---
title: "AI at Arbol, in plain words"
description: "What an AI agent is and isn't, what it's made of, why it can be wrong, how Arbol limits that and what gets recorded, explained for nurses, doctors and the whole team."
---

> Documentation Index
> Fetch the complete documentation index at: https://docs.getarbol.com/llms.txt
> Use this file to discover all available pages before exploring further.

# AI at Arbol, in plain words

Arbol uses artificial intelligence (AI) to look after your patients and to help your team. You don't need to know about technology to work well with it: it's enough to understand what it does, where what it says comes from, and when it needs a person.

## What an AI agent is

An **agent** is a program that holds conversations: it understands what someone writes or says and answers in its own words, the way someone at the front desk would. At Arbol, agents live in [Raíz](/en/raiz/) (in the menu, **Workers**) and look after your patients by phone, WhatsApp or your website's chat, on behalf of your organization.

| An agent does… | An agent doesn't… |
| --- | --- |
| Follow the instructions and tone your team gives it. | Replace a person or a health professional. |
| Check your organization's documents before answering. | Know what it wasn't taught: beyond its instructions and documents it only knows general things, which may be outdated or wrong. |
| Carry out tasks in connected systems, such as finding slots or booking an appointment. | Decide for your team: diagnosing, prescribing a treatment or making an exception is a person's job. |
| Hand the conversation to your team when a person is needed. | See more than your organization connected it to. |
| Keep a record of every conversation. | Learn from its mistakes on its own: it improves when your team corrects its instructions or knowledge. |

**Arbi** is a different AI: your team's assistant inside the applications. You ask it in your own words for a report, a list or a search. It works with your own permissions; in Raíz, Agenda and Account it asks you to approve before it makes a change (in Ámbar it saves notes, tasks and drafts without asking; see [Approve what Arbi proposes](/en/arbi/approvals/)) and, in Raíz and Agenda, shows this notice below where you type: "Arbi can use your connected apps and can be wrong. Check anything that matters."

## What an agent is made of

An agent is built from three pieces, and your organization chooses all three:

- **Instructions**, also called the *prompt*. It's the text where your team tells it who it is, what it must achieve, how to speak, what it must not do and when to hand the case to a person. It's like the job description for someone new in the role. To write them well, read [What prompting is](/en/raiz/prompting/) and [Write an agent's instructions](/en/raiz/write-instructions/).
- **Knowledge.** The documents it checks before answering: services, hours, locations, test preparation. If a document is outdated, the agent answers with the old information. For something temporary, such as a holiday, you use **bulletins**. See [Give your agents knowledge](/en/raiz/knowledge/).
- **Tools.** What it can do in the systems connected to your organization: look something up, such as a service's free slots, or change something, such as booking an appointment. They come from **integrations**, which Arbol's team connects. See [See the integrations](/en/account/integrations/).

## Why it can be wrong

AI writes the answer that seems most likely to it. It's almost always right, but not always:

- **It makes things up confidently.** If it's missing a fact, it may fill the gap with something that sounds right, such as a price or a time that doesn't exist. This is called a *hallucination*. That's why its instructions should tell it, for example, "Never invent prices, timelines or availability".
- **It uses old information.** If a document wasn't updated, it answers with what the document says.
- **It misunderstands.** A noisy audio message, an ambiguous sentence or a message with two requests can confuse it.
- **A tool fails.** If the connected system doesn't respond, the agent can't look things up or book. In the conversation, the tool shows as **Failed**.

## How Arbol limits it

- **Your instructions set the limits.** What the agent doesn't do and when it hands the case to a person is written by your team, and can be changed whenever needed.
- **A person can always take over.** When someone asks to talk to a person, the agent hands the conversation to your team, even if its instructions don't say so.
- **People decide.** In Ámbar, Ámbar proposes and a person decides: nobody enters a cohort, no care is authorized and no clinical alert is closed without a decision from someone on the team. In Raíz, Agenda and Account, Arbi asks you to approve each change with **Approve** or **Deny**.
- **Missing data isn't guessed.** In Ámbar, when information is missing, it shows as **Sin evaluar** (not evaluated) or **Por verificar** (to be verified), never as a "no".
- **Changes to an agent are tested first.** When your team corrects a reply, Raíz tests the improvement against the agent's earlier conversations and applies it only if someone approves it. See [Improve an agent's replies](/en/raiz/improve-replies/).
- **Everything can be reviewed.** Every conversation is kept for your team to read, as explained below.

## What escalating means

**Escalating** is when the agent hands the conversation to your team because it can't or shouldn't resolve it alone. It isn't a failure: it's the agent doing the right thing. The conversation shows **Conversation handed to the team**, with one of these reasons:

- "The contact asked for a person. AI handed over the conversation."
- "AI requested help from the team to continue with this request.", for example because the case is one its instructions say must go to a person.
- "The team continues because the agent reached its usage limit."

On WhatsApp, when the agent escalates, the AI stops replying in that conversation until someone on your team turns it back on. You find it in **Conversations** with the **Escalated by the AI** filter, and your team gets a notice if it turned notices on. See [Handle a WhatsApp conversation](/en/raiz/take-over-a-conversation/).

## An AI conversation, step by step

By WhatsApp, phone or your website's chat.

Its instructions and its knowledge.

For example, it finds slots or books the appointment.

If a person is needed, it hands the case to your team.

Transcript, summary and outcome, ready to review.

## What gets recorded

- **In Raíz**, every conversation keeps the messages and who wrote each one (the agent, the person or your **Team**), the call recording if the agent records, and the tools it used with their result. When it closes, Raíz adds a summary and an outcome: **Success**, **Failure** or **Unknown**. Anyone with a role that can see conversations reviews them in **Conversations**. See [Follow conversations](/en/raiz/conversations/).
- **The feedback** your team leaves on an agent's replies stays on its page, under **Team improvements**.
- **In Ámbar**, every decision is kept with its evidence and its date.
- **With Arbi**, your conversations stay in its **History**, and Arbi only saves notes in its **Memory** when you ask it to.

A conversation's summary and outcome are made by AI too: they help you review quickly, but they aren't the record. If something matters, read the transcript.

## Keep going

- **What prompting is** — How to write good instructions for an agent.
- **Safe use and patients' data** — The person decides, what never to write and how to report a wrong answer.
- **Raíz** — The agents, their conversations and how to improve them.
- **Ámbar** — Governance: Ámbar proposes; a person decides.

Source: https://docs.getarbol.com/en/get-started/ai-basics/index.mdx