# Create an AI Agent

> Create a Raven AI agent in Revora: write its instructions, train it on your website and documents, set actions, test it and deploy it to your channels.

Product · Raven

# Creating an AI Agent in Raven

Raven lets you build an AI agent that answers your customers, uses your business knowledge, and hands conversations to your team when needed. Every agent is set up in five sections: **Instructions**, **Resources**, **Actions**, **Test** and **Deploy**.

| Section | What you do there |
| --- | --- |
| Instructions | Tell the agent who it is, what it handles, how it talks, and when to hand over |
| Resources | Train the agent on your website, documents and Q&A |
| Actions | Let the agent take actions such as handing over or closing tickets |
| Test | Chat with your agent, publish your changes, and activate it |
| Deploy | Connect the agent to your channels |

## Before you start

1. Open the **Raven** module.
2. Select **Create an Agent**.
3. Choose the **agent type**. You can't change the type later, so pick the one that matches where the agent will work:

| Agent type | Use it for |
| --- | --- |
| DM | Private conversations, such as WhatsApp, Messenger, Instagram DMs and web chat |
| Community | Public comments on your Facebook and Instagram posts |
| Email | Email tickets |

4. Add a **name** for your agent.
5. Upload an **avatar** (an image of up to 2 MB). This is the picture customers see next to the agent's replies.
6. Save. Your agent is created as a draft, and you can now set it up in the five sections below.

**Good to know**

- **AI credit:** Each new customer conversation uses AI credit. If your credit runs out, the agent stops starting new conversations. You can top up from your billing page.
- **Plan limits:** Your plan sets how many active AI agents and connected channels you can have.
- **Have your content ready:** your website URL, policy documents (PDF, DOCX, DOC or TXT, up to 5 MB each), and a list of common questions with answers.

## 1. Instructions

The Instructions box is where you describe how your agent should behave. It is the most important part of the setup, so be clear and specific. The Best practices section at the end explains how to write good instructions.

A good set of instructions covers:

- **Role and scope:** who the agent is, which business it represents, and what it should and shouldn't help with.
- **Tone and language:** for example, "Be friendly and concise. Reply in the customer's language."
- **Key policies:** delivery times, returns, pricing rules and anything the agent must follow exactly.
- **When to hand over:** for example, complaints, refund requests, or when the customer asks for a person.
- **What never to do:** for example, promise discounts, guess an order's status, or share internal information.

If you don't mention tone or language, the agent uses a professional, warm tone and replies in the customer's language.

**Example**

> You are Mia, the customer support assistant for Acme Store, an online fashion shop in Dubai. Help customers with product questions, delivery, returns and order status. Be friendly and concise, and reply in the customer's language.
>
> Delivery takes 2–4 working days within the UAE. Returns are accepted within 14 days if items are unused.
>
> If a customer is unhappy, asks for a refund, or asks to speak to a person, assign the conversation to the Support team. Never offer discounts or make promises that aren't in our policies.

## 2. Resources

Resources are the knowledge your agent uses to answer questions. A resource is only used after it has been **trained** and is **active**. Uploading it is not enough.

| Type | How to add it | Good to know |
| --- | --- | --- |
| Website | Enter your website address and start the crawl. Review the pages found, select the ones you want, then click **Train** | Up to 5,000 pages per crawl. When you retrain, only pages that have changed are updated |
| Files | Upload your documents, then click **Train** | PDF, DOCX, DOC or TXT, up to 5 MB per file |
| Q&A | Add questions with their answers, then click **Train** | Best for exact answers such as prices, opening hours or return windows. Questions can be up to 500 characters |

**Tips**

- Training runs in the background. Wait until it finishes before testing.
- To stop using a resource without deleting it, switch it to inactive.
- The Files and Q&A lists show untrained items by default. If a trained item seems to be missing, change the filter.
- Keep your resources up to date. When a policy changes, update the file or Q&A and train it again.

## 3. Actions

Actions let your agent do things, not just answer questions. For each action you add, say in your Instructions when the agent should use it.

| Action | What it does |
| --- | --- |
| Assign ticket | Hands the conversation to a team member or group |
| Close ticket | Closes the ticket once the customer's issue is resolved |
| API | Connects to another system, for example to look up customer details |
| Product search, order placement, order tracking | Helps shoppers find products, place orders and track them. Requires a connected Shopify, Salla or Zid store |
| Reply, private reply, hide or delete comment | Manages comments on your Facebook and Instagram posts (Community agents only) |

## 4. Test

Use the test chat to try your agent before customers see it. Tests run on your draft, so your live agent isn't affected.

