Patient portals are now mainstream — about 60% of patients accessed them in 2022 vs 25% in 2014. The downside: primary care teams are getting 5x more portal messages than surgeons, with no extra staffing model for 24/7 access. That workload is linked to burnout.
Like many systems, UCHealth (University of Colorado Health) tried to fight inbox overload with AI.
What UCHealth did
From Sept 2023 to March 2024, UCHealth ran a quality improvement study across 9 clinics (6 primary care, 3 specialty).
They embedded GPT-4 in Epic, dubbed PAM Chat (Patient Advice Message Chatbot reply). For every incoming patient portal message, the LLM drafted a reply for a nurse, medical assistant (MA), or clinician (MD / APP) to edit.
21,323 drafts generated
2,596 used = 12% usage rate
166 users: 12 RNs, 14 MAs, 93 clinicians
Iterative prompt engineering: adding the Assessment & Plan from the last progress note and emphasizing conciseness made drafts more useful
Transparency: patients were told content was computer-generated before clinician editing
Who liked it?
Results diverged sharply by role.
Nurses were the fans: 92% said PAM Chat helped improve efficiency, empathy and tone. They were significantly more likely to say it reduced need to forward to MDs and helped them stay within scope of practice.
MAs, physicians, and APPs were less favorable.
Why MDs didn't like it
1. They get the hardest messages.
In team-based triage, MAs and RNs filter out routine stuff first. By the time a message reaches an MD/APP, it's complex, ambiguous, off-protocol — exactly what LLMs are worst at. The authors note physicians "may preferentially receive complex messages, which are more difficult for the LLM."
2. The drafts created more work, not less.
12% overall use means 88% of the time, users ignored the draft. For MDs, editing a long, generic, slightly-wrong draft was slower than writing a short, precise answer themselves.
3. Wrong context and nonsensical actions.
The editorial calls out instances where PAM Chat gave a response "more suited for a specialty clinic" for a primary care question, or "recommended a nonsensical action" requiring prompt rewrites.
4. Hallucinations and scope creep.
For higher complexity queries, drafts could be detrimental if they included hallucinations or suggestions "that go beyond their appropriate scope of knowledge." MD work is precisely those higher complexity queries.
5. No value for staying in scope.
This is the inverse of why nurses liked it. Nurses agreed it helped them handle more messages without forwarding to MDs and stay within scope. For MDs, there is no one to forward to — they are the scope. That benefit disappears.
6. Empathy ≠ clinical value for MDs.
92% of nurses valued improved empathy and tone — useful for 50 routine replies a day. For an MD answering chest pain, medication risk, or abnormal labs, empathy is table stakes. What they need is clinical precision and nuanced decision-making. GPT-4 didn't deliver that, even after adding the last A/P note.
Result: lowest Net Promoter Score, least likely to say it improved efficiency, least likely to recommend.
The negative cases - where AI failed
1. Mixed real-world feedback:
"Sometimes the answers are astoundingly helpful, other times they are not helpful at all."
2. Wrong clinical context: Drafts that sounded like a specialty answer for primary care.
3. Nonsensical or out-of-scope actions: Requiring post-hoc prompt fixes.
4. Hallucinations and scope creep: Risk of suggesting management beyond appropriate knowledge.
5. Low adoption as signal: 88% ignored means not a time saver for most.
Takeaway for health systems
Include recent A/P from last note — it helped at UCHealth.
Keep a human in the loop. As the editorial notes, it "may be some time — or perhaps it will never be time — for practices to reply directly to patients without any human in the loop."
AI drafts can add empathy and speed for routine work, but they are a supportive aid, not a panacea for the inbox crisis.
References:
Primary study:
https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2824738
Commentary / editorial:
https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2824739