SarAI — what it does, what it doesn't, and what's on you

SarAI is an artificial-intelligence assistant. It answers your patients on the channels you connect, offers them appointment times, and books, moves or cancels appointments.

This guide is what you need in order to supervise it — because from 22 October 2026 the Terms say that supervision is yours.

What it does on its own

• Says it is an AI in its first message of every conversation, says it can be wrong, and says the patient can ask for a person.
• Answers only from your published catalogue. It does not invent services, prices or preparation instructions. If it does not have the fact, it says it will confirm.
• Asks for the data authorization (Ley 1581) before registering anyone.
• Hands the conversation to your team the moment someone asks for a person.

What it does not do

It does not give medical advice, diagnose, or interpret results. It does not decide who receives care or at what price. And it does not keep answering after someone has asked to speak to a human.

What can actually go wrong

1. It can be wrong. Its answers come from a language model. It sounds equally confident when it is right and when it is not.
2. It only knows what you publish. An out-of-date price in your catalogue is an out-of-date answer to a patient. Most wrong answers start here, not in the model.
3. The times it offers are a sample, not the whole day. It says so, but a patient may read a short list as the only options.
4. A WhatsApp template message cannot carry the AI notice. Meta approves those templates and rejects altered text, so in those threads the notice arrives in the next message.

What is on you

• Read it. Conversations are in the live inbox and your team can take over any of them at any moment.
• Verify before you rely on it. What it drafts for you is a draft. Content badged AI-generated has not been reviewed by a person yet.
• Keep your catalogue current. It is the most common cause of a bad answer.
• Decide how far it goes. You can switch the AI off for one conversation, or not connect a channel at all.

What we record

For every turn we record the model, the usage, and the prompt version that produced the answer. That is what makes it possible to reconstruct a conversation if a patient or a regulator asks what happened.

We do not train models on your data or your patients’ data, and our model providers are contractually barred from doing so.

If something goes wrong

Take over the conversation first — that stops the assistant immediately. Then write to contacto@deusrex.io with the conversation: we can retrieve exactly what was sent and which prompt produced it.