1. Ask the questions your customers usually ask. Also try unusual questions, questions your resources don't cover, and situations that should trigger a hand-over.
2. Check that answers match your resources and that actions happen at the right moments.
3. Use **Clear chat** to start a fresh conversation between tests.
4. When you're happy with the results, click **Publish** to make your changes live. To throw away your changes instead, click **Discard**.
5. Turn on the **Active** toggle at the bottom of the Test section to activate your agent.

After your agent is live, any change you make to Instructions or Actions is saved as a new draft. Customers keep getting the published version until you publish again. Changes to Resources take effect as soon as training finishes.

## 5. Deploy

Once your agent is active, connect it to the channels it should handle.

1. Go to **Integrations** and find the channel you want to use, such as WhatsApp, Messenger, Instagram, web chat or email.
2. Click the **three dots (⋮)** on that channel.
3. Select **Ticket Settings**.
4. **Disable the chatbot**, then **select your AI agent**.
5. Save, then repeat for any other channels.

Each channel can be connected to only one AI agent at a time. Your agent now replies to customers on the connected channels. Review the first conversations in your inbox, and keep improving your Instructions and Resources over time.

## Best practices

### Writing instructions

**Define the agent's role**
Clearly state who or what the agent is and what domain it operates in.
*Example: "You are an expert travel planner specializing in budget-friendly vacations."*

**Provide context**
Summarize any background, policies or constraints the agent needs to know.
*Example: "We are developing marketing copy for eco-friendly products aimed at millennials. The tone should be positive and inspiring."*

**Be specific with the request**
Specify exactly what you want: the format, style or depth of the answer.
*Example: "Help the user with their grievances and resolutions in a brief but informative manner."*

**Use step-by-step instructions**
Guide the agent with a clear, ordered approach to how it should think or respond.
*Example: "First, identify the main issue. Next, evaluate possible solutions. Finally, recommend one with a clear justification."*

**Provide examples ("few-shot" demonstrations)**
When possible, give examples of good replies (and optionally, poor ones). This shows the agent the style and quality you expect.
*Example: "Here are two examples of well-structured product descriptions. Follow a similar tone and structure."*

**Encourage clarification when needed**
Tell the agent to ask for more information if certain details are missing.
*Example: "If you are unsure about the product's color variants, ask a clarifying question."*

**Include constraints and policies**
Clearly state any constraints, such as length limits, style guidelines, or compliance rules (privacy, brand voice and so on).
*Example: "Your answer must be fewer than 500 words, follow the Oxford style guide, and avoid any confidential data."*

**Refer to your Actions by name**
If the agent can use Actions, say what each one is for and when to use it.
*Example: "Use the Order Tracking action to get the order status. Use the Assign Ticket action to hand the conversation to the Support team if the order is delayed by more than 5 days."*

**Tell the agent what to do when it doesn't know**
Agents should never guess. Give it a fallback.
*Example: "If the answer isn't in your knowledge, say you'll check with the team and assign the ticket to Support."*

**Keep instructions consistent**
Avoid rules that contradict each other or your Resources. If two rules conflict, the agent may follow either one.

**Write for the channel**
Chat customers expect short replies. Ask for brief messages on WhatsApp, Messenger and web chat, and fuller, well-structured replies for email.

### Managing knowledge

- **Keep facts in Resources and behaviour in Instructions.** Put prices, policies and product details in Resources so you can update them without rewriting your instructions.
- **Use Q&A for answers that must be exact,** such as prices, opening hours and return windows.
- **Only train relevant pages.** Leave out pages like login, cart, or blog posts that don't help customers. Irrelevant content can lead to off-topic answers.
- **Avoid duplicate or outdated content.** If a policy appears in several places with different details, the agent may give mixed answers. Remove or deactivate old versions.
- **Retrain after changes.** When your website or policies change, train the affected resources again.

### Testing and improving

- **Test like a real customer.** Use real questions from past conversations, including spelling mistakes, mixed languages and vague requests.
- **Test the hand-over.** Make sure the agent passes conversations to your team when it should.
- **Change one thing at a time.** After each change, test again so you know what made the difference.
- **Review live conversations regularly,** especially in the first few weeks, and turn common gaps into new Q&A or clearer instructions.
- **Keep an eye on your AI credit** so your agent doesn't stop replying to new customers.

---
HTML: https://userevora.com/docs/product/create-an-ai-agent
