Explainer

HIPAA and AI phone agents: the checklist before you sign

September 8, 2026 · 4 minute read

A shield deflecting arrows

The short answer

An AI phone agent handling patient calls is a business associate under HIPAA and requires a signed business associate agreement before it takes a single call. The 2026 Security Rule updates added encryption, multi-factor authentication and annual penetration testing requirements. Civil penalties run up to $50,000 per violation.

Why a phone call counts

Practices often assume HIPAA attaches to records rather than conversations. It attaches to protected health information wherever it lives, and a call in which somebody gives their name and asks about an appointment is PHI.

That means any vendor answering your phone is handling PHI on your behalf, which makes them a business associate. Business associates require a signed business associate agreement before they begin, not after the pilot goes well.

This is general guidance rather than legal advice. Whoever handles compliance at your practice should read any agreement before it is signed.

What changed in 2026

The Security Rule updates raised the technical floor and made several things explicit that were previously interpreted.

Encryption, multi-factor authentication and annual penetration testing are now expected for systems handling PHI. For a vendor that means specifics you can ask about rather than assurances.

The distinction worth holding: every vendor will say they are HIPAA compliant, because compliance is a posture rather than a certificate. What separates them is what they can evidence.

Compliance is a posture, not a certificate. Ask what they can evidence.
Questions being asked

The checklist

Work through this before signing anything. A vendor who cannot answer these quickly is telling you something.

Ask, and get it in writing

  • Will you sign a BAA, and may I see it before committing rather than after
  • Where are call recordings and transcripts stored, and for how long
  • Is call data used to train your models, and can that be switched off
  • Is data encrypted at rest as well as in transit
  • Who at your company can access recordings, and is that access logged
  • Do you subcontract any part of this, and are those subcontractors covered by the BAA
  • What happens to our data if we leave, and how quickly is it deleted
  • Have you had a penetration test in the last twelve months, and will you share the summary

The red flags

A vendor that says HIPAA compliant on the website but has no BAA ready is the most common one. If the agreement has to be requested from legal and takes a week, they are not doing this routinely.

Vague answers about subcontractors are the second. Many AI phone products are built on a general speech provider, and if that provider is not covered by the chain of agreements, your PHI is somewhere nobody has papered.

Training on your call data is the third. Some products improve by learning from customer calls, and that is a reasonable business model in most industries and a serious problem in this one. It should be off by default and stated in writing.

What good looks like

A vendor doing this properly will have a BAA on file that they send without being chased, a documented retention period, an option to disable training, and a named subprocessor list.

They will also usually be more expensive, which is worth understanding rather than resenting. The cheapest tier of any AI answering product is generally built for businesses without a compliance obligation, and a clinic buying that tier is buying the wrong product.

Human-staffed services such as Smith.ai, at roughly $9.75 a call, have been operating under BAAs for years and are a reasonable benchmark for both price and process.

The narrower option worth considering

If the compliance overhead feels disproportionate to the problem, narrow the problem. An AI that never discusses anything clinical, only takes a name and a callback number and books from an availability list, handles far less sensitive information than one answering questions about treatment.

It still requires a BAA, because a name attached to a healthcare provider is PHI. But the surface area is smaller, the vendor list is wider and the conversation with your compliance person is shorter.

That is often the sensible first step: solve the missed calls, which is where the money is, without handing a vendor your whole clinical conversation.

Common questions

Does an AI phone agent need to be HIPAA compliant?
Yes. A call in which a patient gives their name and asks about an appointment is protected health information, which makes the vendor a business associate requiring a signed BAA before they take any calls.
What changed with HIPAA in 2026?
The Security Rule updates made encryption, multi-factor authentication and annual penetration testing expected for systems handling PHI, which gives you specific things to ask a vendor about rather than general assurances.
What are the penalties?
Civil penalties run up to $50,000 per violation. The practical exposure is usually larger than the fine, since a breach involving patient calls is also a reputational event in a local market.
What is the biggest red flag in a vendor?
Saying HIPAA compliant on the website but having no BAA ready to send. If the agreement takes a week to produce from legal, they are not doing this routinely.
Can I use a cheaper AI answering service?
The cheapest tiers are generally built for businesses with no compliance obligation, so a clinic buying one is buying the wrong product. Narrowing the job to booking only, with no clinical discussion, widens the field of usable vendors.