Using AI Agent Actions

While AI Agent is replying to a Contact, it can perform specific platform actions like assigning the conversation, closing it, or updating contact information. Understanding these actions helps you design AI Agents that handle conversations smoothly and logically.

This guide covers exactly how each action works, how to write effective instructions for them, and how to avoid common pitfalls like conflicts with your existing Workflows.

Why Actions Are Important

Your AI Agent’s replies help handle customer conversations smoothly. But often, replying alone isn’t enough—your Agent also needs to take clear next steps to move Leads across sales or support pipeline.

That’s exactly what Actions are for. They help your AI Agent move beyond just replying, ensuring conversations progress logically down the conversion funnel, customer details stay accurate, and teams always know who should handle each chat next.

By using the right actions, your AI Agent knows exactly what to do while replying—giving your team peace of mind and keeping your conversations organized and effective.

Actions Your AI Agent Can Take

In this guide, you’ll learn how to configure them by prompting and directing your AI Agent. The key to agentic behavior lies in providing the AI Agent with enough context—clearly defining what to do, and when to do it.

During a conversation, your AI Agent can take the following actions:

Close conversations

What this action does:

This action can be used to close a conversation once the Contact’s request has been fully resolved. Think of it as your AI Agent politely wrapping up a chat in person.

• You can also provide a closing note—a label that explains why the conversation ended. To do this, just mention the closing note category in your instructions. For example:
  

  - If the contact has no more questions, close the conversation and select closing note "Issue Resolved"

• When AI Agent closes a conversation, it will also automatically summarize the conversation and select a closing note.

  Note: If you don’t want summaries or notes, you can instruct: “Do not generate a summary or choose a closing note.” For example:
  

  - If the contact has no more questions, close the conversation and do not generate a summary nor choose a closing note. 

When to use:

• Use this action when the AI Agent has completed the customer’s request—like answering FAQs, collecting feedback, or solving a simple issue—so conversations don’t stay open unnecessarily.

Best practices:

• Always include the exact instruction “Close the conversation” so AI Agent knows to trigger it.

• Clearly explain the specific scenario where the conversation should close (e.g., “If the customer confirms they have no more questions…”).

• Pair this action with a friendly, definitive closing response for AI Agent to use such as: “I hope that helps! If you have other questions, feel free to message again.”

Limitations:

• AI Agents can only choose from existing closing notes. They cannot create new closing notes or modify your existing list. To ensure the correct closing note is applied, use the exact closing note name in your instructions.

• AI Agents can only summarize the latest 20 messages in a conversation. Older messages will not be included in the summary. If important context appears earlier in the conversation, instruct the AI to ask a clarifying question or restate key details before closing.

Assign to agent or team

What this action does:

Assigns the right human agent, team, or even another AI Agent once the AI Agent has done its part. This is especially useful for escalating to a human when the AI Agent can’t resolve an issue, ensuring customers always get the right level of support.

You can also control how team members get assigned. There are two methods:

• Round Robin: Assigns conversations in rotation so workload is shared evenly.

• Least Open Conversation: Assigns to whoever currently has the fewest active conversations.

  Note: If no method is specified, AI Agent defaults to round robin.

- If the contact asked about their appointment schedule, assign to @Scheduling Team by round robin.

You can also assign to individual agents or even unassign a conversation by instructing, "Unassign the conversation".

When to use:

• Use this action when the conversation needs to move beyond the AI Agent’s scope—for example, to a human agent for deeper support, to a specialized team (like Billing or Sales), or even to another AI Agent designed for a different task.

Best practices:

• Use these terms in the actions instruction: “Assign to @User/Team Name” to trigger this action for a scenario. For example, “If {something happened}, assign to @User/Team Name”.

• Ensure your AI Agent asks relevant qualification questions to determine the right team.

• Clearly define conditions in your instructions (e.g., “Assign to Sales if customer mentions pricing or demo”).

Limitations:

• AI Agents cannot see who is online. They have no visibility into agent availability or presence, so they cannot choose the “next available” or “currently active” user. Always specify the exact team or agent name if availability matters.

• AI Agents cannot schedule assignments for a future time. Instructions like “Assign to Sales at 5pm” will not work because the AI cannot delay an assignment. However, immediate assignment logic based on current conditions (e.g., “Assign to Sales if within business hours”) will work, as the AI evaluates the condition in real time and assigns immediately.

• AI Agents cannot run round robin across multiple teams. They can only distribute conversations within one specific team at a time.

• AI Agents do not choose the assignment method (round robin vs least open). However, round robin can only be applied within a team, not across individual users. This means agents must be part of a team for round robin to work. If you want to use least open or any other method, you must explicitly specify it in your instructions; otherwise, the AI will default to round robin at the team level.

• Explicitly instruct AI Agent to stop responding after assignment. To prevent AI Agent from replying again after handing the conversation to a human, include a clear instruction such as:

  Do not respond to the Contact when assigning conversations to @Sales Team 

  This helps avoid unexpected AI replies after handoff and ensures a clean transition of ownership to the human agent or team.

