How should a clinic use AI for patient communication?

Patient communication divides more cleanly than most business messaging. On one side is administrative content: you have an appointment at this time, please complete this form, your balance is this, we need this document. Every element of it exists in a record already, which makes it assembled rather than generated and puts it in the category safe to send without a person composing it.

On the other side is anything a patient could act on medically. A question about symptoms, a query about medication, a result, a change to treatment, or advice about whether to come in. These require a clinician, not because the wording is difficult but because the content is a clinical act. A system that answers them helpfully has performed one, and the fact that it did so in a chat window rather than a consulting room does not change what it was.

Unsolicited outbound messaging carries a separate set of rules and they are strict. Automated calls and texts to patients are governed by federal telephone consumer protection rules with specific provisions for healthcare-related messages, and the relevant questions are what consent exists, what the message contains and how a patient opts out. These are conditions on the message rather than guidance, and they apply identically whether a person or a system composed it.

The messages worth automating first are the ones that reduce a measurable loss. Appointment reminders reduce non-attendance, which is a direct and quantifiable cost in a clinic. Form completion before arrival reduces the time at the desk. Recall for a due appointment generates visits that would not otherwise happen. Each has a defined trigger in the record and a message drawn from it, which is exactly the safe category.

Language and reading level are worth an explicit decision because they affect who the communication reaches. Generated text tends towards a register more formal than most patient communication should be, and a clinic serving a population with mixed language needs has an accessibility question that predates any automation. Setting the target register and reading level once, and checking against it, is a small task with a real effect.

Finally, everything here sits behind the same threshold question as any other tool in a practice: whether the vendor handling the information has signed for it. Patient names, appointment details and clinical context are protected information, and a messaging service that transmits them on the practice's behalf is handling them regardless of how brief the message is.