Key takeaways
- Separate three kinds of texting: fixed templates, AI drafts a person approves, and AI replies sent on their own. Each needs different rules.
- FCC rules for exempt healthcare texts to mobile numbers include one message a day, three a week, 160 characters or less and an easy STOP opt-out[1].
- Opt-out requests must be honored within a reasonable time not exceeding ten business days[1], and TCPA damages start at $500 per violation[2].
- Keep health details out of texts. Use texts to prompt, confirm and route, and send personal information through the portal.
Why AI is moving into patient texting
Message volume keeps rising. In an MGMA poll, 70% of medical groups said patient portal message volume increased in 2024, and only 1% said it fell[3]. A time-motion study of primary care physicians found inbox management alone took 85 minutes a day, nearly a quarter of their EHR time[4]. Texts add to the same pile.
- Increased 70%
- About the same 29%
- Decreased 1%
Source: Medical Group Management Association (MGMA), 2024[3]
Show the data
| Item | Value |
|---|---|
| Increased | 70% |
| About the same | 29% |
| Decreased | 1% |
AI drafting is the obvious response, and physicians are already trying it: 19% told the AMA in 2026 that they use AI to draft responses to patient portal messages[5]. The same tools can draft texts to patients. That can save real time, but a text arrives on a personal phone, is easily forwarded and is covered by telecom rules that do not apply to portal messages. It needs its own guardrails.
Three kinds of AI texting
| Type | Example | Risk | Review rule |
|---|---|---|---|
| Fixed template, automated send | Appointment reminder with date, time and location | Low if templates are approved | Template approved once; content changes re-approved |
| AI draft, person approves | Reply to a patient asking what to bring to a first visit | Medium | Staff read and approve every draft before sending |
| AI reply sent on its own | Answer to where do I park? sent instantly | Higher | Only from an approved answer list; anything else routes to staff |
The line that matters most is between administrative and clinical content. An automated reply about parking is low stakes. An automated reply that interprets a symptom or a result is not acceptable, however well the model writes. Any patient text that mentions symptoms, results or medication should move to a person, and if it suggests an emergency, the reply should be a fixed message telling the patient to call 911.
Federal rules that shape healthcare texting
Texts count as calls under the Telephone Consumer Protection Act. FCC rules contain an exemption for certain healthcare messages to mobile numbers that are not charged to the patient, and they set conditions for providers relying on it[1]:
- Send only to the mobile number the patient provided.
- State the name and contact information of the healthcare provider.
- Stay within the permitted healthcare purposes, such as appointment reminders and lab result notices, with no telemarketing, solicitation or advertising, and no billing, debt-collection or other financial content.
- Send no more than one message a day and three a week to each patient.
- Keep texts to 160 characters or less, and voice calls generally to a minute or less.
- Offer an easy opt-out in every message; for texts, replying STOP must work.
Separately, a patient may revoke consent by any reasonable method, including replying stop or unsubscribe, and the request must be honored within a reasonable time not exceeding ten business days[1]. The law allows people to sue for actual loss or $500 per violation, whichever is greater, and courts may increase damages for willful violations[2]. Marketing texts, such as promoting a new cosmetic service, fall outside the healthcare exemption and need their own consent. Ask your counsel to review your texting program, including any state rules that apply.
A minimum-information template library
Every template should pass three tests: it fits in 160 characters, it names the practice and a contact number, and it would cause no harm if a family member read it over the patient's shoulder.
- Reminder: โ[Practice]: reminder of your visit [day] at [time]. Reply C to confirm, R to reschedule. [phone]. Reply STOP to opt out.โ
- Results ready: โ[Practice]: you have a new message in your patient portal. Questions? Call [phone]. Reply STOP to opt out.โ
- Form prompt: โ[Practice]: please complete your forms before Tuesday's visit using the link we emailed. Help: [phone]. Reply STOP to opt out.โ
- Callback notice: โ[Practice]: we got your message and will call you back today by 4pm. Emergencies call 911. Reply STOP to opt out.โ
Fill the bracketed fields from your system and count characters after filling, since a long practice name can push a text over the limit. Notice what the templates leave out: the type of visit, the provider's specialty, the test name and any result. Those details belong in the portal, where identity is checked.
Review rules for AI-drafted replies
- Every draft that mentions a symptom, a medication, a result or a diagnosis is reviewed by clinical staff, not front desk staff.
- Drafts that change an appointment, a balance or a referral are checked against the system before sending.
- Staff edit the draft rather than send and correct afterwards.
