Why CoralEHR Doesn't Require Recording Your Sessions

There are two completely different things people mean when they say "AI in therapy," and therapists feel very differently about each.

The first is AI-as-therapist — a chatbot that talks to a client instead of you. That idea gets rejected, often viscerally, and for good reasons: relationship, clinical judgment, and crisis safety. The numbers back up the discomfort. A 2025 YouGov survey found only about 11% of Americans are open to AI-powered mental health tools, and just 8% say they trust AI to give accurate mental-health recommendations.

The second is AI-for-paperwork — something that helps you turn a session into a note faster. Here, most clinicians are pragmatic. They'll take the help. But they fence it with a hard line: the clinician signs and is responsible, consent is documented, and the audio or data is not retained or used to train somebody's model.

CoralEHR is built entirely on the second side of that line. And the single design decision that matters most is this: AI note help does not require recording the session.

The recording problem nobody wants to say out loud

Walk through how the major AI scribes work, using only what they say on their own pages. SimplePractice's Note Taker, Upheal, and Mentalyc all generate notes the same fundamental way — primarily by capturing the session as audio (or audio and video), transcribing it, and drafting a note from the transcript. That recording-first capture is the core of the product (some also accept dictation or a typed recap). It's a reasonable design, and for some clinicians it works fine.

But recording carries three costs that have nothing to do with how good the note is.

Recording changes the room. The moment a client knows the hour is being captured, the dynamic shifts. People self-edit. They ask where the recording goes. For trauma work, for clients with any history around surveillance, for adolescents — the microphone is not neutral. You can disclose it well and most clients will still consent, but you've added a thing to the room that wasn't there before.

Recording is a per-session consent burden. Capturing audio of a clinical session generally means getting and documenting consent — every client, and in practice every time the arrangement changes. It's a small administrative tax that recurs forever, and it's one more place for a gap to open up between what your policy says and what actually happened in the room last Tuesday.

Stored audio and transcripts are a discovery risk. A recording or transcript is a new, highly sensitive record that didn't exist before — and that can be subpoenaed. In a custody dispute, a malpractice claim, or any litigation that touches a client, a stored verbatim transcript of a therapy hour is a far richer (and riskier) target than a clinician-authored progress note. The safest record is the one that was never created.

"No recording" answers all three at once. There's nothing to consent to recording, nothing to change the room, and no stored audio to be compelled.

What "your words train the model" looks like in practice

This isn't theoretical anymore. The market leader made it concrete. SimplePractice has announced that, starting June 16, 2026, it will begin retaining de-identified, de-coupled transcripts of Note Taker sessions to help improve its AI features, including its Care Aide assistant. Per SimplePractice's own transcript-retention FAQ, clinicians who used Note Taker before that date are opted out by default, but clinicians who enable Note Taker on or after June 16, 2026 are opted in by default and have to opt out.

To be fair and precise: SimplePractice says it de-identifies and de-couples the transcripts, will never sell them, and won't share them with third parties for commercial purposes. That is a meaningfully careful policy, and we're not suggesting otherwise. But the structural fact remains — your clients' words, captured as a transcript, can now flow into model improvement unless someone remembers to flip a setting. For a lot of therapists, that's exactly the line they didn't want to cross. The reason it feels different is that the input was a recording of the actual session in the first place.

If there's no recording, there's no transcript to retain, de-identified or otherwise.

How CoralEHR actually drafts a note

Here's the input model, described exactly as the code works.

When you want AI help on a note, CoralEHR drafts from a typed scratchpad — short cells of text you jot during or after the session — combined with structured chart fields the patient already has. The request that goes to the AI carries your typed cells. It does not carry session audio, and it does not carry a transcript. There is no audio field in the request at all.

A fictional, composite example: during a session you might jot three quick cells — "client reports sleep improved to ~6 hrs, still waking early," "practiced cognitive reframe re: work email anxiety, engaged well," "homework: continue thought log, add one paced-breathing rep daily." CoralEHR turns those into a structured SOAP or DAP draft. You typed every word the AI saw. (This is an illustration, not a real patient.)

Three things follow from that design, and all three are enforced in the product, not just promised in marketing:

• The draft is preliminary until you sign it. Every AI-drafted note is saved with a preliminary status and stays there until a licensed clinician reviews and signs it. There is no auto-sign and no auto-accept anywhere. The AI drafts; you decide and you sign. (More on that in How CoralEHR's AI Writes Notes You Actually Sign.)
• The AI is constrained to your input. Hard guardrails on every AI surface forbid it from diagnosing, recommending treatments or medications, or predicting outcomes, and from adding facts that aren't in what you gave it. It reorganizes your observations into a note — it doesn't invent clinical content.
• Your data trains nothing. Inference runs on Anthropic's Claude API under a signed Business Associate Agreement, and under those terms patient data is not used to train models. There's no separate training or fine-tuning pipeline in the product — the only thing the system does with your input is draft the note you asked for.

A note on honesty: is there any recording capability at all?

We'd rather tell you exactly where things stand than sell you a slogan.

There is unfinished scaffolding in the frontend for an optional ambient-scribe path — UI plumbing that, before it could capture anything, requires explicit dual consent. We're being straight about it for two reasons. First, it is not how note help works today: the drafting feature you'd actually use is the typed-scratchpad path described above, and it needs zero recording. Second, that scribe path has no backend behind it — there is no transcription service, no HealthScribe job runner, no stored-transcript pipeline implemented in the system. It can't function end to end. If we ever finish it, it will ship as an explicit, consent-gated opt-in, and we'll write about it plainly — not slip it into a default.

So the accurate claim is the careful one: no recording is required. Not "we have never written a line of code that touches a microphone." The honest version is the more useful one anyway, because it's the one you can verify against how the product behaves.

Why this is the wedge

Therapists have largely made peace with AI doing the paperwork. The friction that's left is almost entirely about recording — the room, the consent, the stored audio. Remove the recording and you remove the friction, without giving up the thing clinicians actually wanted: faster notes that they still author and sign.

That's the whole positioning. Not cheapest. Not "most AI." Just: you don't have to put a microphone in the room to get help with your notes, your clients' words don't train anything, and you sign every note yourself.

If you want the fuller picture, see Ethical AI for therapists for the "two kinds of AI" frame, and AI trust & data for exactly where your clients' data goes. CoralEHR is HIPAA-compliant and signs BAAs; we're pursuing SOC 2 and HITRUST as the practice grows.

You can try it on a 30-day free trial with no credit card.

Related reading: Therapists don't hate AI — they hate AI that pretends to be a therapist · De-identified is a door, not a wall · The No AI Therapist Pledge.

This article is for general informational purposes only and is not legal, clinical, or compliance advice. Documentation, consent, and record-retention requirements vary by jurisdiction, license, and payer — consult your own legal counsel and professional guidelines before making decisions for your practice.