After configuring your AI agent’s behavior and Knowledge, test how it handles customer conversations before making it available on your messaging channels.

This process has two stages:

1. Test and improve: Check responses, identify problems, and refine the agent.
2. Deploy: Choose where, when, and for which conversations the agent should respond.

A deployment controls:

• What starts the conversation
• The messaging channels the agent uses
• The conversations the agent can handle
• When the agent is active
• How messages are delivered
• What happens after the agent exits a conversation

Before testing and deploying your agent:

• Complete the steps in “Set up your AI agent’s behavior and knowledge”.
• Make sure your agent has Rules and at least one Knowledge source. Testing is locked until both are configured.
• Review any actions, Context and response settings, and exit conditions you want to test.
• Connect the messaging channels you want to use. Only connected channels are available for deployment.

Testing does not send messages to real customers or modify real customer data.

Part 1: Test and improve your agent

Test your agent with different customer questions and scenarios before publishing it. Use the results to improve its Playbook, Knowledge, actions, and exit conditions.

Once you have completed configuring your AI agent, you can click on the “Preview” tab at the top of the page to go to the Testing page

Choose a testing method

You can test the agent in two modes:

Use Chat to explore how the agent handles a conversation in detail. Use Batch test to evaluate a broader range of common questions, business rules, and edge cases.

Each question in a batch test runs independently as a single-round test. Batch tests do not support multi-message conversations.

Test in Chat mode

Chat mode lets you have a multi-message test conversation with the agent in a sandboxed environment. Test messages are not sent to real customers and do not modify customer data.

The header shows how many assigned Knowledge sources are currently readable by the agent.

To start a test conversation:

1. Select the Chat tab.
2. Choose a suggested question or enter your own in Message your agent
3. Send the message and review the response
4. Enter follow-up messages to check how the agent maintains context across the conversation.

Press “Shift + Enter” to add a new line to your message.

Click “Reset” when you want to clear the current conversation and start a new test.

Test scenarios such as:

• Common customer questions
• Follow-up questions
• Incomplete or unclear requests
• Requests with missing information
• Questions that depend on assigned Knowledge sources
• Requests that should trigger an action
• Requests that should cause the agent to exit or escalate
• Questions the agent should not answer
• Variations in wording, tone, and spelling

Review execution details

Expand the execution details beneath a response to understand how the agent handled the message.

Depending on the response, the details may include:

• Response routing: The response route selected by the agent and its confidence level
• Why this route was selected: An explanation of how the agent interpreted the request and why it selected that route
• Cited data sources: The Knowledge sources considered when generating the response
• Logic trace: The steps completed while processing the message, such as planning the response or searching Knowledge

Use these details to check whether the agent selected the expected response route and referenced relevant Knowledge.

If the response is incorrect or incomplete, check whether the required information exists in the assigned Knowledge sources and whether the Playbook clearly explains how the agent should handle the request. Update the relevant settings, reset the test conversation, and test again.

Run a batch test

Batch test lets you run multiple questions and review their results together.

To open it, click “Batch test” in the top-right corner.

You can add questions by:

• Generating suggested questions
• Writing your own questions
• Importing questions from a CSV file

Generate suggested questions

Click “Generate suggested” to let AgentFlow create questions based on the agent’s current configuration, including its:

• ProceduresPlaybook
• Knowledge
• Rules
• Persona
• Enabled integrations

The number of questions may vary depending on the agent’s setup.

Questions appear progressively in the batch test table and run automatically. You can review completed results while AgentFlow continues generating and running other questions.

Generate a new set after making significant changes to the agent so the suggested questions reflect its latest configuration.

Stop question generation

While suggested questions are being generated, click Stop to prevent AgentFlow from adding more questions.

Questions that have already been generated remain in the table. Questions that are running or have already been answered are also retained.

Stopping generation does not remove existing questions or results.

Write your own questions

Click “Write your own” to add questions based on your business requirements and actual customer scenarios.

Add questions that test:

• Common customer requests
• Business-specific rules
• Unusual or difficult cases
• Questions that previously received an incorrect response
• Knowledge-dependent answers
• Actions and integrations
• Exit and escalation scenarios

After you add the questions, they appear in the batch test table and run automatically.

Import questions from a CSV file

Use “Import CSV” when you already have a set of test questions or want to reuse the same questions across agents.

You can download the sample CSV to review the required structure.