- Drafts never promise outcomes, such as your doctor will approve this refill.
- If a draft cites a policy, the policy must exist in your approved list.
- Log who approved each message and when.
Handling opt-outs properly
Recognize every form
Treat STOP, unsubscribe, cancel, quit, end and plain-language requests such as please stop texting me as opt-outs.
Confirm once
Send a single confirmation that texts have stopped and how to reach the practice by phone, then send nothing more on that program.
Update every system
Opt-outs must reach the reminder platform, the AI texting tool and the practice management record, not just one of them.
Offer an alternative
Flag the record so staff use calls or the portal for reminders, and ask at the next visit whether the patient wants texts back on.
Audit monthly
Sample opted-out numbers and confirm no messages were sent after the opt-out date.
Configuration checks before the first message
Most texting problems come from configuration, not from the AI's writing. Before any AI-drafted or automated text reaches a patient, walk through these checks with whoever administers the tool, and repeat them after every vendor update.
- Frequency caps are set per patient across all programs combined, so a reminder, a form prompt and a recall message cannot stack up on the same day.
- Quiet hours are configured so automated texts are not sent late at night or early in the morning in the patient's time zone.
- The clinical keyword list that forces a draft to clinical review is written by a nurse or physician, not the vendor, and includes common misspellings.
- The emergency keyword list triggers a fixed 911 reply and an alert to staff during hours.
- Inbound texts to the practice number that the tool cannot classify go to a person, not to an auto-reply.
- Minors' records are flagged so texts go to the parent or guardian number your policy specifies.
- Patients whose preferred language is not English get templates translated and reviewed by a qualified person.
- Every outbound message is logged with its template or draft, the approver and the time sent.
Assign one owner for the texting program. That person approves new templates, reviews the monthly opt-out audit and signs off configuration changes. Without an owner, guardrails decay quietly as staff add one-off messages to solve today's problem.
Example: a fictional reply queue
Example: a fictional pediatric practice lets AI draft replies to inbound texts, with front desk staff approving administrative drafts and a nurse approving anything clinical. In one morning, the queue holds a parking question, two reschedule requests, a question about whether a child with a rash can attend a vaccination visit, and a message that says stop sending these.
Example, continued: the parking reply goes out from the approved list. The reschedules are checked against the schedule and approved. The rash question is routed to the nurse, who calls the parent rather than replying by text. The stop message is processed as an opt-out across all systems, and the audit trail shows the confirmation sent within minutes. The practice notes that the AI draft for the rash question had suggested the visit could go ahead; the nurse rejected it, and the team adds rash to the clinical keyword list so such drafts are never offered to front desk staff.
What to measure
- Share of drafts approved without edits, by message type.
- Clinical messages correctly routed to clinical staff.
- Time from patient text to reply during business hours.
- Opt-out requests processed within your internal target, well inside the ten-business-day limit.
- Messages sent to opted-out numbers, which should be zero.
Whether patients should be told a reply was drafted by AI is covered in disclosing AI to patients. How chat on your website fits the same rules is in website chat assistants.
Frequently asked questions
Can a medical practice text patients without consent?
FCC rules contain an exemption for certain healthcare texts to mobile numbers, with conditions such as one message a day, three a week, 160 characters, no marketing or billing content and an easy opt-out. Marketing texts need consent. Have counsel review your program.
Should AI-drafted patient texts be reviewed?
Administrative replies from an approved list can go out automatically. Anything mentioning symptoms, results, medication or diagnosis should be reviewed by clinical staff before sending.
How quickly must a practice honor a STOP request?
FCC rules require revocation requests to be honored within a reasonable time not exceeding ten business days. Most practices should process them the same day across every texting system.
References
Legal Information Institute, Cornell Law School (Electronic Code of Federal Regulations text) (current). 47 CFR ยง 64.1200 - Delivery restrictions (opens in a new tab) โ
Legal Information Institute, Cornell Law School (current). 47 U.S. Code ยง 227 - Restrictions on use of telephone equipment (opens in a new tab) โ
Medical Group Management Association (MGMA) (2024). Ensuring accurate coding and billing for patient portal messages as e-visits (opens in a new tab) โ
Annals of Family Medicine (2017). Tethered to the EHR: Primary Care Physician Workload Assessment Using EHR Event Log Data and Time-Motion Observations (Annals of Family Medicine, 2017) (opens in a new tab) โ
American Medical Association (2026). More than 80% of physicians use AI professionally: AMA survey (opens in a new tab) โ