AI Notetaker for Therapists: No Recording Means No Audio File

10/19/2026

By Kyle Nelson, Founder, Fazit

ogTitle: "AI Notetaker for Therapists: No Recording Is an Architecture"

An AI notetaker for therapists that doesn't record is not a feature list item. It's an architecture. Some tools avoid recording by never capturing session audio at all. Others capture audio, transcribe it, and delete the audio afterward. For a therapist, the distinction between those two postures is the entire risk surface.

Fazit is a native macOS menubar app that captures 1:1 client calls without a bot, transcribes on the Mac, and writes Markdown notes into an Obsidian vault. The call audio exists only in a fixed-size RAM ring buffer with no write API, and it's destroyed on every exit path including errors. Every note records audio_retained: false in its frontmatter. That's the architecture, and it's the only honest answer to the search query.

What "No Recording" Actually Means in Therapy AI Tools

The therapy AI-scribe market has settled into a pattern. Most tools lead with "HIPAA-compliant," and the label does a lot of work. But HHS recognizes no such certification. The label means the vendor will sign a business associate agreement, which is paperwork that makes your client's session leaving your practice lawful, not paperwork that prevents it. I wrote about this in A HIPAA-Compliant AI Notetaker: What the Label Buys, and What It Cannot.

The tools that specifically claim "no recording" fall into two camps. The first camp, represented by Quill, doesn't capture session audio at all. The therapist provides a short post-session summary, and the system turns that into a formatted progress note or treatment plan. According to a Twofold blog post, Quill stores nothing after that transformation. That's a real no-recording posture, but it's also not a notetaker in the capture sense. You're still writing the summary yourself.

The second camp captures audio, transcribes it, and then deletes the audio. The University of the Incarnate Word counseling consent form describes this pattern with Scope Health: audio is automatically deleted after a transcript is created, typically within minutes. The transcript is retained for the shorter of seven days or until the therapist signs and locks the progress note. That's a defensible workflow, but it still creates two artifacts that a therapist now has to control: the audio file, however briefly, and the transcript.

The Audio Path Decides the Risk

For a therapist, the question isn't whether the vendor promises to delete the audio. The question is whether the tool creates an audio file at all. A file that exists for five minutes is still a file. It can be intercepted, subpoenaed, or mishandled in that window. The Granola Record Audio post walks through what happens when a tool records from your microphone and system audio, sends it to a transcription provider, and deletes it after. The question is what happens in between.

Fazit's pipeline is different. The microphone and the call app's audio are captured as two separate streams into a RAM ring buffer. Parakeet via CoreML transcribes both streams on the Mac. A local model on localhost generates the note. Then the buffer is reset. There is deliberately no write(to:) method on the buffer. The audio never becomes a file, so there's nothing to delete, nothing to leak, and nothing to subpoena. The never-records architecture post explains why this is an invariant, not a marketing claim.

This matters for therapists in a specific way. A therapy session contains the most sensitive material a professional can handle. The client's trauma history, family dynamics, substance use, suicidal ideation. A cloud notetaker that records the session and sends it to a vendor has created a discoverable record of that material. The AI Notetaker Legal Risk post makes the point directly: the recording you never made can't be subpoenaed.

Consent Is Still Your Duty

None of this removes the therapist's own consent obligation. California's CIPA is all-party. Germany's §201 StGB is criminal. A RAM-only pipeline that transcribes in volatile memory and never writes an audio file generally does not meet the elements of §201 StGB, according to German counsel cited in the Fazit research document, because there is no "recording on a sound carrier." But that's a defense against a specific charge, not a substitute for telling the client what you're doing.

The University of the Incarnate Word consent form is a useful model. It tells clients that AI tools may be used under the therapist's direct supervision, gives them the option to decline, and states that declining will not affect the quality of care. A SimplePractice guide recommends obtaining written, revocable consent before using any recording or transcription tool. The Is It Legal to Record Client Calls post maps the state-by-state consent landscape and notes that the rules are stricter for therapists than for most professionals.

This is not legal advice. Consent and professional-conduct duties apply to the person running the call, and a therapist's licensing board may have its own rules about AI documentation tools.

What a Therapist's Working Day Actually Looks Like

A therapist running six sessions back to back doesn't have time to write a summary after each one. That's the appeal of an AI notetaker. The problem is that the tools that save the most time are the ones that capture the most audio.

Fazit's workflow is built for that day. You start the capture from the menubar before the session, talk normally, and click stop when the session ends. A Markdown note appears in your Obsidian vault with a summary, key moments, action items with owners, and a paste-ready follow-up email. The note is a plain file in your vault, not a row in a vendor's database. The Obsidian Meeting Notes Should Be Files post explains why that ownership matters.

The transcript is attributed You/Them because the two audio streams are physically separate. That's a small detail that matters for a therapist reviewing a session. You can see exactly what you said and what the client said, without diarization guesswork.

The Comparison That Matters

Tool	Captures audio?	Creates an audio file?	Where does transcription run?	Note lives
Quill	No	No	Not stated in source	Not stated in source
Scope Health (per UIW form)	Yes	Yes, deleted after transcript	Not stated in source	Not stated in source
Granola	Yes	Yes, deleted after transcription	Cloud	Vendor cloud
Fazit	Yes	No	On the Mac	Your Obsidian vault

The table reorders around one question: does an audio file ever exist? Two of these tools create one, however briefly. Fazit is the row where the file was never eligible to exist.

The HIPAA Label Doesn't Answer the Question

Every therapy AI scribe leads with HIPAA-compliant. The label means the vendor will sign a business associate agreement. It does not mean the vendor never receives your client's session audio. A behavioral-health AI-scribe guide from OmniMD says a compliant tool should use a signed BAA, encryption, access controls, audit logs, and data minimization. That's a list of controls around data that has already left your office.

Fazit makes no HIPAA certification claim. It's an early-stage product with no SOC 2 or HIPAA certification. What it offers is a different posture: no third party ever receives the call content, so no business associate relationship exists for the audio path. The Business Associate Agreements for AI Scribes post covers what the paperwork actually governs.

A therapist who wants the time savings of an AI notetaker without the recording risk has two honest options. Use a tool like Quill that never captures audio and accept that you're still writing the summary. Or use a tool like Fazit that captures audio but never lets it become a file. The middle path, capture and delete later, is the one that creates the most paperwork for the least protection.