The CSV file must:

• Include one question per row
• Place the question in the first column
• Optionally include a header row

To import questions:

1. Click “Import CSV”.
2. Select the CSV file from your device.
3. Wait for AgentFlow to import the questions.

Imported questions appear in the batch test table with a “Pending” status. Run the pending questions when you are ready.

If the file does not contain any valid questions, AgentFlow displays an error.

Each imported question runs independently. Use Chat mode if you need to test a multi-message conversation.

Review and rate responses

The batch test table shows each question’s execution status and rating. Use the “Status” and “Rating” filters to find questions that still need attention.

Click a question to open the Evaluate answer panel. The panel shows:

• The test question
• The response time
• The agent’s response
• Actions taken
• Guided Workflow
• Options for rating the response

Expand “Actions taken” or “Guided Workflow” to review how the agent handled the question.

Check whether the response:

• Answers the question correctly
• Includes the necessary information
• Uses relevant Knowledge sources
• Follows the expected workflow
• Performs the required actions
• Uses an appropriate tone
• Exits or escalates when required

Click “Re-run” at the top right corner to run the question again using the agent’s latest configuration.

Rate a response

Under Rate the response, select:

When you select “Poor”, choose the reason that best describes the issue:

• Wrong answer
• Missing info
• Cited wrong source
• Tone off
• Didn’t take action
• Other

You can also enter an optional “Internal note” with more context about your rating or the expected response. Internal notes are not shown to customers.

Save the rating before moving to another question.

Improve Poor-rated responses with Co-pilot

After rating one or more responses “Poor”, use Co-pilot to review those cases and identify possible improvements.

If some responses have not been rated, AgentFlow shows:

• The number of cases that still need a rating
• The number of Poor-rated cases included in the review

Choose:

• Keep rating to return to the batch test and continue evaluating responses.
• Review [number] cases to continue with the currently Poor-rated cases.

Only Poor-rated cases are included in the review. Good, Acceptable, and unrated cases are excluded.

Co-pilot uses the Poor rating, selected reason, response details, and any saved internal notes to investigate the issue. It may:

• Identify patterns across Poor-rated responses
• Ask what behavior you expected
• Suggest changes to the Playbook
• Recommend adding or updating Knowledge
• Direct you to the relevant agent settings
• Suggest changes to actions or exit conditions

Review each suggestion before applying it. After making changes, click “Re-run” for the affected questions or use “Re-test all” to check whether the responses have improved.

Filter batch test results

Use the Status and Rating filters to find results that need attention.

For example, you can filter for:

• Pending or running questions
• Completed questions
• Poor-rated responses
• Unrated responses

Manage batch tests

Click “Manage” to access the following options:

“Re-test all” keeps the questions but replaces their results with new executions.

“Reset results” keeps the questions but clears their existing results, ratings, and notes.

“Regenerate suggested” replaces the previously suggested questions. Manually written and imported questions are not regenerated.

The exported report includes the questions, agent responses, statuses, ratings, notes, sources, and response performance. “Download report” is unavailable until at least one question has been answered.

The exported report includes the questions, agent responses, statuses, ratings, notes, sources, and response performance.

Test reports are session-based. Once you reset the test or leave the testing session, you can no longer access the report. If you want to keep a record of the results, download the report after completing the test and any ratings or notes.

To test only one question again, open its result and click “Re-run”.

Check whether the agent is ready

Before deployment, confirm that:

• Common questions receive accurate and complete answers.
• The agent follows your business rules.
• Responses use the expected Knowledge.
• The agent asks for missing information when necessary.
• Actions occur at the correct point in the conversation.
• Exit and escalation conditions work as intended.
• The tone and response style match your requirements.
• Poor-rated results have been reviewed and resolved.
• Changes have been retested.

Part 2: Deploy your agent

Deploy your agent when it is ready to handle customer conversations.

Configure how conversations start, which channels the agent uses, who it can interact with, when it is active, and what happens after it exits a conversation. When the deployment is ready, click “Publish”.

Configure the trigger

The “Trigger” section determines what starts a conversation with the agent. Each deployment has one trigger. Both trigger options are available on every plan.

Choose one of the following:

A new deployment uses Incoming message by default.

When you select Click-to-WhatsApp ad, the agent receives information about the ad the customer clicked and can use it to answer questions during the conversation.

Under Ads, choose:

For Specific ads:

1. Copy the ad ID from Meta Ads Manager. Use the ad ID, not the ad set or campaign ID. An ad ID looks like 120200000000000000.
2. Paste the ID into the ad ID field.
3. Click Add ad ID to add another ad. You can enter up to five ad IDs per deployment. Add ad ID is no longer offered after you add five.

