A practice manager's first reaction to an AI assistant is rarely excitement. It is a specific fear: that the thing will answer a medical question at eleven at night and somebody will act on it.
That fear is correct, and it is also the wrong reason to say no. Look at what a clinic actually gets asked out of hours and most of it is not clinical at all. It is hours, parking, insurance, preparation, cancellation. The question is not whether to automate the inbox. It is where the line falls inside it.
Four messages, one line
The line is not about difficulty. It is about whether the answer exists in writing and applies to everybody, or whether it depends on one person's body and one person's file.
The pile that is safe to answer
This is longer than most practice managers expect, and it is where the whole return lives.
- Opening hours, including holidays, and which site a clinician works from on which day
- Address, parking, step-free access, which entrance to use
- Which insurers and plans are accepted, and what needs a referral
- How to book, reschedule or cancel, and the notice your policy requires
- Preparation instructions: fasting, medication to pause, what to bring
- Self-pay prices for the services you publish them for
- What to expect at a first appointment, and how long it takes
None of that is medical advice, all of it is written down somewhere already, and all of it is asked at the hours the front desk is closed.
The pile that must never be answered
Four categories, and they should be refused explicitly rather than handled carefully.
The temptation, once the assistant says it cannot see results, is to give it the records. Everything changes at that moment: what a wrong answer costs, what a breach means, and which agreements you need in place. The safer design is the one where no patient data is involved at all, and it covers the questions in the first pile completely.
Refusing well is a skill you configure
A refusal is not a failure, and it is not a dead end either. A good one does three things in two sentences: it says what it cannot do, it says what happens instead, and for anything urgent it points at emergency care before anything else.
Write those words yourself rather than leaving them to a model. The instructions for a clinic assistant should name the refusals explicitly: no symptoms, no results, no medication, no opinions about treatment, and an emergency line that appears whenever somebody describes something acute.
Before it goes live, ask it the ten worst questions you can think of, the ones a worried patient asks at midnight. What it says to those is the only acceptance test that matters, and it takes twenty minutes.
What the clinic gets back
Two things, and the second is the one that surprises people.
The first is the obvious one: the questions that used to become voicemails get answered, in the language the patient wrote in, at the hour they asked.
The second is a list of what your website does not explain. Every question the assistant could not answer is logged, and in a clinic that list is remarkably consistent: parking, whether a relative can come in, what happens if you are late, what a first appointment costs. A month of that list is a better content plan than any agency brief.
welcomeai.devAI chatbot for clinics and practicesQuestions people ask
Start from the pile that is already written down, publish the refusals as deliberately as the answers, and read the unanswered list at the end of the first month. That order keeps the risk where a clinic can actually manage it.
welcomeai.devHow to stop your AI chatbot from hallucinating
