AgentFlow lets you automate customer conversations, resolve enquiries instantly, qualify leads, and support your team around the clock. By configuring your AI agent in SleekFlow, you can teach it your business knowledge, shape how it speaks, set guardrails for sensitive topics, and define when it should hand off to humans.

This guide walks you through each setup step so you can build an AI agent that improves response speed, reduces manual workload, and delivers consistent, on-brand support across every conversation.
 

Before adding AgentFlow to your customer journeys, you’ll first configure how your AI agent understands your business, responds to customers, and works within your workflows. Each setup tab focuses on a specific part of AgentFlow’s behaviour — from the knowledge it relies on, to the actions it can perform, to how it’s deployed in your flows.

Here’s what you can configure:

1. Knowledge: Add and manage the information your AI agent can reference when generating answers.
2. Integrations: Connect external tools or platforms so your agent can use synced business data in conversations.
3. Instructions: Define how the AI should respond, what it should avoid, and when it should exit a conversation.
4. Advanced actions: Enable optional capabilities like lead scoring or automatic labeling to enhance your workflows.
5. Performance testing: Validate your agent’s accuracy with auto-generated test cases before going live.
6. Flow deployment: Deploy your configured AI agent into Flow Builder so it can start assisting customers.

1. Knowledge

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.

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.

2. Integrations

The “Integrations” tab allows your AI agent to pull data directly from your connected third-party platforms. This helps AgentFlow provide more accurate and contextual responses based on live information from your external systems.

Currently, Shopify is supported. You can connect your Shopify store to enable product recommendations and allow your AI agent to reference your latest catalog data.

To learn how to set up and manage the Shopify integration in detail, see our full guide on Shopify integration in AgentFlow

3. Instructions

The “Instructions” tab defines how your AI agent responds, what tone it uses, how it behaves in different situations, and when it should exit a conversation. Most fields come pre-filled with recommended defaults, so you can set up your agent quickly without needing to write prompts manually. You can keep the defaults or adjust them to match your brand and use case.

How to respond

This section defines the core communication style of your AI agent — how it replies, what tone it uses, and how it introduces itself to customers. Most fields are pre-filled with recommended defaults, and you can either use them as-is or customise them to better match your brand.

Response type

Choose how your AI agent should prioritise its responses. This affects the speed, depth, and level of reasoning in each reply.

Available options include:

• Prioritize speed (default): Uses basic knowledge retrieval to generate fast replies. Suitable for straightforward FAQs and direct questions.
• Prioritize response quality: Uses advanced checks and personalisation for more thoughtful, context-aware replies. Recommended for sales or complex support use cases.

Select the option that aligns with your conversational goals; you can change it anytime.

Personality

Choose the tone and style your AI agent should use when replying to customers. Each preset automatically fills in the tone guidance box.

Options include:

• Professional and friendly (default): Balanced, approachable tone suitable for general support.
• Persuasive and confident: Helpful for sales-driven conversations when the AI needs to guide customers toward decisions.
• Custom: Create a custom personality.
  • When selected, you must describe how the AI should sound — for example: “Friendly and casual”, “Warm and supportive”, or “Direct and concise”.

The tone guidance text box appears underneath the dropdown and becomes required when “Custom” is chosen.

AI intro message (optional)

Write a welcome message the AI sends the first time it joins a conversation. Use this to introduce the AI, set expectations, or offer starting guidance.

• This field is optional.
• You may follow the pre-filled example or write your own.
• You can enter up to 1,000 characters.

Unsupported media type reply (optional)

Set the message the AI should send when it receives media formats it cannot process — such as voice notes, videos, or documents.


 

• This reply is optional.
• A suggested example is included for your convenience.
• You can enter up to 500 characters.

This helps ensure the conversation stays clear and consistent even when unsupported files are sent.

Behaviour

Define the primary role and responsibility of your AI agent. This determines the type of enquiries it should handle and how it should assist users.

Preset options include:

• Act as first-line support for product-related inquiries (default)
• Act as a proactive sales assistant to guide leads toward a purchase

Each preset automatically fills in a description explaining the expected behaviour, response style, and when the AI should hand over to a human.

You can also select:

• Custom: When chosen, you must describe the intended behaviour — including what tasks the agent should prioritise, how it should respond, and when to hand off to a human agent.

If your agent needs to cover multiple responsibilities, click “Add behaviour” to include additional roles.

What to avoid

Use this section to define situations your AI agent should not respond to directly.

These settings help the agent recognise out-of-scope, sensitive, or high-risk messages and guide customers to the right next step instead of giving an incorrect or inappropriate reply.

Each entry includes three parts:

• Guardrail — what category of messages the AI should steer away from
• Observe for — signals or message types that should trigger this rule
• How to react — the safe, guided response the AI should use

Most fields come pre-filled when you choose a preset, and you can customize them as needed. If your AI agent needs to avoid multiple topics, click “Add guardrail” to create additional entries.

