Private AI for Therapists: How to Vet Vendors on Encryption, PHI, and HIPAA Before You Sign

Quick Answer

Before adopting any AI tool in a behavioral health practice, require a signed Business Associate Agreement (BAA), confirm that Protected Health Information (PHI) is never sent to third-party AI models without a BAA in place, verify encryption at rest and in transit, and ask exactly how long data is retained and under what conditions humans can access it. "Privacy-first" is a marketing claim — a BAA, a documented data flow, and a clear breach-notification policy are facts.

Why Practice Owners Need a Higher Bar Than Other AI Buyers

Most AI vendor due-diligence guides are written for general business buyers. Behavioral health practices face a different set of obligations.

You are a HIPAA Covered Entity. That status is not optional, and it applies to every technology vendor who touches patient data on your behalf. A standard SaaS privacy policy is not a substitute for a BAA. A tool that is excellent for a law firm or marketing agency may expose your practice to regulatory penalties, civil liability, and reputational harm if it lacks proper PHI safeguards.

The Office for Civil Rights (OCR) has issued settlements and corrective action plans specifically tied to insecure third-party software use. The average HIPAA settlement in recent years has exceeded $1 million. Small practices are not exempt.

This guide gives you a structured checklist — built around HIPAA requirements and the realities of AI product architecture — to evaluate any AI tool before you sign a contract or enter a single note.

This post is part of our complete guide to HIPAA-Compliant AI for Behavioral Health Practices.

What "Private AI" Actually Means for a Therapy Practice

The phrase "private AI" is used to market everything from fully on-premise models to standard cloud tools with a checkbox privacy policy. As a practice owner, you need a working definition that ties to your legal obligations.

For behavioral health purposes, private AI means:

• No PHI leaves your control without a signed BAA covering the receiving party.
• PHI is not used to train third-party models, ever, unless the patient has explicitly consented in writing.
• You can demonstrate to an auditor exactly where data goes, how long it stays, and who can access it.
• You receive timely breach notification — within the 60-day window required by HITECH — if a vendor incident involves your patients' data.

Encryption is one piece of this picture, not the whole picture. A tool can encrypt data at rest and in transit and still be non-compliant if it lacks a BAA, retains PHI indefinitely, or routes session data through unprotected analytics pipelines.

The HIPAA AI Vendor Checklist for Practice Owners

Work through each section when evaluating any AI tool that may touch clinical content, intake data, scheduling records, or any information that could identify a patient.

1. Business Associate Agreement (BAA)

• Does the vendor offer a signed BAA before you input any PHI?
• Does the BAA specifically cover the AI functionality you intend to use (not just cloud storage or billing)?
• Is the BAA available in the standard contract or does it require enterprise negotiation?
• Does the vendor list its own subprocessors and confirm subprocessors are also under BAA?

A vendor who hesitates on a BAA, or who says their tool is "HIPAA compliant" without offering a BAA, should be disqualified from consideration for any PHI-touching workflow. For a worked example of how the BAA situation plays out for ChatGPT specifically, see Is ChatGPT HIPAA Compliant? 2026 Guide.

2. PHI Handling and Data Flow

• Can the vendor provide a written data flow diagram showing exactly where PHI travels?
• Is PHI sent to any third-party AI model (e.g., OpenAI, Anthropic, Google) for inference or processing?
• If third-party AI models are used, does the vendor have a BAA with those providers, and have they confirmed those providers will not use your data for training?
• Are there product tiers where PHI handling differs? (Free or trial tiers commonly have weaker protections.)

3. Encryption Standards

• Is data encrypted in transit using TLS 1.2 or higher?
• Is data encrypted at rest? What encryption standard (AES-256 is the current baseline)?
• Are database backups encrypted under the same standard?
• Who controls the encryption keys — the vendor or your practice?

4. Retention, Deletion, and Minimum Necessary

• What is the default data retention period for session data, notes, and uploaded files?
• Can you configure retention to match your practice's policies?
• Is there a self-service deletion path, or does deletion require a support ticket?
• How long does full deletion propagate, including from backups?
• Does the tool apply a minimum-necessary principle — or does it log everything by default?

5. Access Controls and Audit Logs

• Is role-based access control (RBAC) available so staff only see data relevant to their role?
• Are access logs available for your review?
• Can you see a history of who accessed patient records within the tool?
• What access does vendor staff have to your data, and under what circumstances?

6. Breach Notification

• What is the vendor's documented incident response procedure?
• Will you receive notification of a potential breach within 60 days as required by HITECH?
• Has the vendor experienced any reportable breaches? What was the outcome?
• Does the BAA specify the vendor's notification obligations?

7. Training Use and Human Review

• Is your data ever used to train or fine-tune AI models?
• Does the vendor conduct human review of session content for quality assurance or safety purposes?
• If human review occurs, is it covered under the BAA and subject to the same access controls?
• Can you opt out of training or analytics data collection while retaining full product functionality?