Each ad ID row must contain a value. If you try to publish with an empty row, the following message appears: “Enter an ad ID, or select All ads.”

The restriction matches the ad the customer actually clicked. An ad ID included as text in the customer’s message does not count as a match.

If an ad no longer exists in Meta, its ID will not match new conversations. Conversations already running with the agent are not interrupted.

What the agent knows about the ad

When a conversation starts from a Click-to-WhatsApp ad, the agent receives the following information separately from the customer’s message:

• The ad’s headline
• The ad’s body text
• The ad ID
• The link destination

The agent keeps this information throughout the conversation. It is not carried into a later conversation that did not start from an ad.

The agent reads the customer’s message as their question. The ad link and reference attached before the first message are not treated as part of that question.

Use the agent’s Rules to define whether and how it mentions the advertised product. For example, you can instruct it to acknowledge the product in the ad before answering and use the ad’s wording for the product name.

The following limitations apply:

• The agent receives the ad’s text, but cannot answer questions about its image or video.
• Ad text provides context for the conversation. It cannot override the agent’s Rules, Procedures, or Knowledge.
• This behavior applies to Click-to-WhatsApp ads. It does not cover click-to-message ads that open Instagram or Messenger.

The conversation’s appearance in Inbox does not change. Existing automations that identify ads by matching message text continue to work.

When the ad isn’t detected

Some ad conversations reach SleekFlow without ad information. WhatsApp on mobile sometimes removes the ad reference before the message arrives. This is outside SleekFlow’s control.

A customer may also start a conversation from an organic Facebook or Instagram post rather than a paid ad. In this case, there is no ad information to provide to the agent.

When no ad information is available:

• An Incoming message deployment can pick up the conversation if it matches the deployment’s other settings.
• A Click-to-WhatsApp ad deployment does not pick up the conversation.
• A Specific ads restriction does not match.
• If an agent handles the conversation through an Incoming message deployment, it responds normally without ad context.

To cover campaign conversations that arrive without ad information, set up a separate Incoming message deployment as a fallback. A deployment restricted to specific ads may not cover every conversation generated by your campaign.

Configure channels

The Channel section determines which messaging channels the agent can use for this deployment.

To add a channel:

1. Click “Add channel”.
2. Select a connected messaging channel.
3. Repeat as needed to add other channels.

If you selected Click-to-WhatsApp ad as the trigger, the Channel section displays: “WhatsApp channels only, because ad clicks arrive on WhatsApp.” Only WhatsApp channels are available for selection. Non-WhatsApp channels are not shown.

You must add a channel before publishing the deployment.

If you switch an existing deployment to Click-to-WhatsApp ad while it has non-WhatsApp channels selected, a confirmation modal appears. It lists the channels that will be removed.

• Click “Cancel” to keep the current trigger and channels.
• Click “Switch trigger” to use the ad trigger and remove the listed channels.

Switching back to “Incoming message” does not require confirmation and does not remove any channels.

Configure reply delivery

The “Reply delivery” section controls how longer replies are sent to customers.

Turn on “Split replies into separate messages” to send a long reply as several chat bubbles instead of one message.

Turn it off to send the reply as a single message.

Configure the target audience

The “Target audience” section determines which incoming conversations the agent can handle.

Choose:

• All incoming conversations: The agent can respond to all eligible conversations received through the selected channels.
• Custom audience: The agent only responds to conversations that match your audience rules.

For a custom audience, you can create rules using:

• Contact labels
• Phone numbers
• Keywords

Review the rules carefully to make sure the agent only joins the intended conversations.

Target audience rules use contact details, such as labels and phone numbers, or keywords in the customer’s message. They do not determine where the conversation started. To restrict a deployment to Click-to-WhatsApp ad conversations or specific ads, use the “Trigger” section.

Configure active hours

The Active hours section determines when the agent can respond. Active hours are based on your company’s time zone.

The agent is available at all times by default. To limit its availability:

1. Turn on “Set specific hours”.
   
2. Select the active days.
3. Set the “From” and “To” times for each day.
4. Review the company time zone displayed in the deployment settings.

Overnight schedules are supported. For example, you can set the agent to respond from 10:00 PM to 6:00 AM.

For more information about company and personal time zone settings, see Personal settings.

Configure post-exit actions

The “Post-exit actions” section determines what happens after the agent exits a conversation under each exit condition.

