Choosing A Vendor: 20 Questions You Must Ask

For most therapists, the administrative burden of clinical documentation is a direct path to burnout. AI therapy notes promise a way out: software that listens, transcribes, and drafts your notes instantly, giving you back your evenings and your focus. It is an enticing offer.

But here is the risk: a bad vendor is worse than no vendor at all. Choosing the wrong AI platform exposes you to data breaches, HIPAA violations, and financial penalties, not to mention the loss of patient trust. To protect your practice, you must look beyond the sales pitch. Here are the 20 non‑negotiable questions every therapist needs to ask before signing a contract.

Security & Compliance: The Non-Negotiables

Before discussing features or pricing, you must verify that the vendor can legally and technically handle mental health data. Under HIPAA, you are the Covered Entity; the scribe vendor is your Business Associate.

1. Will You Sign a Business Associate Agreement (BAA) Before We Start?

A Business Associate Agreement (BAA) is a legally binding contract that extends HIPAA liability to the vendor.

• Context: Without a signed BAA, the vendor has no legal obligation to protect your patients' data. If they suffer a breach, the penalties fall entirely on you for disclosing PHI to a non-compliant third party.

2. Follow-up Question

• "Do you sign the standard HHS-compliant BAA without modifications, or does your agreement contain caveats that limit your liability?" Be wary of vendors who cap liability at the cost of your subscription.

3. Where Is My Data Stored? (Data Residency)

Data lives on physical servers, and those locations determine which laws apply.

• Context: If you practice in the US, your data must remain on US soil to stay under HIPAA jurisdiction. Data stored abroad may be subject to foreign privacy laws or seizure by foreign governments.

4. Follow-Up Question

"Are you using a major provider like AWS US‑East or a third‑party data center?" Vagueness here is a red flag.

5. What Happens to My Recordings?

You need to know exactly what happens to the audio file after the AI finishes its job.

• Example: Ask specifically: "Are audio files deleted immediately after transcription, or are they stored? If stored, even temporarily, can you confirm they are encrypted at rest using AES-256 and encrypted in transit with TLS 1.3?"

6. Is the Platform "Zero-Retention" by Design?

Beyond storage, you need to know if your data is being used to train commercial AI models.

• Context: Some vendors view therapy transcripts as free training data. Using patient PHI to improve a public-facing model is a breach of confidentiality, even if they claim it is "anonymized."

Accuracy & Clinical Safety

An AI that writes fast is useless if it writes wrong. In clinical settings, inaccuracies are clinical risks. You need to understand how the model thinks, where it fails, and how it handles the complexity of therapeutic dialogue.

7. How Do You Handle "Hallucinations"?

A "hallucination" is when the model generates information that was never spoken. It doesn't just mishear; it makes things up.

• Context: Imagine an AI adding "Patient reports suicidal ideation" to a note when the topic never arose. This is a hallucination, and it creates a dangerous legal and clinical record. All LLMs hallucinate to some degree, but medical-grade tools must have safeguards.

8. Follow-Up Question

• “How exactly is your model fine-tuned specifically to prevent adding fictitious symptoms, narratives, or medical terminology that wasn't present in the session?"

9. Can It Understand Therapeutic Nuance?

Therapy is not linear. patients talk about the past, present, and future, often in the same sentence. A basic AI lacks the context to place events correctly.

• Context: A patient might say, "I felt just as scared as I did back in that car accident." A generic model might chart this as a current event ("patient reports fear related to car accident today"). A clinically aware model understands the distinction between reporting a memory and experiencing a current event.

10. Follow- Up Question

"How does your model differentiate between a patient describing a past trauma, recounting a dream, or reporting a current event? Can it handle temporal context accurately?"

11. Does the Model Learn My Specialty?

An AI trained on general conversation or even general medical notes will struggle with the language of mental health.

Technical Detail: There is a massive difference between a general large language model (like standard ChatGPT) and a fine‑tuned model trained exclusively on mental health transcripts. A fine‑tuned model understands terms like "CBT," "exposure therapy," or "transference" in context.

12. Follow-Up Question

“How does your model handle specialty modalities like DBT, EMDR, or play therapy?"

13. How Does It Handle Multiple Speakers in a Noisy Environment?

Real‑world therapy is not like a podcast studio. There are overlapping voices, crying, background noise, and sometimes poor microphone quality.

• Context: Speaker diarization is the AI's ability to label "who spoke when." High accuracy here is essential to avoid charting the therapist's questions as patient symptoms.

Workflow & EHR Integration

The goal is seamless integration.

14. Does It Integrate Natively With My EHR?

• Context: "Integration" can mean anything from a bidirectional sync to a basic copy-paste function. True integration uses HL7 or FHIR APIs to map data directly into the correct fields of your specific EHR.

15. Follow-Up Question

"Can you show me a live demo with my specific EHR?"

16. Can I Edit the Note Before It Goes Into the Chart?

Automation should not mean loss of control. You are the clinician; the AI is the assistant.

• Context: Some platforms push notes directly into the chart to save clicks. This is dangerous. You must have a chance to review, redact, and rewrite before the note becomes a permanent legal document.

17. What Note Templates Do You Support?

Different payers, supervisors, and modalities require different note structures.

• Context: You need to know if the AI can match your existing template requirements, whether that is SOAP, DAP, GIRP, or BIRP. Customization is even better, allowing you to tailor sections for specific patient populations.

Business & Contractuals

Once you are satisfied that the product works, you must treat the vendor like any other business partner. Understand the costs, the risks, and the exit strategy before you sign.

18. How Does Pricing Work? (Per Provider vs. Per Session)

Pricing models vary.

• Context: Some vendors charge a flat monthly fee per provider. Others charge per "session" or per "note," which can become expensive if you have a high-volume practice. You need to model the cost against your actual caseload.

19. Can I Export My Data If I Leave?

If you decide to leave, your clinical notes must leave with you.

• Context: Your notes belong to you and your patients, not the software company. You need to ensure you can retrieve all your data in a usable format if you cancel your subscription, whether due to cost, feature gaps, or a practice closure.

20. Follow-Up Question

"What is your data portability policy? If I cancel my subscription, can I export all my generated notes and transcripts in a standard, readable format like JSON or CSV? Is there an 'export fee'?"

Conclusion

Selecting an AI scribe is more than a software purchase; it is a partnership that impacts your clinical integrity, your patients' privacy, and your professional liability. The demo and time‑saving promises are meaningless if the vendor fails on security, accuracy, or transparency. By asking these 20 non‑negotiable questions, you move from being a passive buyer to an informed advocate for your practice.