Configuring your AI agent with “Advanced mode”

Learn how to configure and train your AI agent with customizable prompts in “Advanced mode”

Written By Frieda Yip (Super Administrator)

Updated at August 6th, 2025

“Advanced mode” gives you full control and customizations over your AI agent’s behavior. Unlike Basic mode, which uses pre-filled templates, Advanced mode lets you write and edit your own prompts across Instructions, Actions, Guardrails, and Labels — so you can fine-tune tone, logic, and response style to match your workflow.

This mode is recommended for users who are confident in designing prompt-based logic or want to handle more complex use cases.

Turning on/off “Advanced mode”

You can toggle Advanced mode on or off from the agent’s Settings page:

  • When “Advanced mode” is ON, all prompt fields become fully editable.
  • When “Advanced mode” is OFF, the agent will return to Basic mode using simplified, non-editable templates.

 

⚠️ Switching off Advanced mode will discard your changes

 

If you’ve edited any prompts in Advanced mode and turn it off, your changes will be permanently deleted. When you turn Advanced mode back on, all prompts will reset to their default templates — your previous edits will not be recoverable.

 

A confirmation modal will appear before the switch is applied:


  • Click “Go back” to cancel and keep your edits
  • Click “Turn off anyway” to confirm and discard all custom prompts

 

 

 

Once you have created your AI agent, you will be redirected to the set up page of The AI agent you have created. 

To manage your AI agent’s configurations, you can click “Manage” to update its actions, instructions, or linked knowledge base at any time.

Here is a list of settings you can configure within an AI agent:

  • Knowledge base: Link relevant content the AI can reference to generate accurate replies.
  • Instructions: Define your agent’s objective, set a welcome message, and add guardrails to guide how The AI agent should behave in different scenarios.
  • Actions: Select what the AI can do in a conversation, such as send messages, score leads, or exit.
  • Flow deployment: Deploy your AI agent in Flow Builder to start using it. The AI agent will only respond in conversations once it’s added to an active flow.

 

Adding data source to AI agent

Before configuring how your AI agent replies, you’ll first need to provide it with the right data to reference. Adding a data source ensures that your AI agent can generate accurate, helpful responses based on your business content — such as your website, uploaded files, or custom answers.

Make sure you’ve completed this step before setting up instructions or actions in Advanced mode.

Refer to our Help Center article on Adding data source to your AI agent for a full guide on connecting and managing your knowledge base.

 

Configuring AI agent’s Instructions

The Instructions tab defines how your AI agent should behave during conversations — including its purpose, and how to respond in sensitive or unexpected situations. These settings help your AI agent stay aligned with your brand and ensure it engages customers effectively and responsibly.

To configure AI agent’s instructions, you can follow the steps below:

  1. In The AI agentFlow page, click “Manage” on the AI agent card
  2. Click “Instructions” on the left-sided menu
  3. You will be redirected to the “Instructions” page
  4. In this page, you can configure the following:
    1. Instructions (overall behavior)
      1. Define the AI agent’s objective and provide detailed context about your business. You can specify who The AI agent represents, what types of enquiries it should handle, and how it should communicate (e.g. tone, language style, level of detail). The more specific your instructions, the more accurate and on-brand the AI’s responses will be.
      2. Example: “You are a friendly and knowledgeable assistant representing Cat Paradise. Our company provides premium cat grooming, boarding, and wellness services. When introducing yourself, say: ‘Hi! I’m from Cat Paradise. I’m here to help you with any questions about our services or products.’ You assist customers with booking spa sessions, explaining our service packages, and answering product-related questions. Use a warm, helpful tone and keep responses concise but informative.”
    2. Welcome message
      1. Write a greeting that appears when a customer first interacts with your AI agent. Use this to set expectations and establish tone.
      2. Example: “Hi there! I’m here to help with anything about our cat care services, grooming bookings, or product info 🐾”
    3. Guardrails
      1. Guardrails help your AI agent identify and respond appropriately to sensitive or complex topics. They guide how The AI agent should steer the conversation when certain types of input are detected — but they do not trigger additional actions like exiting or handing off the chat. Each guardrail includes:
        1. Observe for: Describe what the AI should watch out for. (For example: refund requests, pricing disputes, complaints, personal information)
        2. How to react: Specify how the AI should respond using tone, phrasing, or clarification questions (For example: acknowledge the concern politely and ask a follow-up question to gather more details)
      2. You can add multiple guardrails based on different types of sensitive content. For example:
        1. Example 1: Order issue
          1. Observe for: “Questions about refund status or order disputes.”
          2. How to react: “Apologize for the inconvenience and ask the customer to clarify their order number so you can better understand the issue.”
        2. Example 2: Damaged product
          1. Observe for: “Mentions of product defects or damaged items.”
          2. How to react: “Acknowledge the issue with empathy and ask the customer to describe the problem or provide a photo if possible.”
        3. Example 3: Data privacy concern
          1. Observe for: “Questions about personal data, privacy, or account security.”
          2. How to react: “Reassure the customer that their data is protected and direct them to review your company’s privacy policy for more details.”
      3. To add a guardrail, click “Add guardrail”, fill in both fields

