How AI clinical notes actually work (and how Signal keeps you in control)
May 13, 2026 · 10 min read · By Signal by TheraPreneur
If you are a licensed clinician and you are skeptical of AI documentation, you are reading the right article. We wrote this for you, not for the early adopter who already trusts the magic.
Below is what is actually happening inside Signal when an AI note appears in your draft queue — the pipeline, the constraints we put on the model, and the seven or eight places we deliberately stopped the system from doing more than it should.
1. The skepticism is warranted
A lot of what gets marketed as "AI therapy notes" right now is a thin wrapper around a general-purpose LLM. You paste in a transcript, the model produces a SOAP note, and you sign it. That workflow has three problems any clinician will spot inside a week:
- It hallucinates clinical detail. General LLMs are trained to sound fluent. When the source material is thin — a quiet client, a fragmented exchange — they fill the gap with plausible content that was never said. In a clinical note, plausible-but-fabricated is malpractice waiting for an audit.
- It flattens modality. A CBT session and an IFS session are different documents. They use different vocabulary, attend to different mechanisms, and a generic prompt collapses them into the same beige paragraph. You can feel it when you read the output. Reviewers will too.
- It treats notes as content, not records. A clinical note is a legal record with retention requirements, audit expectations, and a signature gate. "ChatGPT for therapy notes" produces text. It does not produce a record.
So when a colleague tells you they tried "ChatGPT for notes" and gave up after a week, they are not being a Luddite. They identified a category mistake. Clinical documentation is not a writing task with a clinical setting. It is a clinical task that produces writing. The two require different systems.
2. The pipeline
Here is what happens between the moment you press record and the moment a draft lands in your queue. Six stages, each with a specific job and a specific failure mode you can inspect.
#1Voice capture
Audio is captured from the in-room microphone or telehealth stream at 16 kHz mono PCM. Capture starts only after explicit in-session consent is logged. The raw audio buffer lives in memory long enough to transcribe and is then dropped — it is never written to long-term storage.
#2Speech-to-text (Deepgram)
Audio is streamed to Deepgram's Nova model. We use streaming rather than batch so words land in the transcript with sub-second latency, which matters because the next stage needs word-level timestamps. The resulting transcript is a list of word objects with start time, end time, confidence, and a placeholder speaker label.
#3Diarization (Pyannote)
A separate Pyannote pipeline runs over the same audio to answer one question: who spoke when. Its segment boundaries are aligned back onto the Deepgram word stream so each word ends up tagged with a speaker ID. Two-speaker sessions resolve cleanly; family or group sessions surface a confidence drop you can review before signing.
#4Emotion and prosody analysis
Our clinical engine analyzes the speaker-tagged transcript and the audio envelope to produce per-utterance markers: affect valence, arousal, hesitation, prosodic flatness, and a small set of clinically meaningful patterns. These are signals, not diagnoses. They appear as evidence the note can cite, not as conclusions the note asserts.
#5Structured note draft
An LLM consumes the diarized transcript, the emotion markers, the patient's recent history, and your selected note format and modality. It produces a draft in the requested structure — never a freeform essay. Every claim it generates carries a pointer back to the transcript range that supports it.
#6Clinician review and sign-off
The draft opens in your editor with the transcript pinned beside it. You can hover any sentence to see the source utterance and timestamp. Edit anything. Reject anything. Nothing is filed, sent to insurance, or marked complete until you sign. Your signature is the gate.
3. What "grounded" means
The single most important constraint in the system is grounding. In Signal, the note can only reference content that was actually said in the session, and every reference carries a timestamp into the transcript.
Mechanically, the drafting model is given the diarized transcript with stable utterance IDs. When it writes a sentence, it must emit the IDs of the utterances that support that sentence. A post-processing step reads those IDs back, confirms each one resolves to a real range in the transcript, and rejects any sentence whose support is missing or fabricated. The rejected sentences never reach you.
In practice
When you open a draft, every sentence is hoverable. Hover reveals the utterance and the timestamp the model used to produce it. If the source does not justify the claim, you delete or rewrite the sentence in place. If you find yourself doing that frequently, that is feedback worth telling us about — and feedback the system uses to recalibrate for your style.
4. Note formats we support
Format is a per-clinician and per-payer choice. Signal does not pick one for you. Here is the menu and a short, honest take on when each one earns its keep.
SOAP
Subjective, Objective, Assessment, Plan
The default for most insurance-billable visits. Familiar to every payer and audit reviewer.
DAP
Data, Assessment, Plan
Behavioral health agencies that prefer a leaner three-section structure. Common in community mental health.
BIRP
Behavior, Intervention, Response, Plan
Tracks what the clinician did and how the client responded. Strong fit for skill-building and CBT-flavored work.
GIRP
Goal, Intervention, Response, Plan
Goal-anchored variant. Useful when treatment plans drive every session and you want goal progress visible at the top.
PIRP
Problem, Intervention, Response, Plan
Problem-list anchored. Common in agencies billing Medicaid where the presenting problem must lead each note.
SIRP
Situation, Intervention, Response, Plan
Crisis and case-management contexts where the precipitating situation matters more than a stable problem list.
PIE
Problem, Intervention, Evaluation
Compact format for brief contacts and care coordination notes.
