# What an AI agent can and cannot do for a clinic

> Most of the pitches you hear are either magic or doom. The useful answer is in the middle, and it is specific.

Author: the Velosyd team
Published: 2026-10-29
URL: https://velosyd.com/articles/what-an-ai-agent-can-and-cannot-do-for-a-clinic

If you run a clinic, you have probably had two kinds of conversations about AI this year. In one, somebody promised that an agent would replace your front desk. In the other, somebody warned that an agent would tell a patient something dangerous. Both conversations skip the question that matters: which specific jobs is an agent good at today, and which ones should stay with a person?

## What it does well

An agent is excellent at the conversations that are frequent, predictable and interruptive. In a clinic, that is most of the phone.

- **Answering the phone after hours.** A missed call at 7 p.m. is usually a patient who books somewhere else. An agent answers, checks the calendar and books.
- **Confirmations and reminders.** The day-before call that nobody on your team enjoys making. An agent makes it, in the patient's language, and updates the appointment when they reschedule.
- **Repetitive questions.** Opening hours, parking, what to bring, whether you accept a given insurance. Answered from your own documents, the same way every time.
- **Writing things down.** Every conversation ends up as a note on the patient's record. Nobody has to transcribe anything.

## What it should not do

An agent should not give clinical advice, triage symptoms or decide what is urgent. Not because the technology is incapable of forming a sentence about it, but because the cost of a wrong sentence is too high and the responsibility is yours. A well-built agent recognises those moments and hands the conversation to a person, with the context attached, instead of guessing.

It also should not pretend to be human. Patients forgive an honest assistant that says what it is. They do not forgive a fake receptionist.

> The question is not whether an agent can talk. It is whether it knows when to stop talking and get a person.

## How to tell a good one from a bad one

Ask three questions before you buy or build anything.

1. **Where does it get its answers?** The right answer is "from your documents and your systems, and nowhere else". If it can improvise, it will.
2. **How was it tested?** You want to hear about a set of real conversations it has to pass before launch and after every change, not a demo that went well once.
3. **What happens when it does not know?** You want a hand-off to a person, with the transcript, quickly. Anything else is a liability.

## Where to start

Start with the single conversation that costs you the most. For most clinics it is the missed call. Put an agent on that one job, measure how many appointments it books in a month, and decide the next step with that number in hand. One job done well beats five jobs done halfway.

[Let's talk](https://velosyd.com/contact)