• Practice your interview with the new SDN AI Interview Coach. Choose a school, answer by voice or typing, and receive a personalized feedback report. Available now to all SDN members. Try the AI Interview Coach.

AI scribes are amazing!

Started by Bartelby
This forum made possible through the generous support of SDN members, donors, and sponsors. Thank you.

Bartelby

Full Member
15+ Year Member
Advertisement - Members don't see this ad
I finally started using an AI scribe and wanted to share that it has had the biggest positive impact on my work day of any technology since high-quality dictation.

The scribe has gotten good enough to create a solid "subjective" portion of the SOAP note. I let it create that, skim and edit anything as needed, then dictate off an MSE, labs, assessment, and update the plan (copy/pasted). The dictated portion only takes a couple of minutes.

While all I have the scribe do is the subjective, it is so freeing! I can finally just talk to the patient without constantly typing and trying to roughly format what they tell me. The visit flows faster and more freely, and the cognitive "drain" is meaningfully less. If a visit happens to run long so that I need to start with the next patient before writing the note, I already have all the key information automatically pasted in so that I can finish off the note in a few minutes later in the day instead of plumbing my memory for all the details (how many times did she wake up each night? How many hours of sleep was it? How much weight did he mention he lost this past month?).

I use Dax Copilot because my organization pays for it, but I think there are other free or low-cost HIPAA-compliant options like OpenEvidence that are good enough to use it the way I do.

Just wanted to share this for those of you who haven't tried it. Being able to finally be more fully "present" with patients rather than typing up notes has been a major unexpected positive.
 
Absolutely, we have Abridge and the killer app was when they let us save preferred prompts for refining the note. Since I usually do a recap + assessment + plan verbally with the patient, it can even output a good psych style integrated assessment. I now use it for all appointments, new and follow-up.
 
Absolutely, we have Abridge and the killer app was when they let us save preferred prompts for refining the note. Since I usually do a recap + assessment + plan verbally with the patient, it can even output a good psych style integrated assessment. I now use it for all appointments, new and follow-up.
How long did it take you to train your agent for Abridge? We have some residents who have been using it for 3-5 months and editing their notes is still a nightmare. Even more so because their notes are now often dissertations because of how many random (often irrelevant) details the AI is including. It’s also problematic because it is often including details that can be counter to what our plan is.

I’d really love to be able to use AI to scribe, but the notes I’m seeing it produce after 3-4 months of training (by 2 strong residents who are quite tech savvy) are still sometimes so bad that it makes me feel like taking the time to train it is just not worthwhile. We also have abridge for scribing.
 
Advertisement - Members don't see this ad
Oh yeah...it would need to have a specific algorithm for managing resident notes, higher bar for inclusion of content. That seems doable? But you definitely can't just apply the same process to an attending and a resident interview. I know when I'm in an observed resident interview, I'm often zoning out. The AI needs to somehow do that intermittently until the resident figures out what is actually relevant and starts only asking those questions. Obviously yes, I give the resident feedback after the interview and let them know the max length of time they have with the patient, but interrupting them all the time doesn't really work with a patient there. If we're short on time, I prefer to write notes while the resident does the interview.
 
How long did it take you to train your agent for Abridge? We have some residents who have been using it for 3-5 months and editing their notes is still a nightmare. Even more so because their notes are now often dissertations because of how many random (often irrelevant) details the AI is including. It’s also problematic because it is often including details that can be counter to what our plan is.

I’d really love to be able to use AI to scribe, but the notes I’m seeing it produce after 3-4 months of training (by 2 strong residents who are quite tech savvy) are still sometimes so bad that it makes me feel like taking the time to train it is just not worthwhile. We also have abridge for scribing.
The training comment is interesting and makes me wonder if your org has a different implementation in some way, as we don't "train" it exactly. I think they do use some degree of RLHF if you choose to rate a note, but I think that's more used for feedback to the developers. They have a way of saving your favorite prompts and that is way more useful. I used to take abridge note + transcript and put it in Copilot (HIPAA secure BAA version) for rewriting but can do it in-app with Abridge now.

They're not perfect, but these are my two prompts:

INTAKE APPOINTMENT (ADHD Assessment Version):
Rewrite the source note into a psychiatrist-style new patient HPI and Assessment & Plan.