Red Flags in AI Privacy Marketing

The following statements appear frequently in AI vendor marketing. None of them, on their own, constitutes HIPAA compliance.

• "We take privacy seriously." This is a values statement, not a policy.
• "Your data is encrypted." Encryption is necessary but not sufficient.
• "We don't sell your data." Sale is not the primary HIPAA concern — training use and third-party sharing are.
• "HIPAA compliant." Compliance is a process and a set of controls, not a certification. Ask what specifically they mean.
• "We use enterprise-grade security." This is undefined.
• "Your data stays private." Ask for the data flow and retention policy in writing.

Any vendor that responds to your specific technical questions with marketing language rather than documented policies is a risk.

Vendor Evaluation Matrix

Use this matrix when comparing tools side by side. Score each category independently before reaching an overall decision.

How Architecture Affects PHI Risk

The way an AI tool is built determines where your PHI exposure actually lives. Two tools can make similar privacy claims while having fundamentally different risk profiles.

Client-side vs. server-side processing

Tools that process data on your device or within your own cloud environment do not route PHI to the vendor's servers. Tools that use server-side inference send data to the vendor (and often to the underlying AI model provider) on every interaction.

Pass-through AI vs. in-house models

Some behavioral health AI tools act as a thin wrapper over a general-purpose model like GPT or Claude. Your clinical content is sent to that third-party provider on every request. Others run inference on their own infrastructure under a BAA. These are fundamentally different architectures from a HIPAA standpoint. For a comprehensive review of AI stack choices, see HIPAA-Safe AI Stack for Behavioral Health.

Memory and context persistence

AI tools with persistent memory or long-term context features store conversation content — often indefinitely — to improve responses over time. For behavioral health use, this creates a PHI retention risk that standard consumer AI tools are not designed to address. Ask specifically whether session content is stored after the session ends and for how long.

PsyFiGPT was built to address exactly this architecture problem. It provides AI-powered clinical documentation — including session notes, treatment summaries, and correspondence — with PHI de-identified before any third-party AI models. The architecture is designed from the ground up for practices that need to maintain HIPAA compliance without abandoning the efficiency gains AI offers.

AI Tools Across the Practice: Where PHI Exposure Lives

Clinical documentation is the most obvious PHI risk, but practice owners should map every workflow where AI is being used or considered.

Intake and scheduling

Intake forms and scheduling workflows collect sensitive information before the first session: presenting concerns, insurance data, medication history, and referral reasons. AI tools that automate intake processing must be evaluated with the same rigor as clinical documentation tools.

PsyFi Assist handles AI-powered intake and scheduling, including therapist matching, in a HIPAA-conscious framework. It gives practices the efficiency of automated intake without routing patient-submitted data through unprotected consumer AI pipelines.

Clinical reporting and analytics

Progress notes, treatment outcome measurements, and aggregate reporting all touch PHI. AI tools that generate clinical reports or analyze practice-level outcome data require their own BAA review and data flow assessment.

PsyFi Reports (psychological evaluation and assessment reports) provides clinical report generation and analytics built for behavioral health practices. It allows practice owners to derive operational insights without manual export to general-purpose spreadsheet or BI tools that lack PHI safeguards.

Conducting Your Own Risk Assessment

HIPAA requires Covered Entities to conduct periodic risk assessments. Adding a new AI tool is a triggering event for a risk assessment update. A minimal AI-specific risk assessment should include:

1. Asset identification. List every patient data type the tool will touch.
2. Threat modeling. Identify ways PHI could be exposed through this tool (data breach, unauthorized access, training ingestion, improper disposal).
3. Current controls. Document what the vendor provides (BAA, encryption, access controls).
4. Residual risk. Identify gaps between your required safeguards and the vendor's actual controls.
5. Mitigation plan. Document steps to close gaps before deployment.
6. Documentation. Keep records. OCR requests documentation during audits and investigations.

You do not need to be a security engineer to complete this process. You do need written answers from your vendor for each item above.

Next Steps for Your Practice

Evaluating AI vendors for a behavioral health practice requires asking questions that most general-purpose software buyers never think to ask. The checklist above gives you the foundation.

Start with the BAA. If a vendor cannot provide one, the conversation ends there. Then work through data flow, retention, encryption, access controls, and breach notification. Put every material answer in writing before you sign.

If you are looking for a place to start with purpose-built tools that were designed for this compliance environment from day one, explore what PsyFi Technologies has built for behavioral health practices:

• PsyFiGPT for HIPAA-conscious clinical documentation — AI-assisted notes, summaries, and correspondence without PHI passing through third-party AI models.
• PsyFi Assist for AI-powered intake, scheduling, and therapist matching with patient data handled appropriately.
• PsyFi Reports for clinical report generation and practice analytics without manual export to unsecured tools.

The goal is the same as it has always been in behavioral health: do right by your patients while running a sustainable practice. The right AI tools make both easier, not harder.