Custom
Your own template
You define the sections and prompts. Useful for supervisors, group practices with house style, or training programs.
5. The 16 modalities
A SOAP note for a CBT session and a SOAP note for an IFS session should not read the same. The mechanism of change is different, the language is different, and what counts as an "intervention" is different. Signal ships with sixteen modality profiles. Each profile shapes vocabulary, the kinds of mechanisms the model attends to, and what gets surfaced as a treatment-relevant moment.
Concretely: in CBT mode, the model will look for cognitive distortions named in session, behavioral experiments assigned, and homework reviewed. In IFS mode, it will track parts language, blending and unblending, and Self-led moments. In Gottman work, it will tag the four horsemen, repair attempts, and bids. The transcript is the same; the lens is different.
6. The clinician sign-off is the gate
Every AI-generated note in Signal is a draft until you sign it. The system never auto-files, never sends to a payer, and never marks an encounter complete on its own. There is no "auto-approve after 24 hours" setting. There never will be.
Pinned transcript
The source transcript stays open beside the draft so review is one glance, not a context switch.
Change tracking
Every edit you make is captured. Over time this teaches the system your voice without ever leaking to other tenants.
Permanent audit log
Who drafted, who edited, what changed, when signed. Append-only and exportable for licensing boards or audits.
No silent autoship
Insurance submission, supervisor routing, and chart filing all require an explicit signed note. The AI cannot bypass them.
7. What gets stored, and what does not
We are deliberate here because storage is where the AI documentation industry has its weakest record. Signal's defaults:
- Raw audio: not retained. Audio exists in memory long enough to transcribe. Once Deepgram returns the transcript, the buffer is dropped. We do not write WAV or MP3 to long-term storage. Ever.
- Transcripts: stored only with explicit consent. Default is consent-on at session start, with a visible toggle. If consent is off, the transcript is used to generate the draft and then deleted in the same request cycle.
- Embeddings: stored for semantic search. The structured note and its embeddings are stored so you can search across a chart for, say, every mention of suicidal ideation. Embeddings are tenant-isolated and never trained against.
- Cross-tenant training: never. No client of yours, and no client of any other practice, ever ends up in a training set — ours, Deepgram's, or any LLM vendor's. This is contractual on every link in the chain.
8. Insurance language without the padding
Most clinicians have learned to write notes that satisfy two readers: themselves and a payer. The payer reader needs medical-necessity language — symptoms tied to functional impairment, intervention tied to a treatment plan, response that justifies continued care.
Signal's drafting prompt is tuned to surface medical-necessity language when the session content supports it, and to stay quiet when it does not. The model will not invent functional impairment that was not discussed. It will, however, take a moment the client described — say, missing two days of work because they could not get out of bed — and render it in language an insurance reviewer recognizes as billable.
The line we hold
Translate, do not embellish. If the session does not justify a CPT code, the note will not pretend it does. Padding to clear an audit threshold is exactly the failure mode we will not ship.
9. PHQ-9 and GAD-7 integration
Signal sends standardized outcome measures to the client through the patient portal — PHQ-9, GAD-7, and any custom instruments you configure. Scores feed back into the chart and become available to the drafting model.
Practically, that means a note about a depression-focused session can pull the client's most recent PHQ-9 and last several scores into the Assessment section, and trend language follows automatically. If the score is rising, the draft notes it; if a critical-item flag (such as item 9) trips, the draft surfaces it for explicit review before you sign. Every integration is one direction: scores inform the draft, the draft never alters scores.
10. Common worries, with honest answers
We hear the same five concerns from skeptical clinicians. None of them have glib answers. Here is how we actually think about each one.
What about the therapeutic alliance — won't the mic get in the way?
It can, especially in the first session. We recommend a brief, plain-language consent script and giving the client visible control over a pause button. Most clients adapt within one or two sessions. Some do not, and that is also a clinical signal worth honoring — Signal lets you turn capture off per-session without losing the patient record.
Hallucination — what if the AI invents something the client didn't say?
Every sentence in the draft is bound to a transcript range. If the model tries to assert something with no source, the binding step fails and that sentence is dropped before you ever see it. You can also click any line in the draft to jump to the source utterance, which makes invented content immediately visible. We do not claim this is impossible — we claim it is auditable.
What happens if Signal goes down mid-session?
Capture buffers locally in the browser for the session. If our backend is unavailable, audio queues and uploads when connection returns; if Deepgram is unavailable, we fall back to a slower in-house transcript. Worst case, you have no AI draft and you write the note the way you always have — your session is never blocked by our infrastructure.
I don't want my voice or my client's voice training someone's model.
It does not. Audio is processed and dropped. Transcripts and embeddings are tenant-isolated and never used to train base models — ours or any vendor's. Deepgram is configured to opt out of model improvement on our account. This is contractual, not aspirational.
If I sign a note I didn't fully read, am I liable for the AI's mistakes?
Yes. That is true of any signed clinical document, AI-assisted or not. Signal's job is to make the draft fast and the review meaningful — pinned transcript, source highlighting, change tracking. Your job is still to read and stand behind what you sign. We will never make that easier to skip.
See the pipeline in your own words
Walk through the full architecture on the how-it-works page, or start a 14-day trial and run a real session through it. No card required, full features, your sign-off is still the gate.