AI tools for therapy documentation are moving fast. Every few months, a new product promises to eliminate note-writing after sessions. Some of them are genuinely useful. Some are compliance risks. Many therapists aren’t sure which is which.
The short answer: AI tools for therapy notes can meaningfully reduce documentation time and help with structure, but they require a signed BAA for HIPAA compliance, transparent client consent, and clinician review of every output. “AI-assisted” does not mean “AI-handled” — the clinical judgment in your notes is still yours.
What AI therapy note tools actually do
Direct answer: There are three categories of AI tools marketed for therapy documentation. Understanding which category a tool belongs to tells you what it can and can’t do.
Category 1: Ambient AI scribes
These tools listen to the therapy session in real time — via a phone, tablet, or computer microphone — and generate a draft note from the audio. Examples: Nabla, Freed, Upheal, Blueprint.
What they produce: a structured session note draft, often in DAP, SOAP, or narrative format, based on what was said in the session.
What they require: audio recording of the session, which raises its own consent and privacy questions (covered below).
Category 2: AI-assisted documentation platforms
These tools don’t record sessions. Instead, you input key details after the session — what was discussed, client status, interventions used — and the AI drafts a complete, formatted note. Examples: TheraMemory and similar documentation platforms.
What they produce: a structured note based on clinician input, in the format you specify.
What they require: the clinician to capture the key clinical elements. The AI structures and expands, but doesn’t replace the clinician’s observation.
Category 3: General AI tools used for documentation
Some therapists use ChatGPT, Claude, or similar tools to draft or improve notes. This is the highest-risk category from a compliance standpoint — most of these tools do not offer HIPAA-compliant configurations for standard accounts.
What none of them do
No current AI tool reliably replaces clinical judgment in therapy documentation. AI can structure what you give it. It cannot:
- Assess suicide risk
- Interpret transference or relational dynamics
- Determine what’s clinically significant versus incidental
- Know that the client’s flat affect contradicted what they said
AI draft notes require clinician review. Every time.
HIPAA compliance: the non-negotiable checklist
Direct answer: Any AI tool that processes Protected Health Information (PHI) — including session notes, audio recordings, or any identifiable client data — must be covered by a signed Business Associate Agreement (BAA). No BAA = HIPAA violation, regardless of how secure the platform claims to be.
The BAA question
A Business Associate Agreement is a contract between you (a HIPAA covered entity) and the technology vendor (a business associate) in which the vendor agrees to protect PHI. Without it, using the tool for any client data is a violation.
Before using any AI documentation tool, ask:
| Question | What a compliant answer looks like |
|---|---|
| Do you sign a Business Associate Agreement? | Yes — and they provide it in writing before you use the product |
| Is client data used to train your AI models? | No — PHI must be excluded from model training |
| Where is data stored? | US-based servers, or jurisdiction you’ve verified |
| How long is data retained? | Defined retention period, not indefinite |
| What happens if there’s a breach? | Written breach notification policy |
Which tools are HIPAA-compliant (and which aren’t)
| Tool type | HIPAA status | Notes |
|---|---|---|
| EHR with built-in AI notes | ✅ Compliant | SimplePractice, TherapyNotes — BAA included |
| Purpose-built AI scribe with BAA | ✅ Compliant if BAA signed | Nabla, Freed, Blueprint — confirm BAA before using |
| TheraMemory and similar platforms | ✅ Compliant | HIPAA-ready, PHI not used for training |
| ChatGPT (free or Plus) | ❌ Not compliant | No BAA; not available for consumer accounts |
| Claude (consumer account) | ❌ Not compliant | No BAA on personal accounts |
| Google Gemini (personal account) | ❌ Not compliant | No BAA; consumer product |
| Otter.ai (standard) | ❌ Not compliant | No BAA on standard plans |
The rule: if you can’t find the BAA within two minutes of looking at the vendor’s website, assume they don’t offer one.
