How AI Notes Actually Work (And Why They Won't Replace Your Clinical Voice)

There's a Reddit thread I know about. Or several, actually. Therapists arguing about AI in clinical practice. Someone brings up the idea of using AI transcription for progress notes and within hours you get the responses. "It's going to make my notes robotic." "The AI won't catch the nuance." "What if it misses something important?" "Isn't the machine training on my client data?" "If the AI writes my notes, am I even doing therapy anymore?"

I built Practice Harbor. I built the AI transcription tool inside it. And I want to tell you directly what I've learned in hundreds of conversations with therapists about this exact fear: the fear that using AI for documentation somehow makes you less of a clinician.

It doesn't. But the way it's sold to you sometimes makes it sound like it might.

What Actually Happens in the Room

Let me start here, because this is the part nobody really explains. When you're using AI notes, the AI isn't writing your note in real time. It isn't making clinical decisions. It isn't watching your client across from you and trying to understand the moment.

You are. You're doing all of that exactly the way you always have.

Here's the technical part, in plain language: the session is recorded. Just audio. Not video. Audio gets transcribed — converted from speech to text by a machine that has learned what human voices sound like. The transcript is fed into a large language model that has learned what therapy notes look like, what information matters clinically, what goes in a SOAP note or a DAP note or a BIRP note. The model reads the transcript and generates a draft note with subjective data pulled from what the client said, objective observations inferred from their words, assessment section written, plan documented.

All of that happens after the session is over. While you're sitting with the emotional weight of what just happened, or between clients, or whenever you get a chance to look at it. Not during.

The draft note appears in your system. You read it. You edit it. You add things the AI missed. You remove things that don't belong. You move paragraphs around. You change language that doesn't sound like you. And then you sign it. You attest that this is your clinical work, your clinical judgment, your record.

The AI didn't make a clinical decision. It made a draft. You made the decision.

What the AI Actually Misses

Let me be honest about what it can't do. The AI reads the transcript. It doesn't feel the shift in your client's breathing. It doesn't see the moment when their eyes got wet but they pretended not to notice. It doesn't know that the thing they didn't say — the thing they've been circling around for three sessions — finally came into view in minute 37, and that's why you made a different intervention choice than you normally would.

These things are invisible in a transcript.

A therapist who uses AI notes has to do one critical thing: review the draft and add the texture. "When discussing her mother, client became tearful. She was silent for thirty seconds." These details don't appear in the transcript. The client said words; they didn't say "I'm feeling emotional about my mother." But you saw it. Your nervous system felt it. You need to put it in the note.

This is actually fine. This is what you were going to do anyway — translate what happened into words. Except you're not doing it from a blank page. You're editing something. And editing is forty percent of the effort of writing from scratch.

The AI also sometimes makes confident wrong guesses. It will misidentify a name (heard "Steve" instead of "Stefan," and now your note has family dynamics of the wrong person). It will occasionally miss the clinical significance of something that seemed trivial to the transcription model but mattered to you. It can format things awkwardly, put information in the wrong section, misunderstand the structure of your chosen note format.

All of these are things you have to catch. So the note you're reviewing isn't "done" in the way a human clinician's note would be done. It's 75-85% there. You add the last 15-25%.

And here's the thing: you're better at that 15-25% when you're not exhausted.

The Fear Nobody Wants to Say Out Loud

I know what you're actually worried about.

You're worried that if AI writes your notes, you stop being a real therapist. That you're outsourcing something core. That you're taking a shortcut. That your notes will sound like a machine wrote them and that someone reading them will know you didn't spend the time doing it yourself.

This is the fear that doesn't get named in the technical discussions. It's underneath everything. A therapist who already feels like they're not doing "enough" — not seeing enough clients, not keeping up with case notes, not charting timely, not getting home until the kids are asleep — sees AI notes as the final proof that they're failing at the job.

But this is backwards. The moment you need to be the most clinical, the most precise, the most present — that's in the session. You're already doing it. The documentation after the session is over is the record of what you did, not the thing itself.