OVERALL GOAL:
Produce documentation that is concise, clinically synthetic, and easy to read. Preserve useful clinical detail from the source note, but compress unnecessary anecdotal or repetitive content. The result should be more concise than a transcript-style note while still retaining clinically useful nuance.


HPI:
- Start with a single sentence in this format:
“[Age] year old [male/female/patient] with a reported psychiatric history of {diagnoses reported prior to appointment during chart review by clinician and/or diagnoses reported by the patient}” and, if relevant "and relevant past medical history of..."
- Then start the HPI with a new paragraph.
- Write in narrative prose only, except where specifically instructed below for ADHD symptom review.
- Use short paragraphs with smooth transitions.
- Start the first paragraph with: “The patient reports…”
- In the opening paragraph, summarize the main presenting problem, relevant duration/course, major symptoms, functional impact, important recent stressors/context, and the patient’s reason for seeking care now.
- If the patient expressed hopes or goals for treatment, include that briefly.
- If collateral information was provided, incorporate it naturally. Do not mention the absence of collateral if none was available.
- Include additional paragraph(s) only for other clinically relevant psychiatric symptom domains that were actually discussed (for example depression, anxiety, trauma-related symptoms, mania, psychosis). Present these as narrative synthesis, not as a review-of-systems checklist.
- If sleep was discussed, include one separate paragraph specifically summarizing sleep pattern, duration, insomnia/hypersomnia, sleep quality, daytime tiredness, and other relevant sleep details.
- Do not include routine negative ROS language in the HPI unless it is clinically important.
- Do not include detailed past psychiatric, social, family, developmental, or substance history in the HPI unless directly relevant to the current presentation, diagnostic reasoning, or treatment plan.
- Prefer clinical abstraction over exhaustive detail: keep examples only when they clarify symptom severity, pattern, or impairment.
- Avoid transcript-like phrasing, redundant qualifiers, and long lists of minor examples.


ADHD EVALUATION RULE:
- If this appointment included an ADHD evaluation, after the narrative HPI paragraphs include a numbered inattentive symptom review section.
- The output must explicitly include the following questions as written below, with concise answers based on the interview:

Inattention
1. Do you often fail to give close attention to details or make careless errors? (Specifically focused on mistakes / errors / incomplete submissions)
2. Do you often fail to sustain attention in tasks and activities? (Distractibility, trailing off, etc.)
3. Do you often appear to not listen to what is being said to you? (Very specific, people having to verbally cue or snap at patient to regain their attention in conversation)
4. Do you often fail to follow through on instructions or to finish tasks? (Primarily focused on assigned tasks rather than self-generated tasks, leaving things undone. Alternatively difficulty following written instructions or procedures.)
5. Are you often impaired in organising tasks and activities? (This is more about difficulty sequencing and prioritizing activities than about being an "organized (neat) person.")
6. Do you often avoid or strongly dislike tasks that require sustained mental effort? (This is about avoidance / procrastination / dislike of mental tasks rather than earlier questions about difficulties faced with those tasks.)
7. Do you often lose things necessary for certain tasks or activities? (Self explanatory)
8. Do you often become easily distracted by external stimuli? (Self explanatory, distracted by other stimuli.)
9. Are you often forgetful in the course of daily activities? (This has more to do with the forgetfulness -- not remembering to do things, where things were placed, forgetting appointments, not checking to-do list, etc.)


- For each item, include both the direct answer if it was asked explicitly and the best supporting examples described elsewhere in the interview.
- If the source note contains relevant examples but no explicit yes/no answer, synthesize the answer from the available clinical content.
- Do not include explanatory parenthetical text after the questions.
- Keep each numbered answer concise but specific enough to show symptom pattern and functional impact.
- Do not omit any of the 9 questions when an ADHD evaluation was performed.


ASSESSMENT:
- Start with a single sentence in this format:
“[Age] year old [male/female/patient] with a psychiatric history of {diagnoses determined during this appointment}” and, if relevant "and relevant past medical history of..."
- Include only diagnoses/history that are relevant to the current evaluation.
- Follow with a new paragraph containing a brief synthesis of the current presentation:
- the most likely diagnostic framing,
- the main symptom clusters,
- key contextual or contributing factors,
- any meaningful diagnostic uncertainty or differential considerations,
- and the rationale for the treatment approach being chosen.
- Keep the assessment concise and interpretive rather than repetitive. Do not simply restate the HPI.


