Overview
This series examines how AI-drafted patient-portal messages move from patient question to clinical release. It explains why drafting may reduce composing burden without proving faster or safer care, how fluent omissions can survive clinician review, and how patient purpose and stakes should shape the channel and response. It then connects accountable review, retrieval-augmented checking, plain disclosure, escalation, and outcome monitoring into a guarded adoption model. Evidence from 40 patient interviews, a simulation of 20 physicians and 18 messages, and a preprint with limited validation supports design conditions—not universal safety, trust, or outcome claims.
Learning outcomes
- Clinician accountability remains necessary, but a send action does not prove effective review.
- Fluency can hide omissions of urgency, safety-net instructions, alternative explanations, or appropriate next actions.
- Patient purpose, clinical stakes, uncertainty, and channel should determine message form and review intensity.
- Guardrails, disclosure, escalation, and monitoring support guarded adoption, but available studies do not prove universal safety or better outcomes.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Can AI-drafted patient-portal messages make care safer?
The available studies do not show that AI drafting makes care safer. They support guarded adoption only when active clinician review, independent checking, appropriate escalation, disclosure, and monitoring surround the draft.
What should a clinician check before sending an AI-drafted portal message?
A clinician should check whether the draft answers the patient’s actual question, uses current guidance, makes urgency and next steps clear, and fits the channel. They should look for omissions and escalate when asynchronous messaging is inadequate.
Why can a fluent AI-drafted message still be unsafe?
A fluent draft can still be unsafe because it may omit urgency, a safety net, an alternative explanation, or a needed action while sounding complete.
What do patients want from AI-drafted portal messages?
Patients in the 40-person qualitative study generally wanted useful, appropriately concise communication, transparency about AI use, and a clinician to remain responsible. Preferences varied with purpose and stakes; some wanted more detail or a call for serious concerns.