Each rule works independently and can be customized with its own observe-and-react behaviour.

Use case 1 — Pricing, payments, or account-specific questions

Choose this when your AI agent should not handle enquiries that involve personal data or financial disputes.

• Guardrail: 
  • Deflect pricing and account-specific inquiries
• Observe for:
  • Questions about payments, refunds, or invoice issues
  • Personal account details or verification
  • Complaints that require human review
• How to react:
  • Acknowledge the question politely
  • Avoid giving a direct answer
  • Offer to transfer the user to a human agent or point them to the correct support channel

Use case 2 — Sensitive or complex enquiries beyond the AI’s scope

Select this when your AI should recognize messages that are unsuitable for automated responses.

• Guardrail: 
  • Deflect messages that go beyond general inquiries
• Observe for:
  • Requests requiring specialised expertise
  • Emotional or sensitive topics
  • Messages unrelated to your service or business
• How to react:
  • Let the user know the AI can’t handle that topic
  • Maintain a professional, safe tone
  • Suggest escalating to a human agent or another resource

Use case 3 — Custom scenarios unique to your business

Choose “Custom” if you have specific topics you don’t want the AI to respond to.

• Guardrail: 
  • Create your own custom rule
• Observe for: 
  • Define keywords, intents, or types of enquiries that should trigger this scenario
• How to react:
  • Write the exact phrasing or behaviour you want the AI to follow — such as redirecting the user, sending compliance statements, or escalating to a human agent

Exiting a conversation

Use this section to define when your AI agent should leave a conversation and hand over to a human agent or fallback step.

Exit conditions are triggered by signals that occur within the chat, ensuring the AI exits gracefully when it’s no longer able or appropriate to continue assisting.

Your agent will exit the conversation if any of the configured conditions are met.

Pre-filled exit conditions

Two exit conditions are included by default:

1. Confidence is low (Locked, cannot be removed or edited)
   
   1. The AI agent will automatically exit when its confidence score falls below a certain threshold.
   2. This prevents the AI from generating uncertain or unreliable answers and ensures the conversation is handed over to a human when needed.
       
2. Timeout (Preset, duration adjustable)
   
   1. The AI exits the conversation after a specific amount of time has passed since the customer entered the flow.
   2. You can adjust the timeout duration (e.g., 24 hours).
   3. This helps prevent long or inactive conversations from remaining assigned to the AI for too long.

Additional exit conditions you can add

Below the pre-filled conditions, you can create additional exit rules tailored to your workflow.

Each added condition includes:

• Condition name
• Condition type
• Exit behaviour (editable text describing what the agent should do)

Common condition options include:

• Human agent requested
  
  • The AI exits when a customer clearly asks to speak with a human
  • A default exit message is provided, and you can customize it if needed.

• Keyword detected
  
  • The AI exits when a specific keyword or phrase appears in the customer message.
  • This is helpful when certain terms or intents should immediately trigger handover.
  • You define both the keyword and the exit behaviour.

Click “+ Add condition” to create as many additional rules as you need.

Each condition operates independently, and the AI will exit as soon as any one of them is triggered.

4. Advanced actions

Advanced actions let your AI agent perform additional tasks beyond replying to messages.

These capabilities are optional and switched off by default. You can enable them to help with lead qualification or automatic conversation organisation.

Two advanced actions are available:

• Calculate lead score
• Add label

Calculate lead score

Toggle on to enable this setting if you want your AI agent to automatically score incoming leads based on their messages.

How it works

• The agent calculates a lead score when the contact replies (this trigger is pre-filled and cannot be changed).
• You define scoring criteria and assign each one a weight.
• All criteria weights must add up to 100% for scoring to work.

What you can configure

• Add criteria (e.g., buying intent, urgency, product interest)
• Set weight (%) for each criterion
• Adjust or remove criteria as needed

When enabled, the AI agent will apply the criteria to each conversation and generate a score automatically.

Add label

Toggle on to use this action to let your AI agent automatically apply conversation labels based on message content.

How it works

• You choose the list of labels the AI can select from.
• The agent will evaluate the conversation and apply one label — whichever is the most relevant.

What you can configure

• Label options for AI
  • Select the labels the agent is allowed to assign (e.g., “Qualified lead”, “Support”, “Urgent”).
• Trigger condition
  • Describe when a label should be assigned — for example: “Assign ‘Qualified lead’ when the customer expresses purchase intent.”

This helps keep conversations organized and can be used to trigger downstream automations or routing rules.

5. Performance testing

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.

Use Performance test to evaluate how well your AI agent can answer questions based on your linked documents, without writing test cases manually.

This tool automatically generates sample questions from your knowledge base, then tests the AI’s responses and highlights any failed or low-confidence answers. It’s useful for validating coverage, checking tone alignment, and spotting gaps before launching your agent.