PLAN:
- Present the plan as clear bullet points.
- Include medications with exact dose, route, frequency, and titration instructions if provided. (e.g. - Bupropion XL 300mg PO Daily or - Sertraline 25mg PO Daily x 1 week then 50mg PO Daily)
- Include psychotherapy and other recommended interventions when relevant.
- Include clinically important counseling, monitoring, follow-up instructions, or contingency planning if discussed.
- Be exceptionally concise and focused.
- Be specific. Avoid unnecessary operational detail.


STYLE CONSTRAINTS:
- Use physician-level psychiatric documentation.
- Be concise, but not skeletal.
- Preserve nuance that affects diagnosis, severity, impairment, or treatment planning.
- Omit clutter, repetition, and low-yield anecdotal detail.
- Do not invent information.
- Do not mention that you are summarizing, rewriting, or reformatting a source note.


SUBJECTIVE / ATTRIBUTIONAL STYLE:
- Write the HPI primarily in subjective style.
- Attribute symptoms, beliefs, and historical interpretations to the patient unless they are independently established facts.
- Prefer phrasing such as:
“The patient reports…”
“The patient states…”
“The patient describes…”
“The patient feels…”
“The patient believes…”
“The patient wonders whether…”
- Do not convert patient-reported possibilities, self-suspicions, or tentative interpretations into definitive diagnostic statements.
- For example, write “The patient reports lifelong difficulty with attention and wonders whether ADHD may explain this” rather than “The patient has had ADHD their whole life.”
- Do not present diagnostic conclusions in the HPI unless they were already clearly established prior diagnoses and are directly relevant.
- Keep the HPI focused on reported symptoms, course, context, and impact rather than the clinician’s interpretation.
- Save diagnostic synthesis, differential considerations, and clinical conclusions for the Assessment section.

Follow-up (HPI only):
Start with a sentence that states when patient was last seen and which medication changes were made at that appointment. That first sentence will start with "The patient was last seen in {month year}..." Include information about medication changes made since that appointment if any have been made (but not medication changes made during this appointment.) There is no need to specify that changes have not been made. Do not include patient's diagnoses. Briefly describe patient's current psychiatric symptoms and note whether there has been any worsening or improvement. Briefly describe any current medication side effects. Briefly summarize any ongoing stressors. If therapy is discussed, mention who patient is seeing and how it is going. There is no need to specify if therapy was not discussed. Do not include separate sections for current medications, psychiatric history, social history or medical history.

Unless the information obtained is extremely complex, typically the total HPI should not be more than 15 sentences across 4 paragraphs, with each paragraph dedicated to a specific topic.

Context: Our department often manages patients via (PATIENT PORTAL NAME) messages or nursing calls between appointments, sometimes making medication changes in the interim. That would be the sort of medication change that would be mentioned in the HPI, not medication changes made during this appointment.

Style Point: Even if the brand names of medications (e.g. Lexapro) were used during the conversation, the generic name should be used in documentation (e.g. escitalopram.)
 
Last edited:
Convenient? Yes. But not sure how beneficial it is from a medicolegal standpoint.

Hypothetical, but imaging you’re Dr. Tufts from the Clancy trial and you used one of these softwares. That sh-t would get subpoenaed to no end. Now, if it turns out that the patient had at some point actually told you they were psychotic and you completely missed it, and the defense pulls up a recording from the app…you’re bone fk’d. They’re going through every second of whatever’s on there.

Also, the one I tried before sometimes missed important particles/negations like “not” or “don’t”. I ended up having to comb thru notes that would’ve taken me the same amount of time to write.
 
Convenient? Yes. But not sure how beneficial it is from a medicolegal standpoint.

Hypothetical, but imaging you’re Dr. Tufts from the Clancy trial and you used one of these softwares. That sh-t would get subpoenaed to no end. Now, if it turns out that the patient had at some point actually told you they were psychotic and you completely missed it, and the defense pulls up a recording from the app…you’re bone fk’d. They’re going through every second of whatever’s on there.

Also, the one I tried before sometimes missed important particles/negations like “not” or “don’t”. I ended up having to comb thru notes that would’ve taken me the same amount of time to write.
The recordings and transcripts are supposedly deleted after something like 2 weeks.

