What problem does it solve?
Patients now message their care team through the electronic health record for almost anything: a medication question, a side effect, a result they do not understand, a request to refill a prescription. NYU Langone's chief medical information officer has described a more than 30 percent annual increase in the number of these messages in recent years, with some physicians receiving more than 150 a day. UC San Diego Health's physicians receive about 200 messages a week. Answering them well takes time and judgment, and composing each reply from a blank box adds to that time on top of deciding what to say. The result lands after hours, on top of a full clinical day, and is widely cited as a driver of clinician burnout.
- Physicians saw a more than 30 percent annual increase in recent years in the number of In Basket messages received daily, with some physicians receiving more than 150 messages a day.AI Tool Successfully Responds to Patient Questions in Electronic Health Record (2024)
How does it work?
- Read the message in context. The system reads the patient's new message together with the thread history, so the draft answers what was actually asked rather than a generic version of it.
- Ground the draft in the chart. It pulls the patient's active problem list, medications and recent notes, so a reply about a symptom or a refill reflects that patient's own record.
- Draft within seconds. A draft reply appears directly in the clinician's normal inbox and reply box, written like a message a clinician would send, not a generic template.
- A human reviews every draft. The clinician, nurse, advanced practice clinician or pharmacist who owns the message reads the draft, edits it as needed, and only then sends it.
- The patient is told, at least in some deployments. In UC San Diego Health's pilot, the sent message discloses that it was drafted with AI assistance and reviewed by the clinician who signed it.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools, Email
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Response time reduction | Too few to pool | about 7% | 1 | 1 organization |
Value drivers: Employee productivity, Customer experience.
Indicative value
A primary care group with 150 physicians and advanced practice clinicians
USD 2600 to USD 312,000
Annual clerical time cost avoided from faster drafting on used messages per year
How this is calculated
Formula: clinicians * messagesPerClinicianPerWeek * 52 * usageShare * secondsSavedPerUsedDraft / 3600 * costPerClinicianHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Physicians and advanced practice clinicians clinicians, clinicians | 150 | 150 | The reference group. |
| Patient portal messages received per clinician per week messagesPerClinicianPerWeek, messages per clinician per week | 150 | 200 | UC San Diego Health reports its physicians receive about 200 messages a week; used as the upper bound, with 150 a week as a more conservative lower bound. Source |
| Share of shown drafts a clinician starts from usageShare, fraction of messages with a draft shown | 0.1 | 0.2 | NYU Langone's ten month usage study found clinicians chose to start with the AI draft in 19.4% of cases where one was shown. Source |
| Seconds of reply time saved when a draft is used secondsSavedPerUsedDraft, seconds per message | 1 | 24 | NYU Langone measured a median reply time of 331 seconds with a used draft versus 355 seconds writing from scratch, a 24 second difference, used as the high end; Stanford Health Care and UC San Diego Health both measured no time saving at all in their own pilots, so the low end is set just above zero rather than at NYU Langone's own figure. Source |
| Fully loaded clinician cost per hour costPerClinicianHour, USD per hour | 80 | 150 | Editorial assumption for a US primary care physician or advanced practice clinician; replace with your own fully loaded cost. |
What it leaves out: Counts only the seconds NYU Langone measured saved on messages where a clinician actually used the draft, and assumes, for simplicity, that a draft is shown for every message received. It leaves out the value of higher message quality and empathy, any extra time spent reviewing or editing drafts that were not used, and the fact that two of the three published pilots cited here found no net time saved at all; some organizations may see zero from this line and gain only the reported reduction in cognitive burden.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
NYU Langone Health
United States · Healthcare · 2024
NYU Langone Health licensed a private instance of GPT-4 in 2023, which researchers used to generate draft replies to patient In Basket messages for a blinded rating study: 16 primary care physicians compared 344 randomly paired AI and human replies, finding no statistical difference in accuracy, completeness or relevance, but AI replies were rated more than twice as likely (125 percent more likely) to be considered empathetic (JAMA Network Open, July 2024). A separate, later ten month study of more than 55,000 messages sent through NYU Langone's patient portal, built on an embedded generative AI drafting tool, found clinicians started from the AI draft in 19.4 percent of cases where one was shown, with adoption rising modestly as the system's prompting improved, and a median reply time of 331 seconds when a draft was used against 355 seconds writing from scratch, though this saving was often offset by time spent reviewing, editing or ignoring drafts (npj Digital Medicine).
- Response time reduction: about 7%, October 2023 to August 2024, messages where a draft was used
"Using a draft shaved roughly 7 percent off response times, a median of 331 seconds versus 355 seconds when drafting from scratch"
Claimed by: organization
Stanford Health Care
United States · Healthcare · 2023
Stanford Health Care piloted AI generated draft replies to patient portal messages for five weeks in July and August 2023, covering 162 primary care and gastroenterology clinicians, nurses, advanced practice clinicians and pharmacists. Drafts appeared in the clinician's electronic health record inbox for review and editing before being sent. A post pilot survey found clinicians reported less cognitive burden and fewer feelings of work exhaustion, although the drafts did not measurably reduce response time. Stanford Health Care said it planned to expand the tool to more clinicians.
No outcome disclosed.
