Loading
Skip to content
Miracle Minds AI
  • Try “HIPAA”, “model risk”, “FinOps” or “agentic”.

Patient engagement

The main risk in patient-facing AI is less a wrong medical fact than a misjudged sense of urgency. An assistant earns its place when it routes the message that needs a clinician now, and stays inside consent and privacy rules for everything else.

Part of AI for Life Sciences & Healthcare. 10 sources, checked .

of US individuals were offered and accessed their online medical records or patient portal in 2024, up from 25% in 2014
65%
of US individuals were offered and accessed their online medical records or patient portal in 2024, up from 25% in 2014
Assistant Secretary for Technology Policy / Office of the National Coordinator for Health IT, HHS, 2025(opens in a new tab)
of physician assessments judged that an LLM-drafted reply to a patient message posed a risk of severe harm (2023 study, 156 assessments)
7.1%
of physician assessments judged that an LLM-drafted reply to a patient message posed a risk of severe harm (2023 study, 156 assessments)
The Lancet Digital Health, 2024(opens in a new tab)
average share of patient message replies where clinicians used the AI draft in a 2023 Stanford pilot, with no change in reply time
20%
average share of patient message replies where clinicians used the AI draft in a 2023 Stanford pilot, with no change in reply time
JAMA Network Open, 2024(opens in a new tab)

The problem

Patients now reach their providers online by default. In 2024, 65% of individuals nationally were offered and accessed their online medical records or patient portal, up from 25% in 2014.1 Every message lands in a clinician's inbox. In 2020, patient messages rose to 157% of the prepandemic average, and each additional message was associated with 2.32 more minutes of EHR time per clinician per day.2

Generative AI drafting is the obvious response, and early results are mixed. In a Stanford pilot, clinicians used AI-drafted replies for an average of 20% of patient messages; task load and work exhaustion scores fell, but reply time did not.3 In a Brigham and Women's study, reviewing physicians judged that LLM drafts posed a risk of severe harm in 7.1% of 156 assessments, mostly because the draft misjudged the urgency of the situation.4

Outreach carries its own legal exposure. The FCC has confirmed that AI-generated voices count as artificial voices under the TCPA,5 and calls using an artificial voice generally need the called party's prior express consent.6 HIPAA excludes treatment and care coordination messages from marketing, while marketing communications require the patient's authorization.7,8 Health apps outside HIPAA fall under the FTC Health Breach Notification Rule, where an unauthorized disclosure counts as a breach.9

Language access is the other design constraint. Section 1557 rules require that machine translation be reviewed by a qualified human translator when the text is critical to a patient's rights, benefits or access, or when accuracy is essential.10 For a multilingual assistant, that means deciding up front which messages can go out automatically and which must pass a human reviewer first.

What we build

  1. 01

    Inbox triage and routing

    Classify incoming portal messages by urgency and type, send refill and scheduling requests to the right team, and put possible emergencies in front of a clinician at once with the reason flagged.

  2. 02

    Clinician-reviewed reply drafts

    Drafts grounded in the patient's chart and your approved content, shown to the clinician to edit and sign. Clinical replies are never sent automatically.

  3. 03

    Consent-aware outreach

    Appointment, screening and follow-up reminders by text or voice that check consent, call-frequency limits and opt-outs before every contact, and log the basis for each one.

  4. 04

    Multilingual patient assistants

    Help with scheduling, directions, billing and visit preparation in the patient's language, with human translator review for critical content and a visible route to a person.

  5. 05

    Post-discharge check-ins

    Structured follow-up conversations after discharge that collect symptoms and medication questions and escalate defined red flags to the care team.

Where it goes wrong

The failure modes we design against from the first week, and the control for each.

  • Missing the urgent message

    Harmful AI drafts most often misjudged acuity rather than medical facts.4 Control: an urgency classifier tuned for sensitivity, hard-coded red-flag escalation, and clinician review of every clinical reply.

  • Clinicians adopting the AI's assessment

    Physicians using LLM drafts shifted their own responses toward the model's content, a sign of automation bias and anchoring.4 Control: present drafts as suggestions, track how much clinicians edit, and audit samples against clinical standards.

  • Outreach that becomes marketing or unconsented calls

    A message promoting a third party's product, or one sent in exchange for payment, can count as HIPAA marketing,7 and an AI voice call without consent can breach the TCPA.5 Control: classify every campaign by purpose with compliance sign-off, and enforce consent and opt-out checks in code.

  • Health data leaking through apps and chat

    Chat transcripts and app analytics can send health data to third parties, and outside HIPAA the FTC treats an unauthorized disclosure as a breach that triggers notice.9 Control: written agreements with every vendor that touches the data, no third-party trackers on authenticated pages, and minimal data in prompts and logs.

