At 11:05 on a Wednesday night, a message arrives in a dental practice's website chat: 'Had a root canal Monday. The left side of my face is swelling and it's getting hard to swallow. Is that normal? Can I just take more ibuprofen?' Two minutes later, another: 'This is Dana Ruiz's mom. What time is her appointment Friday? She's 24.'
Neither message is unusual. The first needs to leave the chat immediately and reach emergency care. The second asks the chatbot to share an adult patient's appointment with someone whose identity and authority it can't confirm.
HIPAA does let a practice use professional judgment to share some information with family members involved in a patient's care, but that's a call for a staff member who knows the situation, not for a chatbot. A chatbot designed for a practice handles both well. One that simply 'answers questions' will get at least one of them wrong.
When does a chatbot handle protected health information?
Not every chat involves protected health information (PHI). A visitor asking about your Saturday hours or parking hasn't told you anything about their health. The picture changes the moment a chat links an identifiable person to their health, their care or their payment for it.
| Message | PHI involved? | What it means for the chatbot |
|---|---|---|
| 'What are your Saturday hours?' | No | Answer from your website content |
| 'Do you take Delta Dental?' | No, if nothing identifies the visitor | Answer from your published insurance list |
| 'I'm Sam Lee and I need to book for tooth pain' | Yes | The vendor is handling PHI for you; a BAA is required |
| 'When is my appointment?' | Yes | Verify identity before answering |
| 'What did my lab results say?' | Yes, and clinical | Portal only, released under your clinicians' rules |
In practice, a patient chatbot that books appointments will handle PHI within its first hour, so plan for that from day one. A vendor that creates, receives, maintains or transmits PHI on your behalf is a business associate, and HHS guidance says that holds even for a cloud provider storing encrypted data it can't view.
Five requirements for a patient chatbot
1. A Business Associate Agreement that covers this product
The BAA should name the chatbot service and require the vendor to safeguard PHI, report incidents, hold its own subcontractors (the AI model provider, the hosting company, the texting service) to the same terms, and return or destroy data when the contract ends. HHS publishes sample BAA provisions you can compare a vendor's contract against.
Ask any chatbot vendor these five questions in writing before patient data flows:
- Will you sign a BAA before we send any patient data, and does it name this product?
- Which subcontractors touch chat data, and are they bound to the same terms?
- Does any model get trained or tuned on our chat transcripts, even de-identified ones?
- How long are transcripts kept, and can we set or shorten that period?
- Can we export the logs for an audit or a patient's request?
2. Minimum necessary, by design
HIPAA's minimum necessary standard asks you to limit PHI to what the purpose requires. For a booking chatbot that means name, date of birth, contact details and a short reason for the visit, not a medical history. Don't connect the chatbot to the whole chart when all it needs is the schedule.
3. Identity checks before anything personal
HIPAA's verification rules require a practice to confirm the identity and authority of someone requesting PHI when it doesn't already know them, with some exceptions. A name and birthday typed into a chat box is thin evidence. Keep personal answers such as appointment details, balances and results behind a portal login or a one-time code sent to the phone number on file, and have the chatbot explain that politely.
4. Emergencies and clinical questions leave the chat
A clinician writes and approves the list of red-flag symptoms, such as swelling that affects breathing or swallowing. The chatbot only matches messages against that list. A message that matches gets one instruction immediately: call 911 or go to the emergency room, plus your after-hours line where appropriate.
A message the chatbot can't sort with confidence goes straight to the on-call clinician or after-hours line, with a plain line to the patient: 'If you think this is an emergency, call 911.' Clinical questions ('Can I take more ibuprofen?') go to your clinical team's queue with the conversation attached. The chatbot doesn't triage, diagnose or advise; it follows the list your clinicians wrote.
5. Logs, access and retention
The HIPAA Security Rule requires audit controls that record activity in systems holding electronic PHI. Make sure every conversation is logged, staff reach the logs through their own accounts, and the vendor's retention period matches your policies.