 

  1. Once you’ve finished configuring your AI agent’s instructions, you can test how it responds using the preview panel on the right.

    Note: You’ll need to have at least one fully trained knowledge base source before testing is available. If no source is ready, the test panel will remain inactive.
    ​​


  1. Once you are done with the configurations, click “Publish” on the top right corner to save the changes

 

💡 Need help writing strong instructions and prompts?

Read our Best practices guide to learn how to write clear agent objectives, tone settings, and guardrails that align with your business.

 

 

Configuring AI agent’s Actions

Actions define what your AI agent can do during a conversation — such as replying to customers, scoring leads, or exiting a chat. Each action is powered by a prompt that tells The AI agent how to respond. You can review and customize these prompts to match your workflow, brand tone, and business objectives.

You can access the AI agent’s Action page by following the steps below:

  1. In The AI agentFlow page, click “Manage” on the AI agent card
  2. Click “Actions” on the left-sided menu
  3. You will be redirected to the “Actions” page
  4. In this page, you can configure the details of each available action for your AI agent:
    1. Send message and Exit conversation are required actions — they are enabled by default and cannot be turned off.
    2. For optional actions like Calculate lead score, and  Add labels, you can use the toggle switch to enable or disable them based on your agent’s role.
    3. Click on any action to view and customize the prompt that controls how the AI performs that task. This lets you fine-tune responses, scoring logic, and internal behavior to match your workflow.

       

“Send message” action

This is the core action that allows your AI agent to respond to incoming customer messages using the connected knowledge base. It is always enabled and runs automatically when a contact sends a message.

In this action, you can configure the following:

  1. Response type: Choose how your AI agent prioritizes its replies:
    1. Prioritize speed – Uses basic knowledge retrieval to generate fast responses. Ideal for straightforward questions or high-volume support.
    2. Prioritize response quality – Uses advanced checks and personalization to generate more thoughtful, context-aware responses. Best for complex or sales-related enquiries.
  2. Message tone of voice: Define how the AI should sound and what kind of behavior it should follow when replying to customers. This prompt guides the AI’s reply style, fallback handling, and content framing. You can include tone descriptors such as:
    1. Friendly and casual
    2. Polite and professional
    3. Conversational but confident
  3. Trigger condition: Defines when this action is triggered during the conversation — for “Send message,” it runs automatically when a customer sends a message.
    1. This action automatically runs when the contact sends a message (#Run when: Contact sends a message)
    2. This condition is fixed and cannot be edited
  4. Once you have completed configuring the instructions of the AI agent, you can test your AI agent on the right screen to see how it responds
  5. Once you are done with the configurations, click “Publish” on the top right corner to save the changes

 

Recommended setup Basic Support agents

Use this setup to handle FAQs and common support enquiries efficiently.

  • Response type: Prioritize speed
  • Tone of voice: Friendly and polite
 

 

Recommended setup for Sales Growth agents

Use this setup to engage leads and uncover sales intent through meaningful dialogue.

  • Response type: Prioritize response quality
  • Tone of voice: Confident and conversational
 

 

“Calculate lead score” action

This action allows your AI agent to evaluate each lead and assign a score between 0 and 100 based on message content, intent, urgency, tone, and fit. It helps your team identify high-quality leads, prioritize follow-ups, and streamline sales qualification.