I have not had the sort of issues you describe. Yes you should read through your note to ensure accuracy. When we first got abridge 2? years ago, I had to do so much editing that I swore it off. Then Abridge released their mental health note version and that saved a good bit of time, but still needed a lot of editing for my preferred style and content. Now that I can just click one button and drop in my preferred formatting and phrasing prompt, I rarely have to do more than add or remove one or two sentences and usually that's just for subjective preference reasons, not omissions or commissions.
 
The recordings and transcripts are supposedly deleted after something like 2 weeks.

Supposedly but I suppose we'll find out how true that is when a company inevitably gets subpoenaed for the data.

It's like virtual appointments, the more digital access there is the more monitoring that can happen. Oh that appointment was 28 minutes long on the virtual platform but you said it was 30? Looks like we're taking back your reimbursement until you resubmit that code properly.

However, I will say in one-party recording states you run that risk anyway. If you're in a one party recording state, patients can certainly record your entire visit without your knowledge, keep it forever and it's entirely legal. Luckily I live in a two party recording state lol.
 
Last edited:
The training comment is interesting and makes me wonder if your org has a different implementation in some way, as we don't "train" it exactly. I think they do use some degree of RLHF if you choose to rate a note, but I think that's more used for feedback to the developers. They have a way of saving your favorite prompts and that is way more useful. I used to take abridge note + transcript and put it in Copilot (HIPAA secure BAA version) for rewriting but can do it in-app with Abridge now.

They're not perfect, but these are my two prompts:

INTAKE APPOINTMENT (ADHD Assessment Version):


Follow-up (HPI only):

I like these prompts. One thing I'll note for others looking to do something similar is that it is absolutely possible to take (de-identified) notes of yours that are particularly good examples of what you like to see in a note and ask a frontier model to come up with a prompt like this to capture what it is you want an agent to do. Far easier to tweak something like that to your specifications than to write a prompt from scratch yourself and probably more effective unless you've spent a fair amount of time doing prompt engineering with the agent in question.

I am also curious about the training comment, I would be surprised if there was a lot of reinforcement learning being done on an individual user basis because that would get expensive quickly.
 
I like these prompts. One thing I'll note for others looking to do something similar is that it is absolutely possible to take (de-identified) notes of yours that are particularly good examples of what you like to see in a note and ask a frontier model to come up with a prompt like this to capture what it is you want an agent to do. Far easier to tweak something like that to your specifications than to write a prompt from scratch yourself and probably more effective unless you've spent a fair amount of time doing prompt engineering with the agent in question.

I am also curious about the training comment, I would be surprised if there was a lot of reinforcement learning being done on an individual user basis because that would get expensive quickly.
That's exactly what I did ^ (Except since it was a HIPAA/BAA secure instance, I didn't have to fully deidentify.)
 
It's interesting hearing the range of experiences here. I think sharing which scribe you used and when you last tried it may help guide people about what to try and what to avoid. I have seen good results from Dax Copilot and OpenEvidence (I didn't use OpenEvidence but saw it in action for someone else).

Admittedly there are hallucinations at times. My favorite (because it is so laughable) is that the model sees notes always start with "56 year old man presents for..." so, because no one says the patient's age out loud, the model just guesses an age! Fortunately those kids of errors are pretty rare, and I can tweak the note into good-enough form with little effort.
 
I am also curious about the training comment, I would be surprised if there was a lot of reinforcement learning being done on an individual user basis because that would get expensive quickly.
Literally watched our resident training the AI not to phrase statments a certain way in assessment and plan sections of the note. He said he’s spent a lot of time over the last 3-4 months trying to train the AI to document the way he wants.

It’s fully possible that because he’s working in so many different settings the AI may have difficulties writing notes based on the setting or goals of treatment, but idk.


It's interesting hearing the range of experiences here. I think sharing which scribe you used and when you last tried it may help guide people about what to try and what to avoid. I have seen good results from Dax Copilot and OpenEvidence (I didn't use OpenEvidence but saw it in action for someone else).

Admittedly there are hallucinations at times. My favorite (because it is so laughable) is that the model sees notes always start with "56 year old man presents for..." so, because no one says the patient's age out loud, the model just guesses an age! Fortunately those kids of errors are pretty rare, and I can tweak the note into good-enough form with little effort.
Like I said above, we have Abridge. It’s okay but I have been far from impressed. I find it easy to tell which attendings are using AI, and frankly I pay less attention to their notes because they’re bloated (even by academics standards) and sometimes inaccurate.