Flowchart
How a patient chatbot sorts an incoming message
- StartMessage arrives in the website chat at 11:05 PM
- DecisionDoes it match the clinician-written red-flag list?YesA personTell the patient to call 911 or go to the ER now; alert the on-call clinicianNo
- DecisionCan the chatbot sort the message with confidence?NoA personRoute to the on-call clinician or after-hours line; tell the patient: if you think it's an emergency, call 911Yes
- DecisionIs it a clinical question?YesA personSend it to the clinical team's queue and tell the patient when to expect a replyNo
- StepFor anything personal, the patient verifies through the portal or a one-time code
- StepChatbot books, confirms or answers from approved content only
- DoneConversation logged; staff review flagged chats each morning
Before and after
A generic chat widget vs a chatbot set up for a practice
Generic chat widget
- 01No BAA, and terms that allow broad use of chat data
- 02Swelling message gets a friendly 'we'll reply tomorrow'
- 03Gives Dana's Friday appointment time to whoever asks
- 04Transcripts sit in a marketing dashboard
ResultA delayed emergency and an improper disclosure
Patient-ready chatbot
- 01BAA signed, with subcontractors bound to the same terms
- 02Swelling message: call 911 or go to the ER, on-call alerted
- 03Appointment details shared only after the patient verifies
- 04Transcripts logged in an access-controlled system
ResultUrgent care first, privacy kept, staff informed
AI disclosure laws: California's AB 3030
Some states now regulate AI-written patient messages directly. Since January 1, 2025, California's AB 3030 has covered health facilities, clinics and physician offices whose patient messages about clinical information are produced by generative AI. Those messages need a disclaimer saying AI generated them, plus clear instructions for reaching a human. In a chat, the disclaimer must stay visible throughout; in a phone call, it's spoken at the start and the end.
Two limits matter. Administrative messages such as scheduling and billing are outside the law, and a message that a licensed or certified provider reads and reviews before it's sent is exempt. The law names health facilities, clinics, physician offices and group practice offices; whether a particular dental office falls within it is a question for your counsel, and disclosing anyway costs nothing.
Even where no law requires it, telling patients up front that they're chatting with an automated assistant is the right default. Requirements vary by state, and nothing here is legal advice, only general information.
A launch checklist
- 01BAA signed, naming the chatbot product and binding its subcontractors.
- 02Scope written down: what it answers, what it books and what it never touches.
- 03Emergency wording approved by a clinician and tested with 20 realistic messages.
- 04Unclear messages routed to the on-call clinician or after-hours line, with the 911 line shown to the patient.
- 05Identity verification in place before any personal detail is shared.
- 06Clinical questions routed to a staffed queue, with a reply time the patient is told.
- 07Automated-assistant disclosure at the start of every chat, plus state-specific disclaimers where required.
- 08Chatbot access limited to the systems it needs, such as the schedule, not the full chart.
- 09Logs reviewed weekly, with anything mishandled fixed in the content or the rules.
Run the Dana Ruiz test before launch: have a colleague pose as a relative asking for an adult patient's appointment, another describe swelling after a procedure, and a third send something ambiguous, such as 'my crown feels weird and my jaw hurts.' If any of them gets the wrong response, the chatbot isn't ready.
Phone assistants follow the same principles; our post on dental front desk hold times covers that side. For a chatbot built this way, see our AI chatbots and assistants and how we work with medical and dental practices.
Sources
- HHS: Business Associate Contracts (sample provisions)
- HHS: Guidance on HIPAA and Cloud Computing
- eCFR: 45 CFR 164.502(b), minimum necessary standard
- eCFR: 45 CFR 164.514(h), verification requirements
- eCFR: 45 CFR 164.510(b), disclosures to family members involved in a patient's care
- eCFR: 45 CFR 164.312, technical safeguards including audit controls
- California Department of Public Health: AFL 25-07, Assembly Bill 3030, Health Care Services: Artificial Intelligence
- California Legislative Information: AB 3030 (2024), Health care services: artificial intelligence