This action is included by default in the Sales Growth template, but can also be toggled on manually when using other templates like Basic Support.

In this action, you can configure the following:

  1. Trigger condition: The lead score is calculated automatically whenever the contact replies to the AI agent. You don’t need to configure this manually.
  2. Scoring criteria and weights: Define the factors your AI should evaluate when scoring a lead. Each criterion is assigned a percentage weight, which determines how much it contributes to the final score.
    1. How scoring weights work:
      1. The AI evaluates each selected criterion based on the customer’s message.
      2. Each criterion contributes up to its assigned weight toward the total score.
      3. For example, if “Intent and interest level” is weighted at 40%, it can contribute up to 40 points out of 100.
      4. The total weight across all criteria must equal exactly 100% to ensure the AI calculates a complete score.
      5. If a criterion doesn’t apply to your workflow, you can remove it by clicking the “–” button in the top-left corner of the card. You must keep at least 1 active criterion for the lead score to be calculated.
  3. Default example criteria (editable):
    1. Intent and interest level – Is the lead asking about features or making comparisons?
    2. Buying signals – Are they inquiring about pricing or expressing urgency?
    3. Depth and specificity – Are they asking focused, insightful questions?
    4. Engagement tone – Are they enthusiastic, hesitant, or neutral?
    5. Customer fit – Does their role or profile align with your ideal customer persona?
  4. If you need to add more criteria, you can click “Add criteria”
  5. Once you have completed configuring the instructions of the AI agent, you can test your AI agent on the right screen to see how it responds
  6. Once you are done with the configurations, click “Publish” on the top right corner to save the changes

⚠️ Important:

The total weight across all lead scoring criteria must equal 100%.

You can adjust the importance of each criterion by changing its percentage, or set it to 0% if you don’t want it to affect the score. The system will not calculate a lead score unless the total adds up to exactly 100%.

 

 

“Exit conversations” action

This action allows your AI agent to leave the conversation once a specific condition is met. Exit conditions are based on signals within the conversation — such as certain keywords or low confidence — rather than customer properties outside the chat. This action is always enabled and ensures smooth handoff or clean exits when appropriate.

In this action, you can configure the following:

  1. Exit conditions: Define specific scenarios where the AI should end the conversation. You can create multiple conditions using natural language triggers such as:
    1. “Speak to human”
    2. “Confidence is low”
    3. “No matching answer found”
  2. Each condition includes:
    1. Condition name – A label to help you identify the rule
    2. Exit condition – A short explanation of what triggers this exit (For example: when the AI is unsure, or the user explicitly asks to talk to a person)
    3. Condition type – Choose how the exit should be triggered:
      1. Exit based on message signal: Triggered when specific keywords or patterns appear in the customer’s message (e.g. “Speak to human”, “No matching answer found”)
    4. Exit based on lead score: This condition is triggered when the contact’s lead score meets a specific threshold. The lead score is calculated based on the criteria you’ve configured under the Calculate lead score action — such as intent, urgency, tone, or customer fit. You can set the AI agent to exit the conversation when the score:
      1. is less than a value
      2. is more than a value
      3. is between two values

        Use this condition to smoothly hand off hot leads or end conversations that don’t meet your qualification thresholds.

  1. Once you are done with the configurations, click “Publish” on the top right corner to save the changes

Here are some of the examples of exit conditions:

Condition name

Exit condition example

Confidence is low

 

When the AI agent can’t answer with confidence based on the knowledge base

Speak to human

When customer mentions they want to talk to a human agent

End of task

When the AI agent has completed a set of actions or fulfilled its task

Hot lead

When the lead shows strong buying intent or reaches a high lead score threshold

Cold lead

When the lead shows low to no interest, based on the lead score calculated from the criteria set in the “Calculate lead score” action. This may also apply when the customer becomes unresponsive after initial engagement.

 

🛠️ Tips for setting up this action for Basic Support agents

Set exit conditions for when the AI cannot answer accurately or when a customer asks to speak to a human.

