Why Rivet doesn't record sessions or use an AI scribe

The AI-scribe category exists. Mentalyc, Heidi, Freed, and a half-dozen competitors record therapy sessions, transcribe them, and generate SOAP/DAP drafts. Most telementalhealth platforms either ship recording themselves or integrate with one of those vendors. Rivet doesn’t. Not as a “we’ll get to it” — as a design choice.

​

The short version

A 50-minute therapy session is among the most sensitive PHI surfaces in healthcare. Recording it and shipping the audio (or transcript) to a cloud LLM means: now there’s a copy of every word your client said, somewhere outside your control, processed by a system you don’t operate. Even with the strictest architecture — local recording, 24-hour deletion, vendor BAA — you’ve just moved a sensitive surface from “in the room” to “in a vendor’s processing pipeline.” The structured-template approach captures the same clinical data without the recording. PHQ-9, GAD-7, K10, DASS-21, PCL-5, target identification, SUDS, thought records, safety plans — these are the artifacts an AI scribe would try to extract from a transcript. Rivet captures them as structured data, in-session, in the templates the client filled out. The note can then auto-fill from that structured data deterministically, with no LLM, no transcript, and no recording. You write the formulation. Rivet provides the supporting structure.

​

The full reasoning

​

Privacy posture stays uncompromised

The SOAP / DAP / treatment-plan path requires zero new persistence. The templates render in your browser only, are never sent to the client’s device, are never cached anywhere durable, and exit via clipboard or print-to-PDF. The audit trail is metadata only — that a note happened, not what was in it. Same posture as the practitioner-private annotations Rivet already supported. A recording pipeline — any recording pipeline — adds a persistent PHI store that Rivet would have to operate, secure, and disclose to clients. The structured-template path avoids that entirely.

​

Clinical quality

A SOAP or DAP note authored by you — with the structured template data sitting alongside in the same session summary — is clinically superior to an LLM reconstruction from a transcript. Your note carries inference and judgment that the transcript doesn’t contain: formulation, treatment direction, risk assessment. AI session-notes’ marketing pitch is “save typing time,” not “produce better notes.” The auto-fill engine cuts the typing without putting the inference in a black box.

​

The structured data is already captured

The clinical template library plus the EMDR 2.0 workspace already capture, in structured form, the data an AI session note would have to extract from a transcript. PHQ-9 scores. GAD-7 scores. K10 / DASS-21 / PCL-5 with cluster sums. EMDR target ID (NC / PC / VOC / SUDS baseline / body sensation). SUDS trajectory across BLS sets. Thought records. Safety plans. Sleep diaries. An LLM on a transcript is doing information extraction from unstructured conversation — an inference task with measurable error rates and known hallucination modes. An auto-fill engine on the templates is doing information reformatting from already-extracted clinical data — a deterministic merge. The first needs an LLM and a disclosure modal. The second needs neither.

​

No hallucinated content reaches the chart

An LLM-summarized note will, eventually, write something the client didn’t say. Or attribute a symptom to the wrong session. Or get the medication wrong. The category’s marketing materials acknowledge this and recommend “always review the generated note before signing.” Practitioners do — but the burden of catching every mistake is real, and the rare-but-serious error rate is non-zero. The deterministic auto-fill engine doesn’t have a hallucination mode. What lands in your draft is exactly what was in the source template, formatted for clinical reading. The only inference in the note is yours.

​

No cloud AI processes PHI

Rivet makes zero LLM calls during your session — no cloud transcription, no vendor BAA to negotiate. The auto-fill engine runs in your browser. The voice-to-text dictation runs on the browser’s built-in dictation (which may use a vendor cloud — same as iOS keyboard dictation — but never receives the session audio; only the words you dictate into the textarea). This isn’t an absolute claim that no cloud AI is involved in any tool you use during your day. It’s a claim about Rivet specifically: no session content, transcribed or summarized, leaves your browser via Rivet.

​

Optionality is preserved

If real-practitioner use of the SOAP / DAP path shows that AI note generation is the missing piece, the path to build it is sketched — de-identified transcript, sent to a Canadian-bound vendor under a data-processing agreement, re-substitution on your device, practitioner review. That said: the signal would have to come from real use. AI-scribe marketing pressure isn’t signal.

​

What this means practically

• No recording. The video session is peer-to-peer. The audio
  • video never reach a Rivet server in a re-playable form. Rivet cannot retroactively transcribe a session.
• No AI scribe. No vendor integration, no “connect your Mentalyc account” button.
• No prompt engineering for clinical content. The structured templates don’t have LLM-generated text in them. The cognitive-distortions list, the safety-plan steps, the PHQ-9 items — all human-authored, citation-backed.
• Your dictation is yours. Voice-to-text uses the browser’s built- in dictation. Rivet never sees the audio. (Your browser might — Chrome routes dictation through Google’s cloud, Safari handles it on-device. Same posture as your iOS keyboard.)
• The chart entry is yours. When you paste the note into Jane or Owl, every clinical interpretation in it was written by you. No “AI-assisted” disclosure to your client, no provenance ambiguity.

​

What if I want an AI scribe anyway?

Use a separate tool alongside Rivet — there are several in the category, and they’re designed to plug into any video platform. Rivet won’t block you. But Rivet won’t help, either: the audio never reaches us, so we can’t pipe it to a vendor on your behalf. The third-party tool would handle the recording, transcription, and AI-summarization end of things; Rivet would still handle the session, the templates, and the structured data. If demand becomes clear, we’ll revisit. For now the answer is: the structured-template path is what Rivet bets on, and it earns the bet every time a note exports cleanly with no hallucinated content.

​

Related articles