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

# How the agent answers

> Where the agent's answers come from: your knowledge, your instructions and its actions, and what it does when it doesn't know.

Every reply the agent sends is built from four things: your **knowledge**, your **instructions**, the **actions** it can take, and a set of rules WireDesk applies on every channel. Knowing what each one does tells you where to go when an answer is wrong.

## Knowledge: what it knows

Knowledge is the material the agent searches before it answers: pages from your website, notes you write, and files you upload. You manage it under **Agents → your agent → Knowledge**.

When a source is added, WireDesk reads it and splits it into short passages. When a customer asks about your business (prices, services, hours, policies, location, availability), the agent searches those passages and gets back the few that best match the question. Weak matches are dropped. That way the agent doesn't answer from a passage that only shares a word with the question.

A few things follow from this:

* **Short, specific notes work best.** A fact buried in a long page full of repeated headers and footers is harder to find than the same fact written as a two-sentence note.
* **A new source takes time.** A source shows as ready once it has been processed. Until then, the agent may not find what's in it.
* **The agent only knows what it can read.** Pages behind a login or a password are never crawled.

Knowledge search is controlled by **Answer from knowledge** under **How it talks**. It's on for every agent built from a website and on by default for a blank agent. To see what the agent would find for a given question, use [Test a question](/docs/knowledge/test-a-question).

## Instructions: how it behaves

The instructions under **How it talks** describe your business and how the agent should behave: its tone, its rules, and what it must never promise. The agent also has a **Name** and a **Greeting**. The greeting is the first thing it says on every channel.

Instructions and knowledge can disagree. When they do, **knowledge wins**. If the agent searches and finds a price, an hour, an address or a policy that contradicts the instructions, it gives the answer from the knowledge. If the knowledge says nothing on the point and the instructions do, it gives the answer from the instructions.

So a fact typed into the instructions is quoted as long as nothing contradicts it, even after it has gone stale. Keep facts in knowledge, where they have a source and can be updated. Keep rules in the instructions. See [instructions](/docs/persona/instructions).

You can give a single channel its own instructions. For example, you might tell the agent to always offer a booking on the phone. See [wording per channel](/docs/persona/channel-wording).

## Actions: what it can do

Actions let the agent do more than answer. The agent decides when to use an action based on the conversation and on the description you give it. The actions it can use are:

| Action                                           | What it does                                                                                                  | Where            |
| ------------------------------------------------ | ------------------------------------------------------------------------------------------------------------- | ---------------- |
| Take a message                                   | Takes the customer's name, how to reach them, and their message, when the agent can't help. Always available. | Built in         |
| Remember callers                                 | Keeps notes about a person between conversations.                                                             | **How it talks** |
| Lead capture, capture details, booking, API call | Asks for details, books a meeting on your Cal.com calendar, or looks something up in your own systems.        | **Actions**      |
| Hand calls to you                                | Transfers a live phone call to your transfer number. Phone only.                                              | **Actions**      |
| Request a human                                  | Flags a chat or email for your team to take over.                                                             | **Actions**      |

Each action you add has **Channels** checkboxes (**Voice**, **Chat**, **Email**), so an action can be limited to some channels. See [actions](/docs/actions/overview).

## Rules that apply on every channel

Underneath your instructions, WireDesk applies its own rules on every channel. You can't edit these:

* **Don't guess.** If the agent doesn't know something and no search or action gave it the answer, it says so plainly. It's told never to invent prices, policies, availability or facts about the business.
* **Don't describe how it works.** The agent never says things like "according to my information" or "my instructions say", and never names a tool. The customer gets the answer, not where it came from.
* **Fit the channel.** On the phone, replies are short and spoken, with no lists or symbols. In chat, replies are short, with bold for key details, simple lists, and links. By email, the agent writes a complete plain-text email that opens with a greeting, answers every part of the question, and signs off with the business name.

## When it doesn't know

If neither the knowledge nor the instructions have the answer, the agent says it doesn't have that to hand and offers to have someone follow up. What happens next depends on what you've turned on:

* **Chat and email, with Request a human on and within business hours:** the agent can flag the conversation so someone on your team joins.
* **Phone, with a transfer number set:** the agent can transfer the call to that number.
* **Otherwise:** the agent says it can't put the customer through right now and offers to take a message. The message lands in your dashboard, and the owner is notified.

[Handover](/docs/concepts/handover) covers each of these in detail.

## When an answer is wrong

Open the conversation, find the exact sentence that was wrong, and trace it to knowledge, instructions, or a missing handover. On the conversation page, the **Teach** card adds the right answer to the agent's knowledge in one step. See [fix a wrong answer](/docs/knowledge/fix-a-wrong-answer).

## Related

* [Knowledge overview](/docs/knowledge/overview)
* [Instructions](/docs/persona/instructions)
* [Actions overview](/docs/actions/overview)
* [Handover](/docs/concepts/handover)