For reference: TheraMemory is HIPAA-ready, does not use client data for model training, and stores session notes exclusively within your account. This is the baseline any compliant tool should meet — use it as a benchmark when evaluating alternatives.
Client consent for AI documentation
Direct answer: Ethically and practically, yes — clients should know if AI tools process their session information. Update your informed consent form before using any AI documentation tool.
What your informed consent should cover
Your consent form doesn’t need to be technical, but it should answer three questions:
- What AI tool is used and what it does (ambient recording vs. note drafting from clinician input)
- What data is processed and how it’s protected
- Whether the client can opt out and what opting out means for their care
Sample consent language:
“I use AI-assisted documentation tools to help structure my session notes. [If using ambient AI: ‘These tools may involve audio processing of our sessions.’] No identifiable information is sold or shared with third parties. This service is HIPAA-compliant and covered by a Business Associate Agreement. If you have questions about this practice or would prefer I document differently, please let me know.”
The opt-out question
Some clients will object to AI-processed notes — for privacy reasons, concerns about data security, or personal values. Have a plan for what “opting out” looks like before a client asks. Usually this means documenting manually for that client. That’s a reasonable accommodation and worth clarifying in your consent process.
What AI does well in therapy documentation
Direct answer: AI is most useful for structure, consistency, and reducing the time burden of post-session documentation. It is least useful for anything requiring clinical interpretation.
Where AI adds real value
Reducing documentation time. For many therapists, note-writing after sessions takes 15–30 minutes per client. AI-assisted documentation typically cuts this to 5–10 minutes — not because AI writes better notes, but because the AI handles the structural scaffolding.
Consistency. AI-generated drafts follow a format every time. Human notes drift — especially at the end of a long day, when it’s tempting to write “continued work on anxiety” and close the file.
Completeness prompts. Good AI documentation tools prompt you to include elements you might skip: risk assessment, homework from last session, plan for next session. The prompt alone improves documentation quality.
Session pattern recognition. Some tools aggregate across notes to show patterns — topics that recur, mood trajectories, homework completion rates. This is clinically useful for treatment review and for preparing referral summaries. TheraMemory keeps the full client history — sessions, progress, homework — in one place, so a transfer summary that would take two hours to compile from scattered notes takes twenty minutes.
Where AI falls short
Clinical judgment. An AI tool can note that “client described low mood.” It cannot determine whether that’s clinically significant, consistent with the treatment trajectory, or a signal that warrants a different intervention.
Relational and nonverbal content. What the client said and what the session felt like are different things. AI works from words. The flat affect, the long pause, the way the client’s energy shifted — that’s yours to capture.
Risk assessment. AI should never be the primary documenter of suicide risk, self-harm, or safety planning. This is the single area where AI assistance introduces the most risk. Document risk assessments yourself, in your own clinical language.
Nuanced phrasing. AI-generated notes often read like AI-generated notes — competent but generic. The clinical voice that captures your actual understanding of a client requires human editing.
Evaluating an AI tool before you commit
Direct answer: Evaluate on five dimensions before using any AI tool with client data. Get written answers, not marketing language.
Five evaluation criteria
1. Compliance
- Does the vendor sign a BAA?
- Is PHI excluded from model training?
- What’s the breach notification process?
2. Data handling
- Where is data stored?
- How long is it retained?
- Can you delete a client’s data, and how?
3. Recording vs. input
- Does the tool require audio recording, or does it work from text input?
- If recording: where is the audio stored, and for how long?
- If input: how much clinician input does it need to produce a usable draft?
4. Output quality
- Ask for a demo with a sample session scenario
- How much editing does the draft typically require?
- Does the output match your documentation format (DAP, SOAP, narrative)?
5. Workflow fit
- Does it integrate with your EHR, or is it a separate step?
- What does the end-to-end workflow look like on a typical day?
- What happens if the tool is unavailable?