UC San Diego Health
United States · Healthcare · 2023
UC San Diego Health ran what it describes as the first randomized prospective evaluation of AI drafted physician messaging, using generative AI built into its Epic electronic health record to draft replies to non emergency patient portal messages from a pilot initiated in April 2023. The study, published in JAMA Network Open in April 2024, found AI generated replies did not reduce physician response time, but produced longer, more empathetic drafts that physicians could edit rather than write from scratch, which lowered their cognitive burden. Every AI assisted reply discloses to the patient that it was automatically generated before a physician reviewed and edited it.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- The patient's active problem list, medications, allergies and recent visit notes
- The full message thread, not just the latest message, so the draft answers what was asked
- A defined list of message types the draft may be shown for, excluding anything that looks urgent
Systems to integrate
- Electronic health record inbox, such as Epic In Basket or an equivalent
- Identity and access control so a draft only ever reaches the clinician who owns the message
Complexity: Medium
Drafting the reply itself is a single model call once the message and the patient's context are available. The real work is safe integration with the electronic health record's inbox, pulling only the chart data needed for that message, and building a review workflow that never sends without a named clinician's edit and approval.
- 1
Start with a narrow set of message types
Launch on a small set of low risk, high volume message types, such as medication refill questions or simple symptom questions, and exclude anything that reads as urgent.
- 2
Ground the draft in the chart, not just the message
Retrieve the patient's problem list, active medications and recent notes so the draft reflects their actual situation instead of a generic answer to the words in the message.
- 3
Make editing effortless
Put the draft directly in the clinician's normal reply box, editable like any other text, so using it is never slower than writing from scratch.
- 4
Route anything urgent around the draft
Detect red flag language, such as severe symptoms or safety concerns, and send those messages straight to a human without a draft or with an escalation flag.
- 5
Disclose the AI assistance to the patient
Tell the patient, inside the sent message, that the reply was drafted with AI assistance and reviewed by their clinician, as UC San Diego Health does.
Guardrails
- No message reaches a patient without a named clinician reading, editing where needed and actively approving it
- Urgent or safety related language is detected and routed to a human before any draft is generated
- The draft is grounded in the patient's own record, not left to the model's general knowledge
KPIs to instrument
- Share of shown drafts the care team starts from, by message type and by role
- Time from message receipt to reply sent, with and without a draft
- Reading grade level and length of sent replies
- Clinician reported cognitive burden and burnout, by survey
Human in the loop
A clinician, nurse, advanced practice clinician or pharmacist reviews and can edit every draft before it reaches the patient. None of the published pilots described here allow a draft to be sent automatically.
Common failure modes
- Accurate but not what the patient needed
- The draft answers the literal question but misses the underlying worry. Require the clinician to read the original message, not only the draft, before sending.
- Draft fatigue
- Generating a draft for every message regardless of fit creates clutter and adds a cognitive cost to reviewing a constant stream of AI output, much of which may be irrelevant, as NYU Langone's own study argues. Generate drafts selectively for the message types where they are known to help, rather than for every message.
- Reading level creep
- AI drafts have tested at a higher reading grade level and greater length than clinician written replies. Set a target reading grade level and flag drafts that exceed it before they are shown.
What are the risks and rules?
EU AI Act
Depends on design
The patient never interacts with the AI system directly, only with the clinician who reviews, edits and sends the reply, so Article 50(1), which covers systems intended to interact directly with natural persons, does not apply here; UC San Diego Health's patient facing disclosure is a voluntary practice, not an Article 50(1) duty. A copilot that only drafts for a clinician to review is not listed in Annex III and is usually minimal or limited risk, although the provider can still owe the Article 50(2) duty to mark the drafted text as AI generated, unless the exception for an assistive function with standard editing where the deploying clinician retains final responsibility applies. It would move toward high risk under Article 6(1) and Annex I only if the software qualifies as a medical device under the EU Medical Device Regulation and needs a notified body assessment, for example because it suggests a diagnosis or treatment rather than only drafting correspondence.
Controls to put in place
- Every reply requires a named clinician's review and approval before it reaches the patient
- Automatic routing of urgent or safety related language to a human without a draft
- AI disclosure inside every message the patient receives
- Draft generation limited to the message's own care team
Frequently asked questions
- Does an AI drafted reply save clinicians time?
- Not necessarily. Stanford Health Care found the drafts did not save clinicians' time, and UC San Diego Health found AI generated replies did not reduce physician response time, though both reported less cognitive burden. NYU Langone measured a modest saving, a median of 331 seconds with a used draft against 355 seconds writing from scratch, and only on the messages where clinicians actually chose to use it.
- How often do clinicians actually use the AI draft?
- In NYU Langone's ten month study of more than 55,000 messages, providers chose to start from the AI draft in 19.4% of cases where one was shown, and adoption rose modestly as the system's prompting improved.
- Are AI drafted replies safe for every kind of message?
- UC San Diego Health limited its AI drafts to non emergency patient questions. All three published pilots required a clinician, nurse, advanced practice clinician or pharmacist to review and edit every draft before it was sent, and none let the AI send a reply on its own.
- Do patients know a reply was AI assisted?
- In UC San Diego Health's pilot, every AI assisted reply included a notice that it was automatically generated before being reviewed and edited by the physician who signed it.
How to cite this page
Blits.ai AI Use Case Library, "AI drafting of clinician replies to patient portal messages", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/patient-portal-message-reply-drafting. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 29 September 2026: First published