Your Doctor's Reply Might Be Written by AI Now. Patients Are Starting to Notice.

This account draws on health system disclosures, vendor product documentation from Epic and independent published survey data on patient attitudes toward AI-assisted communication. Specific patient experiences described are illustrative of documented industry patterns rather than a single reported case.

A growing number of health systems using the Epic electronic health record have quietly turned on a feature that drafts AI-generated replies to patient portal messages before a clinician reviews and sends them. The clinician still has to approve the message, often editing it, but the first draft, the words a patient eventually reads as coming from their doctor, increasingly starts life as a language model's output rather than a human being's.

This is a distinct category from ambient scribes, which transcribe and summarize a clinical visit that already happened. This is generative text written to a patient, in a doctor's voice, about their own health question, often without the patient ever being told which parts, if any, were written by AI.

Why health systems are doing this

The portal message load on physicians has become one of the most-cited drivers of burnout in primary care. Patients now message their doctors about everything: medication questions, symptom concerns, requests for refills, results interpretation, follow up questions after a visit. Answering each of these thoughtfully takes real physician time, uncompensated time in many fee-for-service models, that has expanded dramatically since patient portals became standard and messaging volume surged, a trend that accelerated further during and after the pandemic.

An AI draft that pulls in the patient's chart context, prior messages, medication list, recent labs, and produces a clinically reasonable first pass at a response can meaningfully cut the minutes a physician spends per message, particularly for the routine, high-volume categories: refill confirmations, normal lab result explanations, standard post-procedure guidance. That efficiency case is genuine, and it mirrors exactly the burnout relief that ambient scribes have already demonstrated in the clinical documentation context.

The trust problem that is specific to this use case

Here is where this differs meaningfully from an ambient scribe, and why it deserves its own scrutiny rather than being treated as a smaller version of the same story. An ambient scribe produces a note that lives inside the medical record, reviewed by other clinicians, largely invisible to the patient. An AI-drafted portal message is read directly by the patient, presented as their doctor's personal response to their personal concern, at a moment when they are often anxious, seeking reassurance, or trying to understand something frightening about their own body.

Patient trust in a physician relationship is built substantially on the belief that the person on the other end of that message actually engaged with their specific situation. Survey data on patient attitudes toward AI in healthcare consistently shows a split reaction: many patients report comfort with AI assisting behind the scenes to make their doctor more efficient, but significantly less comfort when they learn a message they believed was personally composed by their physician was substantially AI-generated, particularly if that was not disclosed.

That gap, between passive institutional acceptance of AI-assisted workflows and patient discomfort upon actually discovering the practice, is the central tension of this story. Most health systems using this functionality do not proactively disclose which specific messages were AI-drafted. Patients generally have no way to distinguish an AI-assisted reply from a fully physician-composed one unless a system voluntarily labels it.

The clinical risk, distinct from the trust risk

Beyond the relational question, there is a distinct clinical safety concern. A physician reviewing and editing dozens of AI-drafted messages per day, under the same time pressure that made the tool attractive in the first place, is at risk of a specific failure mode: superficial review. If the draft looks reasonable, matches the general shape of what a clinician would say, there is a real risk that the reviewing physician skims rather than genuinely evaluates whether it correctly addresses this specific patient's situation, particularly for messages that involve subtle clinical nuance a language model might miss, such as a medication interaction not obviously flagged in the visible chart summary the model was given.

This mirrors precisely the automation bias concern that has already been documented with ambient scribes and diagnostic AI more broadly: tools designed to reduce cognitive burden can, paradoxically, reduce the very scrutiny that was supposed to remain a human safeguard, especially as clinicians build trust in the tool's typical reliability over time.

The systemic tension

The honest tension here is between two legitimate goods that are currently in conflict. Physician time is a finite, expensive, burnout-prone resource, and anything that gives clinicians meaningful time back without compromising care quality is worth pursuing seriously. Patient trust in the authenticity of their relationship with their doctor is also a finite resource, one that, once eroded by a sense of being quietly handled by a machine, is difficult to rebuild.

Health systems are currently resolving this tension by default, through non-disclosure, rather than by deliberate policy. That is the part that should concern anyone thinking seriously about how this technology gets deployed responsibly. The efficiency case for AI-drafted patient communication is strong enough that adoption is not going to stop. The question worth pushing on, loudly, is whether patients get a say in knowing when they are reading one.

What responsible deployment would actually look like

A small number of health systems have begun experimenting with lightweight disclosure: a subtle note indicating a message was drafted with AI assistance and reviewed by the care team, without implying the physician was absent from the process entirely. Early internal feedback from systems trying this approach suggests disclosure does not meaningfully reduce patient satisfaction with the response itself, so long as patients understand a human clinician remained accountable for what was sent. What patients object to is not the assistance. It is discovering it after the fact, without having been told.

Beyond disclosure, the more durable safeguard is workflow design: structuring the review step so it requires active engagement rather than a one-click approval, flagging draft messages that touch on higher-risk categories, new symptoms, medication changes, abnormal results, for a more deliberate review rather than the same rapid-fire treatment as a routine refill confirmation.

The takeaway

AI-drafted patient messaging is not a hypothetical future risk. It is already running quietly inside major health systems today, generating a meaningful share of the words patients read as coming directly from their physician. The efficiency argument for it is legitimate and the burnout it addresses is real. But healthcare has generally treated informed consent as foundational to the doctor-patient relationship, and there is no principled reason that standard should stop applying the moment the conversation moves from an exam room to a portal message. Patients deserve to know when they are talking to a draft. The technology to tell them already exists. What is missing is the policy requiring it.