Practical workflow: using AI in your documentation practice
Here’s what a sustainable AI-assisted documentation workflow looks like in practice — not the marketing version:
DURING SESSION
→ Focus on the session, not documentation
→ [If using ambient AI]: start recording with client consent confirmed
→ [If using input-based AI]: jot 3–5 key phrases or themes as quick notes
IMMEDIATELY AFTER SESSION (5–10 min)
→ Open your documentation tool
→ Input key elements: what was discussed, interventions, client status,
homework reviewed, homework assigned, risk assessment, plan
→ Review AI-generated draft
→ Edit for clinical accuracy and your voice
→ Finalize and sign
END OF DAY
→ All session notes complete — no documentation backlog
→ Flag any notes that need follow-up (risk items, consultation needed)The workflow only works if note-writing happens immediately after each session. AI doesn’t fix documentation backlog if you’re not doing the notes same-day — it just makes each individual note faster.
Frequently asked questions
Can I use ChatGPT to help write therapy notes if I remove identifying information? De-identified information is not considered PHI under HIPAA, so using a truly de-identified note with a general AI tool does not violate HIPAA. The practical challenge: genuinely de-identifying a therapy note is harder than it sounds. Removing a name isn’t enough — age, location, diagnosis, presenting problem, and specific life circumstances can all be re-identifying in combination. If you’re using general AI tools for documentation, the safer practice is to use a HIPAA-compliant tool with a BAA and avoid the de-identification problem entirely.
Should I tell clients that their notes are partly AI-generated? Ethically, yes — at the level of process (“I use AI tools to help structure my documentation”), not necessarily at the level of individual note content. Clients don’t have a right to approve each note, but they do have a right to know how their information is processed. Disclose this in informed consent.
What if an AI tool generates something clinically inaccurate in a note I sign? You are responsible for everything in your clinical record — including AI-generated content you sign. “The AI wrote it” is not a clinical or legal defense. Review every AI draft before finalizing. If you’re regularly editing heavily for accuracy, that’s a signal the tool isn’t the right fit for your practice.
Is there liability risk in using AI for therapy notes? Yes, in two ways. First, compliance risk: using a non-HIPAA-compliant tool with PHI is a violation that carries penalties regardless of whether a breach occurs. Second, documentation quality risk: AI-generated notes that are generic, inaccurate, or miss clinically significant content could be a liability if records are ever subpoenaed or reviewed by a licensing board. The answer to both is the same: use a compliant tool, and review every note before signing.
AI tool evaluation checklist
COMPLIANCE
[ ] Vendor signs a Business Associate Agreement (BAA)
[ ] PHI is not used for model training — confirmed in writing
[ ] Data storage location confirmed
[ ] Breach notification policy exists and is written
CONSENT
[ ] Informed consent updated to disclose AI documentation use
[ ] Opt-out process defined for clients who object
[ ] [If ambient recording]: session recording consent obtained from each client
WORKFLOW
[ ] Tool integrates with (or doesn't add friction to) existing workflow
[ ] Output quality tested with sample session scenarios
[ ] Editing time per note estimated and acceptable
[ ] Backup process defined if tool is unavailable
CLINICAL QUALITY
[ ] Risk assessment documented by clinician, not delegated to AI
[ ] AI drafts reviewed before signing — every time
[ ] Notes reflect clinical voice, not generic AI outputAI documentation tools are a legitimate productivity tool for therapists when used correctly. The ones that work are HIPAA-compliant, transparent about data handling, and honest about what AI can and can’t do. The ones that create risk cut corners on compliance, promise too much, and leave clinical judgment to the software. Know the difference before you sign up.
Try TheraMemory free for 14 days
AI-assisted therapy documentation that's HIPAA-ready — structured note drafts from your input, no session recording required. Built for therapists who want faster documentation without compliance risk.
Document every session in under 10 minutes.
Try TheraMemory