HIPAA-Compliant AI Voice Agent: What to Ask Before You Buy

You're a clinic owner about to sign for an AI receptionist, and the vendor's site says "HIPAA compliant" in the hero. That phrase is doing a lot of work, and most of what it hides sits in the plumbing you never see: the pieces of the voice pipeline that each touch a patient's protected health information and each need their own paperwork.

Quick answer: A HIPAA compliant AI voice agent is only compliant if every layer that touches PHI — the language model, speech-to-text, text-to-speech, and telephony — is covered by a signed Business Associate Agreement (BAA), not just the vendor you sign with. Consumer Claude and ChatGPT are not HIPAA eligible, so an agent built on them cannot be compliant no matter what the marketing says.

I build voice agents for a living. "Emily," the agent we run on a law firm's phone line, booked 453 new clients in eight months and caught all 176 after-hours calls with a 96.5% self-serve rate. Healthcare raises the stakes: get the compliance wrong and a booking win turns into a breach. This is the buyer-education piece the vendor lists skip, and the usual disclaimer applies — I'm not a lawyer, and none of this is legal advice. Run your final decision past counsel who knows your state.

Why "HIPAA compliant" on a vendor's homepage isn't enough

A voice agent is not one product; it's a chain of services, and HIPAA follows the PHI down every link of that chain. When a patient says their name, date of birth, and why they're calling, that audio and its transcript pass through several separate companies before the call ends.

Here's the pipeline, and why each layer is a HIPAA question, not a technical one:

Each of those is a separate company, and under HIPAA each one that handles PHI is a business associate that needs its own signed BAA. This is the sub-processor chain, and it's exactly what the pitch glosses over. A vendor can truthfully say "we signed a BAA" while one link in their pipeline — often the cheapest STT or TTS provider — has no BAA at all. The whole chain is only as compliant as its weakest signed agreement.

The single most common way this breaks: an agent wired to consumer Claude or ChatGPT. Anthropic and OpenAI both offer HIPAA-eligible paths under an enterprise agreement with a signed BAA, but the consumer apps and default API tiers are not covered. An agent quietly calling api.openai.com on a standard account is sending PHI to a company that has not agreed to protect it. That's not a gray area.

The pre-purchase checklist: what to ask before you sign

Ask for each item below in writing before money changes hands. A vendor who actually built for healthcare answers these in one email; a vendor who painted "HIPAA compliant" on the homepage stalls.

The vendor SERP converges on the first three of these — Simba, Greetmate, Dialzara, and Thoughtly all name BAA coverage, sub-processor chains, and minimum-necessary handling. The last three (audit logs, retention, scope limits) are where even the good lists thin out, and where a bad deployment quietly leaks.

DIY on consumer AI vs a built agent on a HIPAA-eligible stack

Wiring your own agent to the consumer Claude or ChatGPT app is the fast path to a compliance problem, because those tiers were never eligible to sign a BAA for your PHI. The honest comparison isn't "build vs buy" — it's "which stack is actually covered."

We're a Gold Retell partner, and we build clinic agents on Retell because it offers a HIPAA path and on-premise SIP — meaning the telephony can stay inside your controlled environment instead of routing PHI through a shared cloud number. If you want the agent handled end to end with the BAA chain sorted before it takes a single call, that's what our team does: get a custom AI agent built. We answer the sub-processor question with the executed agreements, not a marketing claim.

Common pitfalls that turn a "compliant" agent into a breach

Most failures aren't the model saying something wrong — they're a gap in the paperwork or the handoff design. These are the ones I see clinics walk into.

The first is trusting the homepage badge. "HIPAA compliant" is not a certification anyone issues; there's no HIPAA seal. It's a claim about your controls and your BAAs, which is why the written proof above matters more than the logo.

The second is the incomplete BAA chain. The vendor signs with you, but their TTS provider or their transcription vendor never signed with them. You won't find this without asking for the sub-processor list explicitly — it's the exact thing Greetmate's "what vendors won't tell you" framing is pointing at.

The third is scope creep into medical advice. A caller asks "should I go to the ER?" and an unconstrained agent answers. Beyond the liability, that's the agent operating outside what you scoped. The fix is the same escalation discipline that separates good voice deployments from bad ones across every vertical: a clean handoff path to a human, triggered the moment the call leaves logistics. If you're setting one up, our walkthrough on how to test your AI voice agent before it goes live covers building those escalation triggers before real patients hit them.

The fourth is skipping the deletion policy because the agent "works." It works right up until an audit asks how long you've been storing call recordings and who can reach them. Decide the retention window before launch, not after a records request.

FAQ

Is a consumer ChatGPT or Claude account HIPAA compliant for a voice agent?

Consumer ChatGPT and Claude accounts are not HIPAA eligible, so a voice agent built on them cannot be compliant. Both Anthropic and OpenAI offer HIPAA-eligible enterprise paths under a signed BAA, but the default consumer apps and standard API tiers do not carry that coverage.

What is a sub-processor BAA chain for a voice agent?

A sub-processor BAA chain is the set of Business Associate Agreements covering every company in the voice pipeline that touches PHI — telephony, speech-to-text, the language model, and text-to-speech. Because each layer is a separate business associate under HIPAA, each one needs its own signed BAA, and the agent is only compliant if the whole chain is covered.

Does a HIPAA compliant AI voice agent need SOC 2 Type II or is Type I enough?

SOC 2 Type II is what you want. A Type I report confirms controls existed on a single day, while a Type II report confirms an auditor observed those controls operating over six to twelve months — evidence the practice actually holds, not just that it was set up once.

Can an AI receptionist give medical advice to callers?

An AI receptionist should be scoped so it never gives medical advice — it books, routes, verifies insurance, and answers logistics, then hands off to a human the moment a caller asks anything clinical. Letting the agent improvise about symptoms creates liability and pushes it outside the role you scoped for it.

Who is responsible if a HIPAA breach happens through the voice agent?

Both the covered entity (your clinic) and the business associate (the vendor and its sub-processors) can carry liability under HIPAA, which is why the BAA matters — it defines each party's obligations. This is exactly the kind of question to run past counsel who knows your state before you sign.

Getting this right is mostly about knowing which questions force a straight answer, and that's the same muscle whether you're buying an agent or building one. If you want to learn how these agents are actually put together — the escalation design, the scoping, the stack choices — the free Claude Community is where builders and founders work through it together, with copy-paste templates and live workshops to shorten the path.