Using an LLM with patient data is only HIPAA compliant when the LLM provider will sign a business associate agreement and the deployment includes the required technical safeguards, since HIPAA treats any vendor handling protected health information on a covered entity's behalf as a business associate regardless of how advanced the underlying technology is. The free, consumer version of ChatGPT and most public chatbot interfaces do not offer a business associate agreement, so pasting patient names, diagnoses, or record numbers into them is a HIPAA violation even if the intent was benign. Enterprise or API tiers from major LLM providers can support HIPAA use once a BAA is executed and the deployment adds encryption in transit and at rest, strict access controls, audit logging of every prompt and response, and a defined data retention and deletion policy. Many healthcare organizations choose a self-hosted, on-premise LLM instead specifically to avoid depending on a vendor's BAA terms and to keep protected health information from ever reaching a third-party network. This is general compliance information and not a substitute for a HIPAA risk assessment. Nanobase AI builds HIPAA-aligned LLM deployments, including fully on-premise options, for healthcare and insurance clients.
The business associate agreement is the gate, not a formality
HIPAA treats any vendor that creates, receives, maintains, or transmits protected health information on behalf of a covered entity as a business associate, and this classification applies to an LLM provider exactly as it applies to a billing vendor or a cloud storage provider. Without a signed business associate agreement in place, using that vendor with PHI is a HIPAA violation regardless of how the vendor markets its security practices. The BAA is a legal precondition, not a best practice, and no amount of technical care substitutes for its absence.
Why the free consumer tier fails immediately
The free, consumer-facing version of ChatGPT, Claude, or Gemini does not offer a BAA to individual users, which means any PHI entered into those interfaces is being processed by a business associate with no HIPAA agreement in place. This holds true regardless of intent, urgency, or how the clinician frames the query. A doctor pasting a de-identified case summary into a consumer chatbot to save time is still creating exposure if any of the supposedly de-identified fields turn out to be identifiable, since HIPAA's de-identification standard has specific technical requirements, not a general sense of "sounds anonymous." The free consumer tier fails the BAA test structurally, regardless of how careful the individual clinician is about what they type.
The full checklist for a HIPAA-compliant LLM deployment
| Requirement | What it looks like in practice |
|---|---|
| Signed BAA with the LLM provider | Confirmed and on file before any PHI is processed, not after |
| Encryption in transit and at rest | TLS for API calls, encrypted storage for any logs or cached prompts |
| Access controls | Role-based access limiting who can query PHI-containing prompts or view logs |
| Audit logging | Records of who accessed what PHI, when, satisfying the HIPAA audit control requirement |
| Minimum necessary standard | Prompts contain only the PHI required for the specific clinical or administrative task |
| Business continuity and breach response | A documented plan for detecting and reporting a breach involving the AI system |
Meeting the BAA requirement alone addresses the contractual gate; the remaining rows are the technical and administrative safeguards HIPAA's Security Rule separately requires. A signed BAA without the surrounding technical safeguards is still an incomplete HIPAA posture, not a finished one.
Practical deployment paths that satisfy the checklist
- Use the LLM provider's enterprise or API offering under a signed BAA, with training on inputs disabled and data retention configured to the minimum needed for operations.
- Route PHI through a de-identification or masking layer before it reaches the model wherever the task allows working with de-identified data.
- For the most sensitive workflows, deploy an open-weight model on the organization's own or a BAA-covered cloud infrastructure, removing the need to send PHI to an external inference API at all.
- Maintain the audit logging and access control layer independently of the model provider's own logging, since HIPAA's audit requirements apply to the covered entity's systems as a whole.
Self-hosting removes the "does the vendor sign a BAA" question entirely for the inference step, but it does not remove the surrounding safeguards HIPAA requires for encryption, access control, and audit logging.
This is general information, not legal advice, and a specific deployment should be reviewed against HIPAA's Security and Privacy Rules by qualified counsel or a compliance officer.
Frequently asked questions
Does de-identifying patient data remove the need for a BAA?
If the de-identification meets HIPAA's Safe Harbor or Expert Determination standard, the resulting data is no longer PHI and HIPAA's BAA requirement no longer applies to it specifically. Partial redaction that leaves indirect identifiers, such as rare diagnoses combined with a small zip code, often fails this standard.
Which LLM providers currently offer a BAA?
Both OpenAI and Anthropic offer a BAA for qualifying enterprise and API customers, detailed further in our comparison of OpenAI and Anthropic's BAA terms, though eligibility and exact terms should be confirmed directly with the vendor.
Is a HIPAA-compliant LLM deployment automatically GDPR or KVKK compliant too?
No, these are separate frameworks with separate requirements. A deployment can satisfy HIPAA's BAA and safeguard requirements while still needing its own GDPR lawful basis or KVKK transfer mechanism if patients are also EU or Turkish residents.
Can an AI scribe tool that transcribes doctor-patient conversations be HIPAA compliant?
Yes, if the vendor signs a BAA and the tool meets the same technical safeguard checklist as any other PHI-processing system. Ambient scribing tools are increasingly common in clinical settings and are evaluated the same way as any other business associate.
How Nanobase AI helps
Nanobase AI, a Silicon Valley enterprise AI engineering company, helps healthcare organizations deploy LLM-based tools, including private, on-premise models for workflows where PHI should never reach an external API, with the encryption, access control, and audit logging HIPAA's Security Rule requires. This is part of our AI security and compliance practice, and it draws on the same on-premise LLM deployment work we do for other regulated industries.
Ready to discuss your project? Contact Nanobase AI or email hello@bumu.tech.