Rules that apply

  • HIPAA Privacy Rule marketing provisions, 45 CFR 164.501 and 164.508(a)(3)

    Primary text on ecfr.gov(opens in a new tab)

    Treatment and care coordination messages are not marketing, but any use of patient information for marketing requires the patient's authorization.

  • Telephone Consumer Protection Act rules, 47 CFR 64.1200

    Primary text on ecfr.gov(opens in a new tab)

    Calls using an artificial or prerecorded voice need prior express consent, with limited exemptions for health care messages from HIPAA covered entities; the FCC has confirmed AI-generated voices are covered.

  • FTC Health Breach Notification Rule, 16 CFR Part 318

    Primary text on ecfr.gov(opens in a new tab)

    Vendors of personal health records not covered by HIPAA must notify consumers after a breach, which includes an unauthorized disclosure of health data.

  • Section 1557 language access, 45 CFR 92.201

    Primary text on ecfr.gov(opens in a new tab)

    Machine translation of critical content, or content where accuracy is essential, must be reviewed by a qualified human translator.

How to measure it

  • Share of urgent messages escalated to a clinician within the target time
  • Draft acceptance and edit rates by clinician and message type
  • Inbox time per clinician per day
  • Share of outreach contacts with a documented consent basis (target: all)
  • Patient requests resolved without a callback, and complaint rate

Questions buyers ask

Can an AI chatbot answer patients' medical questions directly?

For general matters such as directions, visit preparation and billing, yes, within clear limits. For anything clinical, keep a clinician in the loop. Studies of AI-drafted replies found the dangerous errors were mostly misjudged urgency, which a patient cannot catch. Route clinical questions to a reviewed draft and escalate red-flag symptoms immediately.

Do we need patient consent for AI-generated reminder calls?

Generally yes. The FCC has confirmed that AI-generated voices count as artificial voices under the TCPA, so calls to mobile numbers need prior express consent. Health care messages from HIPAA covered entities have some exemptions, with call-frequency limits for some lines and a duty to honor opt-outs. Record the consent basis for each patient and honor opt-outs across every channel.

Is a patient-facing AI assistant covered by HIPAA?

If a covered entity or its business associate runs it, yes, and any vendor handling the data must give written assurances, usually a business associate agreement. If a company offers a health app directly to consumers outside HIPAA, the FTC Health Breach Notification Rule can apply, and sharing data without authorization can count as a breach. Map every data flow before launch.

Sources

Every figure on this page comes from the sources below. We checked each link on .

  1. 1.Individuals' Access and Use of Patient Portals and Smartphone Health Apps, 2024 (ASTP Data Brief No. 77)(opens in a new tab)Assistant Secretary for Technology Policy / Office of the National Coordinator for Health IT, HHS, July 2025 · healthit.gov
  2. 2.Assessing the impact of the COVID-19 pandemic on clinician ambulatory electronic health record use(opens in a new tab)Journal of the American Medical Informatics Association, January 29, 2022 · pmc.ncbi.nlm.nih.gov
  3. 3.Artificial Intelligence-Generated Draft Replies to Patient Inbox Messages(opens in a new tab)JAMA Network Open, March 4, 2024 · pmc.ncbi.nlm.nih.gov
  4. 4.The effect of using a large language model to respond to patient messages(opens in a new tab)The Lancet Digital Health, April 24, 2024 · pmc.ncbi.nlm.nih.gov
  5. 5.FCC Confirms that TCPA Applies to AI Technologies that Generate Human Voices (Declaratory Ruling, FCC 24-17)(opens in a new tab)Federal Communications Commission, February 8, 2024 · fcc.gov
  6. 6.47 CFR 64.1200: Delivery restrictions(opens in a new tab)Electronic Code of Federal Regulations (FCC) · ecfr.gov
  7. 7.45 CFR 164.501: Definitions (marketing)(opens in a new tab)Electronic Code of Federal Regulations (HHS) · ecfr.gov
  8. 8.45 CFR 164.508: Uses and disclosures for which an authorization is required(opens in a new tab)Electronic Code of Federal Regulations (HHS) · ecfr.gov
  9. 9.16 CFR Part 318: Health Breach Notification Rule(opens in a new tab)Electronic Code of Federal Regulations (FTC), May 30, 2024 · ecfr.gov
  10. 10.45 CFR 92.201: Meaningful access for individuals with limited English proficiency(opens in a new tab)Electronic Code of Federal Regulations (HHS) · ecfr.gov