  • Recommended conditions: “Confidence is low”, “Speak to human”
  • Consider using keyword detection (e.g. “agent”, “real person”, “help now”) to trigger exits — this is commonly used in basic FAQ bots.
  • Use polite, helpful language to let customers know the conversation is ending or being redirected.
 

 

💼 Tips for setting up this action for Sales Growth agents

Configure exit points for when a lead has been qualified or if further nurturing should be handled by a human.

  • Recommended conditions: “Lead qualified”, “Ask for pricing”
  • You can also exit after scoring the lead to keep the handover seamless.
 

 

“Add labels” action

This action allows your AI agent to automatically tag conversations based on what was discussed. Labels help your team organize conversations, trigger automations, or segment contacts based on interest or intent. The AI will choose the most relevant label from the list you provide and apply it when the trigger condition is met.

This action is optional and can be toggled on or off depending on whether you use labels in your workspace.

 

In this action, you can configure the following:

  1. Label options for AI: Select the list of label options the AI can choose from. The system will evaluate the conversation and apply the most relevant label automatically.
    1. Labels can represent topics (For example: Pricing enquiry, Product interest), customer intent (For example: Ready to buy, Needs follow-up), or internal processes (For example: Escalation, Trial user). 
    2. You can create or select labels from your existing workspace label set
  2. Trigger condition: Define when the AI should assign a label based on what’s said in the conversation. You can use general criteria like:
    1. When a specific keyword or topic is mentioned
    2. After a particular question is answered
    3. At the end of the conversation
    4. Example condition: “Assign a label if the user mentions pricing, interest in grooming services, or requests a follow-up.”
  3. Once you are done with the configurations, click “Publish” on the top right corner to save the changes

 

Configuring AI agent’s Flow deployment

Once your AI agent is configured, you can deploy it inside a conversation flow to start using it with real customers. The “Flow deployment” tab helps you view and manage all the flows where your AI agent is currently active.

To configure your AI agent’s Flow deployment, you can follow the steps below:

  1. In The AI agentFlow page, click “Manage” on the AI agent card
  2. Click “Flow deployment” on the left-sided menu
  3. You will be redirected to the Flow deployment page, where you will see a list of flows where The AI agent is currently in use. You’ll find details such as:
    1. Flow name
    2. Status (Active / Inactive)
    3. Logs
    4. Created by / Updated by
    5. Last updated time

If your agent hasn’t been added to any flows yet, the list will be empty and you’ll see an option to Create new flow.

You can follow the steps below to create a flow in Flow Builder and deploy your AI agent:

  1. In Flow Builder’s editor, click on the “AgentFlow (Beta)” action node to open up its form, which will appear on the right side of the screen
  2. Select the AI agent you would like to use in this flow
  3. Select the WhatsApp channel
  4. Set triggers to determine when the AI joins the conversation
  5. Define exit conditions and fallback actions to control handover logic
  6. Click Save and then Publish your flow to make the deployment live

 

📘 Next step:

Explore Best practices for configuring AI agents to keep your agent reliable, helpful, and on-brand.

 

 

Testing out your AI agent

After you’ve configured your AI agent, you can use the built-in test panel to try out how it responds in real conversations. This helps you review how well your prompts and knowledge base are working, before deploying the agent in live flows.

You’ll find the Test your AI agent panel on the right side of the configuration screen.

You can:

  • Send sample messages as if you were a customer
  • See how the AI responds using the connected knowledge base and configured actions
  • Quickly spot tone or logic issues before going live

This test environment is designed to reflect real-time behavior based on your current setup.

 

What you can test in the playground

In the test panel, you can simulate:

  • Greeting and welcome flows
  • Knowledge-based replies (e.g. product info, FAQs)
  • Guardrail behavior (e.g. when the AI should deflect or escalate)
  • Prompt tone and response clarity

     

Excerpt-level citation

When the AI generates a response based on your knowledge base, you’ll see a “source” label beneath the message. Click “relevant excerpts” to view the exact chunks of content the AI used.

Each excerpt is a chunk: a small segment of data pulled from your uploaded files or website URLs.

This allows you:

  • Verify that the AI is referencing accurate, approved content
  • Spot any gaps or outdated information in your source materials
  • Understand why the AI responded the way it did