can actually make, you make it, you write one line about the response, and the tool assembles the add-on note in the structure the code requires, documented separately from the E/M with the time you recorded. You are not reverse engineering a defensible 90833 out of a visit you have already forgotten. Whether it bills is still your call and your payer's, and the Coder will tell you whether the documentation holds up. What changes is that the work happened, the note shows it, and you got a little better at the intervention while you were doing it. You learn while you practice.
Some things do not belong to an AI at all. Your telehealth attestation is inserted in a fixed position by the code, so the model cannot reword it or strip it out, because an attestation you signed is not something a model should be rephrasing.
The same goes for your patients' data. No patient information is retained on our servers. Notes are processed transiently, recordings are uploaded only for transcription under our BAA and deleted after the transcript comes back, and the only thing saved to your Vault is the reusable template you review and approve, never the patient note it was built from. For crash recovery a working draft is held on your own device, it self-expires after eighteen hours, and it clears when you finish the patient.
Recording came last on purpose. Ambient capture was the most obvious thing to add and the most dangerous, because a listening tool that fabricates is worse than no tool. It went in after the anti-fabrication work was already load-bearing, and it is still optional. Typing and pasting are always on.
The finished note is what you take into the Chart Audit and Coder, same patient, same visit. They are separate on purpose. You review and sign your note first, then reach for the Coder when you actually want it, and as you get sharper you lean on it less and use it for the calls you are unsure of.
It reads your note the way an auditor would. One pass extracts the coding facts. Another checks those facts back against the note rather than forming its own competing impression. The MDM levels are then computed in code under the 2-of-3 rule rather than guessed, so once the coding facts are confirmed, the level calculation itself does not drift. A separate documentation audit looks for contradictions, unsupported statements, and thin rationale. It also understands the temporal structure of a follow-up note, so it will not flag your pre-visit HPI dose against the post-change dose in your plan, which is the most common false positive in every auditing tool I have used. And it will not inflate your risk level just because a Schedule II appears in the list.
Run it a few times and something else happens. You start seeing the reasoning before the tool shows it to you, and you find out where your own notes run thin. It is a teacher as much as a tool.
The ROS, MSE, and Plan sections are the least finished part of this and I want to be straight about that rather than let you discover it.
I have them turned off for myself. When I used them they read as noise, and it is faster to update those sections by hand than to review what came back and decide whether I agreed with it. Once that is true, the automation is not earning its place. There is also a real tension underneath. An ROS and an MSE are attestations of what you asked and what you observed, which means every line needs your eyes on it before you sign. That is a different kind of promise than the HPI makes, and I have not decided whether the difference is acceptable or whether these should work some other way entirely. The Plan sits in the same bucket.
So they are toggles. Turn them on and see what you think, or leave them off like I do and keep owning them yourself. If someone has a clear picture of how these should behave, I would rather hear it than keep guessing.
The other open question is the therapy piece. The psychotherapy guide and the Therapy Coach are two answers to the same problem and I do not know yet whether one of them should win. Try both and tell me which one you would keep.
For the next two weeks this is open to every paying member, forum tier included. The guided walkthrough inside the tool runs the whole flow on a demo patient if you want to see it before you use it. Start with one or two patients to get the feel of the flow, then let it into your regular day once it clicks. Tell me what is off, what is missing, what you would never give up, and where you land on guide versus Coach, in the Bugs and Feedback thread: https://thinkbeyondpractice.com/platform?post=f7101242-cfb9-428a-8af0-a64cd97f0bcb
Nothing runs until you have signed the current BAA and accepted the Terms, and the tool walks you through both the first time. You review and sign every note yourself. You are the clinician of record and that does not change.
~Michael