An AI agent in production: what I learned booking clinic appointments over WhatsApp

An AI agent in production serves real people, on real data, and finishes the job. At SuiteMedica that means one agent answers patients over WhatsApp, phone, Instagram, Messenger, email and web chat, and puts the appointment in the clinic’s calendar. Most of the work to get there sits around the model: context, real availability, authorization between clinics, tests and on-call.

By , CTO · Blanco BrandPublished 4 min read

The problem: patients who write when nobody answers

Patients ask for an appointment in a WhatsApp message at eleven at night, in a call nobody picks up, or in an Instagram comment. By the time reception sees it, the patient has often booked somewhere else.

The clinics we work with also had their information scattered: scheduling in one system, clinical records in another, billing in a third, and reminders done by hand. An agent that holds a good conversation is not much use if it cannot see the calendar.

One agent for six channels

The first decision was one agent for every channel, instead of one bot per channel. The patient writes or calls wherever they like and the conversation lands in the same place, with the same context. Channels come in through Twilio and WhatsApp.

With one bot per channel, each ends up with its own rules, its own logs and its own bugs. With one agent, a fix is made once and applies to all six.

Make it actually book

The agent books the appointment. It checks real availability in the clinic’s calendar and prioritises by the patient’s health insurance and condition, as a good receptionist would.

That works because SuiteMedica is a single multi-tenant platform for the clinic’s whole operation: scheduling, clinical records, orders, consents, billing and point of sale, statutory care guarantees (GES), surveys and reminders. The agent works on the same data the clinic’s staff use, with no separate copy to drift out of sync.

My general advice is to start there: before choosing a model, settle where the agent gets the truth about the calendar and how it writes to it.

The model has to be replaceable

The agent lives in its own package in the monorepo, agente-citas, and reaches language models through OpenRouter. That lets us change the model without touching the rest of the system.

Models change in price and quality every few months. If the agent is tied to one provider, every change is a project; if it sits behind your own interface, it is a configuration change and a round of testing.

The delicate part: authorization between clinics

Every clinic is a tenant. An agent serving several clinics must never be able to see another clinic’s calendar or patient data. In a multi-tenant system with AI, that is the risk I worry about most.

That is why continuous integration runs dedicated end-to-end authorization tests, on top of build, type-check, lint and unit tests, before any change is accepted. The project has 100 test files.

Running it: on-call, alerts and monitoring

An agent that books appointments can fail at any hour, like any other production system. SuiteMedica has an on-call rota, alerting and monitoring, and it has been in production and under active development since March 2026, with a team of three.

The agent also depends on its channels, which come in through Twilio and WhatsApp, and the platform has other integrations in production: IMED electronic insurance vouchers and Haulmer/TUU payment terminals. If an integration goes down, the agent cannot do its job. In a system like this, the alert has to arrive before the patient’s complaint does.

What I do not publish

This case does not publish evaluation, latency or cost metrics for the agent, or usage figures per clinic. I would rather say so than put numbers in public that I cannot back up.

What I can describe is the approach. In the agent on this site, which is much smaller, the model answers only from what is published, cites the source, and an automatic check reviews every answer before it is shown. If it does not cite a recognisable source, it is not shown. Any agent that talks to people needs that kind of check after the model.

Other uses of AI at the same clinics

Booking is the most visible case, but not the only one. In MED360 Presupuestos, an OpenAI model writes weekly summaries of the surgical quotes that operators at Clínica CBB follow up, and patient campaigns go out over WhatsApp. At CCDM, omnichannel campaigns and the WhatsApp inbox for referrals are in production for Clínica Ciudad del Mar.

The same rule applies to all of them: AI goes in when it does a specific task a person does today. If an agent needs to read data from a system that cannot be modified, the approach is in AI on top of a legacy ERP.

If you are putting an agent into production in healthcare and want to compare approaches, write to me at hola@bcristi.com.

Let’s talk. If you want to discuss AI, healthcare, fintech or research, propose a collaboration or invite me to give a talk, write to me.

  • Asturias, Spain
  • CET hours
  • Spanish native, English C1

Also on WhatsApp · GitHub