Benefits and common use cases

Why use Performance testing:

• Validate document coverage: Confirm the AI is using your knowledge base correctly.
• Catch low-confidence replies: Identify where the AI struggles to answer clearly or accurately.
• Avoid manual QA work: No need to write your own test questions — they’re generated for you.

When to use it:

• After uploading FAQ or product documents
• After editing instructions or guardrails
• Before publishing your agent to customers

How to run a performance test

You can follow the steps below to run a performance test:

1. In the AI agent’s settings page, click “Performance testing” on the left-sided panel
   
2. You will be redirected to the “Performance testing” page
   
3. Click “Generate test” at the top right corner to start
4. A pop-up modal will appear
   
5. In this pop-up modal, you’ll be required to fill in:
   1. Enter test name
      1. It will be pre-filled. You can also give your test case a clear and descriptive name so it’s easy to identify later
   2. Language configuration
      1.  Select a preferred language before generating test cases. The AI will automatically create sample questions and answers in that language.
   3. Select linked data sources
      1. Select 1 or more data sources that linked to this AI agent. These sources will be used to automatically generate test questions. The agent’s responses will be evaluated against the content from these sources to calculate performance. 
      2. If a source has already been used in a different performance test, it will appear as disabled and cannot be selected again.
6. Once you have filled in the details, click “Generate”
7. You will be redirected back to the "Performance testing” page, where you will see the summary of the performance test
   
   
   Here’s what you’ll find on the page:
   1. Test name and run timestamp
      This helps you track when the test was last run and what version of the AI agent it used.
   2. Total questions generated
      The number of test questions automatically created from your selected data sources.
   3. Accuracy 
      Displays how accurate the answers generated are
   4. Filters and search bar
      Narrow down results by question status (e.g. Passed, Failed, In Progress) or rating.
   5. List of generated questions
      For each test entry, you’ll see:
      1. Question and answer: The AI’s generated response to the question
      2. Status: Whether the answer was marked as Passed, Failed, or still In Progress
      3. Answer confidence: Indicates how confident the AI was in its response (0–100%)
         1. The score is calculated using two main factors:
            1. Similarity – how semantically close the AI’s answer is to the expected one.
            2. Groundedness – how many supporting facts from your knowledge base are correctly referenced.
         2. These two factors are combined into a single percentage score to indicate the overall answer accuracy for that test.
8. You can click on any questions to view the full response details. 
   1. Once clicked, a pop-up modal will appear
      
   2. Here you can:
      1. Compare the AI’s answer with the expected answer (if available)
      2. See the exact confidence score for the response
      3. View which data sources the AI referenced when generating its answer.
      4. Give direct feedback using the 👍 or 👎 icons next to each expected answer. This helps your team and the system identify which generated answers meet your quality standards and which ones need adjustments.

6. Flow deployment

Your AI agent will only start responding to customers after it is added to an active flow in Flow Builder.

The Flow deployment section gives you visibility into where your agent is currently being used and helps you set up new flows.

In this section, you’ll see a list of all flows where your AI agent is currently deployed. For each flow, you can view:

• Flow name
• Status (Active / Inactive)
• Flow log
• Created by
• Last updated by
• Last updated time

If the agent is not yet used anywhere, you’ll see an empty state with an option to create a new flow.

Deploying your AI agent in Advanced Flow Builder

To start using your AI agent in conversations, you must add it to a flow using the AI agent node and configure where and when it should activate.

You can follow the steps below to create a flow with the AI agent node:

1. Click on “Create new flow” in the “Flow deployment” tab
2. You will be redirected to the Advanced Flow Builder, with a flow template applied for you
   
3. When you add the AI agent node, Flow Builder will automatically detect all exit conditions you have configured in the Instructions tab and generate a corresponding AI agent exit step.
4. This step will include branches for conditions such as:
   1. Confidence is low
   2. Timeout
   3. Human agent requested
   4. Keyword detected
   5. System errors
   6. Else (default)
   7. Each branch can be connected to the next step in your flow — such as handing over to a human agent, sending a follow-up message, or concluding the flow.

Finish and publish your flow

Once your flow is set up, you can

1. Click “Save as draft” in the top right corner, or:
2. Click “Publish” to activate the flow.

Your AI agent will begin responding to customers according to the flow and configuration you’ve defined.

Testing out your AI agent in playground

Use the “Playground” to simulate conversations and review how your AI agent responds before going live, with full visibility into what’s happening behind the scenes.

This environment lets you test tone, logic, and coverage by sending sample questions. You’ll also see real-time logs showing which knowledge base sources were used, if lead scores were applied, or whether labels and contact updates were triggered.

You can follow the steps below to use AI agent playground:

1. In the AI agent settings page, the AI agent playground will appear on the right side of the screen
   
2. Once you have finished setting up your AI agent, you can type a question in the chat input field
   
3. Press “Enter” to send
4. The AI agent will respond based on its current configuration