Each exit condition appears as a separate card. For example, the deployment page can include conditions such as:

• Timeout
• Human Escalation
• Customer goal completed

Each card shows the exit condition and any actions that will run after the agent exits.

To configure a post-exit action:

1. Find the exit condition you want to configure.
2. Click “Add action”
   .
3. Select an action.
4. Complete the action settings.
5. Add another action if needed.
6. Review the actions before publishing.

Available post-exit actions include:

For example, when a customer requests a human:

1. Use Send message to tell the customer that a teammate will continue the conversation.
2. Use Assign to to assign the conversation to the Support team.

Post-exit actions are optional. If you do not configure one, the agent stops handling the conversation without performing another action.

The two-action limit applies to post-exit actions configured on this deployment page. Advanced workflows created in Flow Builder are not limited by this deployment-page setting.

Review and publish

Before publishing, review:

• Trigger. If you selected Click-to-WhatsApp ad, check whether the deployment covers All ads or Specific ads. For Specific ads, make sure every ad ID row is filled in.
• Selected channels
• Target audience
• Active hours
• Reply delivery
• Post-exit actions

Click “Publish” when the deployment is ready.

If a required field is incomplete, AgentFlow prevents publishing until you resolve the validation issue.

After publication, the agent can respond to new conversations that match the deployment’s trigger, channels, audience settings, and active hours.

Changes made after publication do not affect the live deployment until you publish them.

Manage deployments

The “Deployments” page shows every deployment configured for the agent. Once you have deployed the AI agent, you will find the “Deployments” tab in the “Edit agent page”

Each deployment card displays:

• The deployment name
• Its current status
• Its trigger: Incoming message, Ad clicks, or a label such as Ad clicks · 3 ads when specific ads are selected
• The connected messaging channel
• The agent’s availability
• A toggle for enabling or disabling the deployment

A deployment managed in Flow Builder is marked Managed in Flow Builder and does not display a trigger label. It may use a trigger that cannot be configured on the deployment page.

Create a deployment

Click “Create deployment” to configure another deployment for the agent.

Use separate deployments when the agent needs different channels, audiences, schedules, or post-exit actions.

Duplicate a deployment

Each deployment has one trigger. To handle both Click-to-WhatsApp ad conversations and direct messages, create two deployments. You can duplicate an existing deployment to reuse its settings.

To duplicate a deployment:

1. Click on the  icon on a deployment card.
   
2. Click “Duplicate”
   
3. The copy opens with its name field ready to edit. Enter a name that describes the conversations it will handle.
4. Review the copied settings. Change the trigger, channel, or target audience as needed to avoid overlap.
5. Review any warnings and configure missing post-exit actions.
6. Click “Publish” when the copy is ready.

The copy includes the original deployment’s trigger and ad settings, channel, target audience, active hours, and post-exit actions. Exit steps managed in Flow Builder are not copied.

Edit a deployment

Click a deployment card to review or update its settings. Publish the changes before they take effect.

Enable or disable a deployment

Use the toggle on the deployment card to control whether the deployment is active.

Disabling a deployment prevents the agent from joining new matching conversations through that deployment.

Avoid overlapping deployments

A Click-to-WhatsApp ad deployment and an Incoming message deployment on the same WhatsApp channel do not overlap. They handle different conversation sources, and Incoming message excludes detected Click-to-WhatsApp ad conversations.

Two Incoming message deployments on the same channel, or two ad deployments with overlapping ad selections, can match the same conversation and enroll the contact more than once if their other settings also match.

Review audience rules before enabling multiple deployments or automations for the same channels and conversations.

If more than one matching flow or deployment is triggered, the same contact may be enrolled with the agent more than once. Use distinct channels, audience rules, labels, or other routing conditions to prevent unintended overlap.

Use Flow Builder for advanced deployments

Use the “Open Flow Builder” when you need advanced triggers, conditional routing, longer workflows, or more complex post-exit behavior.

You can also place an AI agent after the “Customer messages you from a Click-to-WhatsApp ad” trigger on the Flow Builder canvas to configure routing beyond the options available on the deployment page.

Manage live AI conversations in Inbox

After deployment, your team can use Human-AI Conversation Control to view and manage live AI sessions in Inbox. Depending on the session state, teammates can take over, pause, resume, or end the AI session.

For setup instructions and session behavior, see Manage AI agent sessions in Inbox with Human-AI Conversation Control.