Every clinic I have spoken to already has some version of an appointment reminder. The problem is almost never whether it sends. It is whether the front desk trusts it enough to stop double checking it themselves.
That distrust is not irrational. A reminder system that gets even a little bit wrong, a wrong time, a message that reads like it is coming from a nurse when it should not be, an emergency that gets swallowed by a bot instead of routed to a person, does more damage in a clinic than in almost any other kind of business. Patients are not customers browsing a shop. They are people who are often anxious, sometimes unwell, and they read every message from a clinic more carefully than they read a marketing email.
The part everyone gets right, and the part everyone skips
Sending a message at the right time to the right number is the easy part. Confirming, rescheduling, sending a pre visit form, chasing a no show, all of that is mechanical once the calendar is connected properly. Where clinics get nervous is the next layer: what happens when the patient replies with something that is not a simple yes or no. "Can I come earlier, I have been in pain since last night." "I need to cancel, my son has a fever." "Is it okay if I take my usual medication before the appointment."
This is the moment a reminder system starts drifting into territory it was never built for. A language model is good at sounding confident, and a confident wrong answer about symptoms or medication is far more dangerous than an honest refusal to answer.
Where I draw the line, and why it is not negotiable
When I build reminder and scheduling automation for a clinic, there are two things I will not let the system do, no matter how capable the underlying model is.
- No automated medical advice, ever. Anything that sounds like a question about symptoms, medication, or what a patient should do about how they feel, goes straight to a doctor or a nurse. The bot does not attempt an answer first and hand off later. It hands off immediately.
- No handling of emergencies. If a message reads as urgent, a human is alerted straight away. The system does not try to collect details, confirm what is wrong, or triage. It escalates and gets out of the way.
A patient cannot tell a confident answer from a correct one, and that is really the whole argument for these two rules. A clinic will forgive a reminder that goes out a few minutes late. It will not forgive a bot that tried to reassure someone describing chest pain instead of getting them to a person.
In my experience, this is what actually earns trust from a practice manager: seeing the system reliably get out of its depth and hand off rather than guess. Once they have watched that happen correctly a couple of times, they stop supervising every message and let it run.
Checking an automated system without rereading everything it does
There is a related problem that shows up anywhere I build something that produces a lot of output automatically, not just in clinics. You cannot manually verify all of it, so you have to be smart about what you check.
I ran into this in a different context, digitising 33 books of handwritten records. Nobody was going to reread all 33 books to check the extracted text was right, so I did not check a random sample. I reviewed the pages the system itself was least confident about, because that is where the real errors live. A random sample mostly tells you the pages that were already fine. I used a side by side editor, the original page image on one side and the extracted text on the other, so corrections took seconds instead of a search. And I still spot checked a few pages in every book, because confidence scores miss things too.
The same logic applies to a reminder system, even though the domain is completely different. You do not need to read every confirmation message a clinic sends. You need to look at the messages the system flagged as unusual, the ones where a patient's reply did not match an expected pattern, the ones near a cancellation or a reschedule, because that is where a bad handoff or a wrong escalation would actually show up. Checking everything is not realistic. Checking the right things is.
What I would ask before turning one of these on
Say a multi-location clinic wants to move from manual phone reminders to something automated across WhatsApp and SMS. Before agreeing that it is ready, I would want clear answers to a short list of questions, not a demo of how nice the messages sound.
- What exactly happens when a reply contains a symptom or a medication question, and can I see that path actually trigger, not just be told it exists
- What counts as an emergency in this system, and who gets alerted, and how fast
- Who reviews the conversations the system was least sure about, and how often
- What happens to a patient who does not reply at all, does someone eventually call them
- Can a member of staff see the full conversation history in one place, not scattered across a phone and a spreadsheet
None of these questions are about the technology being impressive. They are about whether the failure modes have been thought through, because a reminder system does not get judged on the ninety nine messages that go fine. It gets judged on the one that did not, and on whether the practice manager finds out about it from the system, or from a patient complaint.
