An automated patient communication sequence should carry the facts a patient is already expecting, and stop at the point where a message asks them to feel something or decide something. Confirmations, directions, preparation instructions, medication prompts and follow-up reminders can all be machine sent, and should be. Results, complications, cancellations, price changes and anything that lands on a frightened person belong to a named human with a clock against their name. The design question is not which messages you are able to automate. It is what happens in the four hours after a patient replies to one.
01
The send is the easy half
The evidence on sending is settled and has been for years. The Cochrane Database of Systematic Reviews looked at mobile phone messaging reminders for healthcare appointments and reported attendance of 67.8 percent with no reminder, 78.6 percent with a text reminder and 80.3 percent with a phone call. Text and phone land in roughly the same place. Text costs less. If you run clinics and you are not sending reminders, that is the cheapest fix available to you, and you should stop reading and go and do it.
The interesting part of that review is what it could not report. The authors noted that none of the included studies reported in detail on specific adverse events such as loss of privacy, misinterpretation of a message, or delivery failure. Years of evidence that the send works, and almost nothing on what happens when it goes wrong.
So the failure side of the design is yours to write. Nobody is going to hand it to you, and no supplier will raise it in a demo.
02
The message ladder
This is the sheet I would put in front of anyone configuring a messaging tool on Monday. Four rungs. Every message in your sequence sits on exactly one of them, and the rung decides who owns it.
Rung one. Facts the patient already holds. Date, time, address, parking, what to bring, who they are seeing. Fully automatic, no approval, nobody in the loop. If this rung is wrong you have a data problem, not a communication problem.
Rung two. Instructions that change what the patient does. Fasting, stopping a medication, washing a site, arriving early. Automatic to send, but a named person owns the exception list, and the message itself carries a number that a person answers. A patient who cannot follow a preparation instruction needs a human, not a link.
Rung three. Questions that expect an answer. Post-procedure check-ins, symptom prompts, consent chasers, feedback requests. Automatic to send, human to read, inside a stated window. Here is the rule that saves you: if you cannot name the person who reads the replies and the hour by which they read them, do not send the message at all. An unanswered check-in is worse than no check-in, because you asked.
Rung four. Anything carrying uncertainty. Results, a complication, a cancellation, a change to price or date, a clinical opinion. Written by a person, sent by a person, signed by a person. Automation here buys you very little and costs you the relationship.
The ladder is not a statement about what software can do. Most tools will happily run all four rungs. It is a statement about what your rota can carry.
Every automated message you send is a promise that somebody is listening at the other end of it.
03
Tone is a policy decision, not a copywriting one
There is a finding here that surprised me, and then on reflection did not.
JAMA Network Open published a study in 2025 of just under 1,500 respondents drawn from a health system patient advisory group, comparing replies to electronic messages drafted by a person against replies drafted by artificial intelligence. The machine-drafted replies scored higher across the board: satisfaction, usefulness, and the sense of being cared for. Eighty five percent were satisfied with the machine-drafted replies, against roughly seventy six percent for the human ones.
Then the same study told respondents which was which, and satisfaction fell where the message was labelled as written by AI. The wording most preferred named the clinician and said the message was written with the support of automated tools.
Two things follow. The first is that the quality objection to drafting assistance is weaker than most clinicians assume, because a tired person at the close of a long list does not write with more warmth than a machine does. The second is that disclosure wording is an operating standard, not a settings screen. Write one line, approve it once, and put it in the document every site works from. Leave it to whoever configures the tool and you will get a different answer in every clinic.
One caution on that study, since operators tend to quote findings past their range. Those respondents were older, more educated and self-selected onto an advisory panel. They are not your patient list.
04
Automation creates inbound
This is the part missing from every business case I have read.
JAMA reported this year on messaging trends across a very large body of electronic record data, more than two thousand hospitals and forty seven thousand clinics. Patient-written portal messages rose from about one per patient per year in 2020 to two and a half in 2025. Office visits rose seventeen percent over the same period, and telephone encounters fell slightly. Messaging did not replace contact. It added a layer of it.
Read that as an operator and it says something uncomfortable. Every channel you open stays open. The reply arrives whether or not you planned for it, it arrives at the least convenient hour, and it does not care that the project was signed off on send volume.
I have watched a reminder programme judged for months on how many messages went out and how many were opened, while replies accumulated against a number nobody had been assigned to watch. The tool worked exactly as sold. The operating condition around it had never been built.
So before you switch anything on, estimate the reply rate, staff it, and give the person who staffs it an escalation route with a name and a timeframe on it.
05
What you control and what you do not
You do not control how much patients want to message you. That number is going up everywhere and no policy of yours will bend it. You do not control whether the phone is shared, whether the number is three years stale, whether the message arrives, or whether it is read by the person it was meant for. You do not control what somebody does with a piece of information at eleven at night.
What you control is all of the rest. Which rung each message sits on. Who reads the replies, and by when. The escalation path for the reply that says I am worried. The disclosure line. The opt-out that works the first time it is used. The hygiene of the list you are sending to. Whether the number printed in the message is answered by a person.
None of that is technology. All of it decides whether the automation helps or quietly does harm.
The tool is bought in an afternoon. The listening is a rota, and the rota is the actual product.
Questions people ask
How do you design an automated patient communication sequence?
Sort every message by what it asks of the patient. Facts they already hold can be fully automatic. Instructions that change behaviour need a named owner for exceptions and a working number in the message. Questions expecting an answer need a person reading replies inside a stated window. Anything carrying uncertainty is written and sent by a human.
Which patient messages should never be automated?
Results, complications, cancellations, changes to price or date, and any clinical opinion. Those messages carry uncertainty, and a patient reading them needs a person who can answer the next question. The practical test is simpler than the category list: if a wrong or badly timed message would frighten someone, a named human writes it, sends it and signs it.
Should you tell patients when a message was drafted by AI?
Yes, and decide the wording once at network level rather than per site. JAMA Network Open reported in 2025 that patients rated machine-drafted replies higher than human-written ones on satisfaction and on feeling cared for, but satisfaction fell when the message was labelled as AI written. The phrasing respondents preferred named the clinician and said automated tools had supported the message.