Key takeaways
- Inbox work takes about a quarter of family physicians' EHR time.
- In a blinded comparison, evaluators preferred chatbot answers to physicians' answers in 78.6% of cases and rated them far more empathetic, but these were public forum questions, not real clinic patients.
- In real clinics, AI-drafted replies reduced burnout and task load but did not reduce the time doctors spent replying.
- The safest uses today are administrative: booking, reminders, directions and approved instructions, with clinical questions handed to staff.
The message flood
Every clinic knows the feeling: WhatsApp messages, portal messages, missed calls and voicemails, all needing a reply. In one health system, inbox management took about 85 minutes a day per full-time family physician, about 24% of all EHR time[1]. For reception teams, most of the volume is predictable: "Can I book?", "Can I move my appointment?", "Where are you?", "What should I bring?"
Can AI write good answers?
A widely cited 2023 study in JAMA Internal Medicine took 195 real questions from a public online forum and compared answers from verified physicians with answers from an AI chatbot. Blinded healthcare professionals preferred the chatbot's answer in 78.6% of evaluations. They rated 78.5% of chatbot answers good or very good, against 22.1% of physician answers, and rated 45.1% of chatbot answers empathetic, against 4.6%[2].
Two big caveats apply. Physician answers on the forum were much shorter, and the questions came from the public rather than from a clinic's own patients with records and history. The authors themselves called for clinical trials before drawing conclusions[2].
What happened when real clinics tried it
Stanford Health Care (2024). 162 primary-care and gastroenterology clinicians received AI-drafted replies to patient messages inside the EHR. Drafts were used about 20% of the time. Clinicians' task load and work exhaustion scores improved significantly, but the time spent on messages did not change[3].
UC San Diego Health (2024). In a randomised waiting-list study with 52 physicians, AI drafts were associated with a 21.8% increase in time spent reading messages, no significant change in reply time and 17.9% longer replies. Physicians still valued an empathetic starting draft[4].
The pattern is consistent: AI drafting reduces the mental load more than the minutes. Reviewing a draft is work too, especially for clinical questions where every sentence matters.
The safety line: administrative help vs clinical advice
Language models can sound confident when they're wrong. Engineering research, including work in the IEEE Journal of Biomedical and Health Informatics, is focused on making medical dialogue systems more clinically accurate, which shows the problem isn't solved[5]. The World Health Organization's 2024 guidance on large multi-modal models in health calls for transparency, human oversight and accountability whenever these systems are used with patients[6].
A sensible line for clinics today:
| Good fit for AI | Hand to the clinic team |
|---|---|
| Booking, confirming, cancelling and rescheduling | Symptoms, test results and medication questions |
| Reminders and pre-visit instructions approved by the clinic | Anything that sounds urgent: tell the patient to call emergency services |
| Opening hours, location, parking, what to bring | Complaints, billing disputes and sensitive conversations |
| Collecting the reason for a visit to route the request | Anything the AI isn't sure about |
The administrative side alone is valuable. Reminder timing on WhatsApp, for example, has a measurable effect on attendance: a single reminder about 24 hours before the visit cut no-shows by about 24% in a 475,000-appointment analysis[7].
Checklist: using AI with patient messages safely
- Tell patients they're talking to an AI assistant, and always offer a route to a person.
- Restrict the AI to clinic-approved information. It shouldn't improvise medical advice.
- Detect urgent or clinical messages and escalate them to staff immediately.
- Have clinicians review any AI draft that touches care before it is sent.
- Collect only the data needed. Know where messages are stored and for how long.
- Review a sample of conversations regularly to catch errors early.
Where Klinmate fits
Klin, Klinmate's AI front desk on WhatsApp, is being built for exactly the "good fit" column: booking, rescheduling, reminders and answers your clinic approves, with everything else handed to your team. It doesn't give medical advice. Klinmate is launching soon; join the waitlist.
This guide is general information for clinic owners and healthcare professionals. It is not medical, legal or regulatory advice, and it is not a substitute for your own professional judgement.
References
- B. G. Arndt, J. W. Beasley, M. D. Watkinson, J. L. Temte, et al., "Tethered to the EHR: primary care physician workload assessment using EHR event log data and time-motion observations," Ann. Fam. Med., vol. 15, no. 5, pp. 419–426, 2017. doi.org/10.1370/afm.2121
- J. W. Ayers, A. Poliak, M. Dredze, E. C. Leas, et al., "Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum," JAMA Intern. Med., vol. 183, no. 6, pp. 589–596, 2023. doi.org/10.1001/jamainternmed.2023.1838
- P. Garcia, S. P. Ma, S. Shah, M. Smith, et al., "Artificial intelligence–generated draft replies to patient inbox messages," JAMA Netw. Open, vol. 7, no. 3, e243201, 2024. doi.org/10.1001/jamanetworkopen.2024.3201
- M. Tai-Seale, S. L. Baxter, F. Vaida, A. Walker, et al., "AI-generated draft replies integrated into health records and physicians' electronic communication," JAMA Netw. Open, vol. 7, no. 4, e246565, 2024. doi.org/10.1001/jamanetworkopen.2024.6565
- Z. Liu, Y. Quan, X. Lyu, M. J. F. Alenazi, et al., "Enhancing clinical accuracy of medical chatbots with large language models," IEEE J. Biomed. Health Inform., vol. 29, no. 9, pp. 6395–6405, 2025. doi.org/10.1109/JBHI.2024.3470323
- World Health Organization, Ethics and Governance of Artificial Intelligence for Health: Guidance on Large Multi-Modal Models. Geneva: WHO, 2024. www.who.int/publications/i/item/9789240084759
- S. Esteban, G. Quintana, M. Sanchez, A. Rubinstein, and A. Szmulewicz, "Comparative effectiveness of single vs. dual WhatsApp reminders on no-shows: a target trial emulation within the public health system of Buenos Aires, Argentina," medRxiv (preprint, not peer reviewed), 2026. doi.org/10.64898/2026.08.17.26360609
How we write guides: every figure is taken from the cited source, and we say where evidence is weaker (preprints, before-and-after or single-site studies). Spotted an error? Email [email protected] and we'll correct it.