As of now, the time it would take to produce notes I’d consider acceptable without significant editing isn’t worth it. Especially since dot phrases and dictation in Epic already make note writing significantly easier.
 
There’s a strategic value of AI, especially in the outpatient psychiatry setting not mentioned yet in this thread. An administrative grind is the repetitive cognitive labor of mapping an encounter to E/M guidelines to justify moderate complexity, particularly for 'stable' established patients. When you're trying to consistently secure Level 4 billing (like a 99214 for medication management, or 99214 + 90833 with psychotherapy) for someone whose chronic conditions are relatively stable, explicitly spelling out the Medical Decision Making (MDM) by hand or even dictation every single time takes a toll.

While dot phrases and dictation in Epic might seem like the easiest fix for these routine follow-ups, relying on them heavily is actually a vulnerability. Using the exact same static macro to document prescription management across dozens of stable patients creates cookie-cutter notes that auditors specifically target as 'cloned documentation.' A properly utilized AI scribe dynamically weaves the distinct nuances of that specific patient's visit into the rigid MDM framework. When calibrated correctly, that extra text isn't bloat—it's audit armor that explicitly defends your moderate complexity coding without triggering the red flags of copy-pasted templates.
 
Advertisement - Members don't see this ad
Literally watched our resident training the AI not to phrase statments a certain way in assessment and plan sections of the note. He said he’s spent a lot of time over the last 3-4 months trying to train the AI to document the way he wants.

I guess the question is when you say "training", what do you mean, exactly? This could refer to a bunch of conceptually disparate things with different consequences.
 
There’s a strategic value of AI, especially in the outpatient psychiatry setting not mentioned yet in this thread. An administrative grind is the repetitive cognitive labor of mapping an encounter to E/M guidelines to justify moderate complexity, particularly for 'stable' established patients. When you're trying to consistently secure Level 4 billing (like a 99214 for medication management, or 99214 + 90833 with psychotherapy) for someone whose chronic conditions are relatively stable, explicitly spelling out the Medical Decision Making (MDM) by hand or even dictation every single time takes a toll.

While dot phrases and dictation in Epic might seem like the easiest fix for these routine follow-ups, relying on them heavily is actually a vulnerability. Using the exact same static macro to document prescription management across dozens of stable patients creates cookie-cutter notes that auditors specifically target as 'cloned documentation.' A properly utilized AI scribe dynamically weaves the distinct nuances of that specific patient's visit into the rigid MDM framework. When calibrated correctly, that extra text isn't bloat—it's audit armor that explicitly defends your moderate complexity coding without triggering the red flags of copy-pasted templates.

I just don't get how this is any simpler than what I do in my plan (not assessment) which is literally going:

1) MDD, recurrent- Chronic, exacerbation/progression
- X treatment change or no treatment change and why

2) ADHD, combined- Chronic, stable
- X treatment
- X treatment

Instructed to folllowup: X weeks

You just literally spell out what chronic conditions you're treating, what you're treating them with and why it meets low vs moderate MDM criteria. It copies forward every visit and I just change the stability of them.



I'm the odd person out in that I just haven't found the scribes to work right enough of the time. I've tried using Heidi and OpenEvidence multiple times and even feeding in various prompts/note examples. My biggest issue is that it'll just make stuff up or draw weird conclusions that I have to catch and you'll never know where it'll stick them.
Most recently, it threw a line in the note that "Her developmental age is approximately one year behind her chronological age due to an early September birthday, which may impact emotional regulation." which is not at all what I said....I said to this patients mom that she's an early september birthday so she's almost a year developmentally behind the oldest kids in her grade. Just random stuff that I would never type or put in my notes or things it mishears I have to catch.

I honestly would love to get them to work somehow because I hate doing notes.
 
Last edited:
I guess the question is when you say "training", what do you mean, exactly? This could refer to a bunch of conceptually disparate things with different consequences.
Editing prompts, instructing on “if, then”, note formatting, instructions on tone, phrases, and ordering. Honestly looked a little like coding to me.

The resident said he was pretty far past the phase of just feeding in examples of notes for the AI to learn from (said that was an early step and fed a couple hundred as examples). Was trying to tweak notes to actually say what he wanted them to but was still getting essays with extraneous infos with random errors. I will add that this resident is nothing like me in the sense that he is not OCPD at all, if anything is the opposite of rigid which makes this even less appealing to me as someone who’s pretty darn OCPD with how I document.