What the assistant must refuse to assess, and how it refuses
Patients describe symptoms to whatever will listen, and a website assistant listens at three in the morning when nothing else does. That is exactly when the usual product instinct - be helpful, give the best answer you can - becomes the wrong instinct. An assistant that offers any assessment of what a symptom means has practised medicine, and the fact that it was probably right is not a defence. The boundary is not a tone setting; it is the single configuration decision that matters most on a clinical site.
The line is between describing the service and assessing the person
An assistant may state what a clinic treats, how long an appointment takes, what a procedure involves, what preparation is required and when the next slot is. It must not say what a described symptom indicates, whether it is urgent, whether it is probably nothing, or which specialist is appropriate for it. The first set is information about the clinic; the second is a clinical judgement about an individual. Write the line down in those terms before evaluating any vendor, because every vendor will claim to respect a line you have not defined.
"Probably nothing" is the failure that will not look like one
The dangerous output is not an alarming answer - alarming answers get escalated and reviewed. It is reassurance. A patient who describes chest tightness and is told it is commonly muscular may not call anyone, and nothing in the logs will record that as an incident. Ask any vendor what their assistant does with a symptom description and listen specifically for whether it ever reassures. If it can reassure, it can reassure wrongly, and that failure is silent.
Emergency wording has to be fixed text, not generated
For the small set of descriptions that must trigger an emergency response, the response cannot be composed fresh each time. It has to be the clinic's own wording, reviewed by someone clinical, emitted verbatim, with the local emergency number for the patient's region rather than a generic one. A generated emergency message is a message nobody approved, which is the opposite of what an emergency path is for. Require the vendor to show you where that text is stored and who can change it.
Refusing well is a skill, and most refusals are bad
"I cannot help with that" loses the patient. A good refusal does three things in one breath: declines the assessment explicitly, says who can make it, and offers the next concrete step - the earliest appointment, the direct number, the out-of-hours route. The difference between a refusal that routes a patient and one that abandons them is entirely in the wording, so make the wording part of what you evaluate rather than something configured later.
Frequently asked questions
Can a website assistant triage symptoms?
What is the most dangerous answer it can give?
Should it have an emergency response at all?
How do we test that the boundary holds?
Does a strict boundary make it useless to patients?
Get in touch
Leave your details and we will reply. You do not need a sales call to get an answer.
Other questions
-
Can the patient actually get in, and who answers that at ten at night
Read the answer -
What happens to patient enquiries when the practice is closed
Read the answer -
A complaint that arrives in a chat window is already a complaint
Read the answer -
A course of treatment is one decision and a dozen appointments
Read the answer -
"Will my insurance cover this?" - the question asked before every booking
Read the answer -
The patient who cannot describe the problem in your language
Read the answer -
What a new patient has to bring, and why nobody can ever find the list
Read the answer -
Can a clinic let an AI assistant near patient data?
Read the answer -
The patient who should have come back and did not
Read the answer -
Repeat prescriptions: the request that must never be answered automatically
Read the answer -
Why patients miss appointments, and which of those reasons you can remove
Read the answer -
Referral requests, and the week they usually spend waiting
Read the answer -
"Are my results back yet?" is a different question from "what do they say?"
Read the answer