HIPAA + AI voice receptionists: what actually matters
Every dental SaaS vendor says they're HIPAA-compliant. Most aren't lying, but the word does different work in different places. Here's what actually matters for an AI voice receptionist handling PHI.
Start with the BAA
A signed Business Associate Agreement is table stakes. Not a PDF on their website — a signed one, executed before any PHI is processed. If the vendor doesn't provide a sample BAA for review, that's a red flag. We host ours at /legal/baa-sample.
The safeguards that actually move the needle
HIPAA Security Rule breaks down into administrative, physical, and technical safeguards. In practice, four of them do 80% of the work:
1. **Encryption at rest and in transit.** AES-256 for stored data, TLS 1.3 for anything on the wire. Non-negotiable.
2. **Access controls with audit trails.** Who touched what PHI, when, from where. Every access, logged, retained 6 years.
3. **Row-level security on every PHI table.** Tenant A's patient data cannot — at the database level — be read by Tenant B's credentials. Ever. Test it with a red-team query.
4. **Signed BAAs with every subprocessor.** Your vendor's vendors also touch PHI. Each one must have a BAA or equivalent downstream contract.
Questions to ask a vendor
Before signing anything, ask these and demand specific answers:
- Who are your subprocessors? (Every one, not the short list.)
- Where is PHI stored, at rest? What encryption?
- How long is audit log retention? 6 years is the minimum.
- What's your breach notification timeline?
- When was your last third-party penetration test? Can we see a summary under NDA?
- What's your SOC 2 status? 'In progress' is acceptable; 'we'll get to it' is not.
The AI wrinkle
AI voice systems introduce a specific risk: prompts can accidentally cross tenant boundaries if engineered poorly. A model trained or cached on Patient A's conversation should never surface in Patient B's call.
Ask: 'Is there any cross-call or cross-tenant retention in your LLM stack?' The correct answer is no — each conversation is ephemeral, the prompt includes only that tenant's context, and any model fine-tuning happens on anonymized data with explicit customer consent.
The call recording question
Voice calls are legally dicier than text. State two-party consent laws apply. Your AI vendor needs to handle the recording disclosure ('this call may be recorded for quality and training purposes') at the start of every call in applicable states. Check that in their demo.
What we do
PracticeIQ runs on Supabase with RLS on every PHI table, AES-256 at rest, TLS 1.3 in transit, audit_log with 6-year retention, and signed BAAs with every subprocessor we touch — Twilio, Vapi, OpenAI/Anthropic, Stripe, Supabase, Vercel.
We're pre-SOC 2. Type 1 targeted Q3 2026, Type 2 Q4 2026. We'll document every control publicly on /trust as we finish it. Ask us anything at security@practiceiq.ai.