Using AI to draft your documentation doesn't mean you're practicing less therapy. It means you're spending time on therapy instead of spending time on typing.

What Happens to Your Data

There's another fear. Hospitals have been training AI on patient data for years. Therapists remember those stories. You're worried that your client's session — the private thing that was supposed to be confidential — is now part of some machine learning training dataset, feeding an AI that will eventually be sold to insurance companies or tech startups or worse.

Let me be direct: that's not how Practice Harbor works, and if you're using a tool that works that way, you should stop.

Here's what actually happens. The audio file arrives at Practice Harbor's servers. It's encrypted in transit. A transcription model converts it to text. A language model generates a note draft. The audio file is deleted. Completely. It doesn't sit on a server. It doesn't get archived. It doesn't get used to train the transcription model or the language model.

The note itself stays in your Practice Harbor account. It's part of your client chart, encrypted, protected by HIPAA, covered by Business Associate Agreements. If a law enforcement agency wants access to it, they need a warrant. If an insurance company wants access, they need authorization from you and your client.

The data doesn't flow to a training pipeline. Your client's story isn't becoming part of the next version of the AI model. You're not making your clients part of a surveillance infrastructure. It's a tool. A private tool. The data is yours.

What a Real AI-Drafted Note Actually Looks Like

Maybe the easiest way to understand this is to see what actually comes out the other end. Here's a real (anonymized) note from Practice Harbor. This is what the AI generated, before the therapist edited it.

That's a good draft. Not perfect — the therapist might change "his perspective" to be more specific, might add that the client cried when discussing her mother (AI picked up softer tone but doesn't capture tears), might shuffle the order of information slightly. But it's 80% of a finished note. The therapist reads it, makes edits, adds the detail only they noticed, and signs it. Total time: five minutes.

Without AI, that therapist writes it from scratch. They start with a blank page. They reconstruct the session from memory. They organize the information into SOAP format. They make sure the language is clinical but readable. They check for spelling and grammar. Total time: 15-20 minutes.

The note itself is probably similar in quality. Maybe slightly better if the therapist wasn't tired. But they just got 10-15 minutes of their evening back.

The Real Question

Here's what I think you should actually be asking, because it's the question that matters:

Will using AI notes make my documentation worse?

The answer is: probably not. If anything, it gets better, because you have more energy left after five minutes of review than after twenty minutes of reconstruction. You're less likely to skip details. You're less likely to write vague notes because your memory is fuzzy. You're more likely to catch the nuance and add it, because you're not cognitively exhausted.

Will it change what happens in session?

No. It won't. You're still sitting across from your client. You're still tracking their affect, their words, the thing they almost said. You're still making interventions based on your clinical judgment. The AI isn't in the room. It comes after. It's the documentation tool. It's not the therapy.

Will it replace my clinical voice?

It can't. Not unless you let it. The draft note is a skeleton. You put meat on the bones. You add the language that sounds like you. You make the clinical calls about what matters and what doesn't. You edit out anything that feels off. The final note has your signature on it, which means it has your judgment in it. Your clinical voice is in the edits, in the decisions about what to keep and what to cut and what to emphasize.

But here's what I'll tell you honestly: if you hate it, don't use it. Not every tool is for every clinician. Some therapists use voice-to-text dictation and love it. Some people still handwrite their notes. Some people use AI transcription and never look back. There's no moral superiority in any of these choices. The right tool is the one that makes documentation feel less like punishment and more like the clinical work that it is.

For most therapists I know, that tool is AI notes. Not because they're lazy. Because they're exhausted. And they're tired of spending their evenings doing work that a machine is better at — the mechanical work of transcription, of organization, of structural formatting — when they could be spending those hours on the work that only a human can do.

You became a therapist to do therapy. Not to type. If a machine can handle the typing so you can do more of the therapy and less of the documentation, that's not a shortcut. That's a sound clinical decision. Your clients get better notes because you're not writing them at 10 PM while your brain is shutting down. And you get your evenings back. Both things can be true at the same time.