Update Contact fields

What this action does:

Updates Contact details (name, phone number, budget, etc.) automatically based on information collected during the conversation. This is especially useful for keeping your Contact’s data up to date, so agents can focus on meaningful interactions instead of admin work.

• You can also mention Contact field names in your instructions to improve accuracy and ensure data is saved in the right place. For example:

- If the Contact provided their budget, save it as Budget field.

When to use:

• For lead qualification, customer onboarding, or booking appointments.

Best practices:

• Clearly provide AI Agent with instructions to ask the customer for specific details (e.g., “Could you share your preferred email address?”).

• Train your AI Agent with knowledge sources so it gives accurate, context-specific answers when replying to Contacts.

Limitations:

• AI Agents cannot detect when a Contact field update fails. If the field name is unclear, misspelled, or doesn’t exist, the update simply won’t happen, and the AI won’t know it failed. To avoid this, always use the exact field name or field ID in your instructions.

• AI Agents cannot create or delete Contact fields. They can only update fields that already exist in your workspace. If you need new fields or structural changes, create them manually before referencing them in your AI Agent instructions.

Update Lifecycle stage

What this action does:

Changes a customer’s Lifecycle stage (e.g., from “New Lead” to “Qualified Lead”) based on conversation context. This is especially useful for automating lead progression and keeping your sales pipeline accurate without manual updates.

When to use:

• When the AI Agent qualifies a lead or achieves a specific milestone in customer engagement.

Best practices:

• Clearly define the criteria needed to advance the Contact’s Lifecycle stage.

• Always provide the exact Lifecycle Stage name so the AI Agent updates to the right stage at the right time. For example:

- Once the customer confirms their interest in a demo, update the Lifecycle stage to Qualified.

Limitations:

• AI Agents cannot create new Lifecycle Stages. They can only update a Contact to stages that already exist in your workspace. If you want to use a new stage, you must create it first before referencing it in your instructions.

Add comments

What this action does:

AI Agents can add internal comments in a conversation to share context with your team—like quick summaries, key details, or escalation reasons—so human agents don’t need to read the entire chat history.

Comments can also include @mentions to notify specific workspace users (team mentions and AI Agent mentions aren’t supported). For example:

- If you escalate to a human agent, add an internal comment summarizing the issue, what’s been done, and tag @Jane Doe.
- Then assign the conversation to @Support Team.

When to use:

Use this action when you want AI Agent to:

• Leave a handoff note before assigning/escalating to a human agent

• Flag important details (e.g., intent, urgency, requirements)

• Tag a teammate when certain conditions are met

Best practices:

• Keep comments short and scannable (e.g., intent + context + recommended next step).

• Be explicit about when the comment should be added (e.g., “before assigning”, “when the customer requests a human”, “when the issue is urgent”).

• If you want a teammate notified, @mention the exact workspace user.

• In Reports, AI Agent comments are tracked like user comments. Where available, the comment will appear under the AI Agent’s name rather than a human user.

Limitations:

• AI Agents can only @mention workspace users in comments. They can’t @mention teams or other AI Agents yet, so always tag a specific person if you need someone’s attention.

• AI Agent internal comments are limited to 2,000 characters.

Trigger Workflow

What this action does:

Starts an existing Workflow from the conversation based on your AI Agent’s instructions. Use this when the AI Agent needs to kick off more complex automations you already manage in Workflows, such as sending HTTP requests, adding a Google Sheets row, or sending a CAPI event. The Workflow then runs as configured, independently of the AI Agent.

When to use:

• When you already rely on Workflows to execute core business processes and want AI Agents to plug into them instead of rebuilding logic with other actions.

• When the next step goes beyond what AI Agent actions can do on their own—for example, pushing data to external tools or orchestrating multi-step automations.

• When you want to keep business logic centralized in Workflows but let the AI Agent decide

  when that logic should run.

Best practices:

• Turn on Trigger Workflows in your AI Agent’s Actions and describe your conditions in the When and how should this action be performed? field.

• Use ! to reference the exact Workflow name in your instructions. For example:

  If the Contact asks for a refund, trigger !refund_processing.

• Make sure the Workflow you reference is published, unpublished Workflows will not run when triggered.

• Avoid building the same automation twice. If a Workflow performs the same actions as your AI Agent (for example, sending a follow-up message or updating the same Contact field), you may see duplicate messages or Contact updates.

You can define multiple conditions in the same instruction field; each condition can trigger the first step of a different Workflow.

Limitations:

• Trigger Workflows only works with existing Workflows in your Workspace. If there are no Workflows, this action cannot be enabled and you’ll be prompted to create a Workflow first.

• Variables are not passed from the AI Agent to the Workflow. The Workflow cannot automatically see conversation variables like collected answers unless you configure that logic inside the Workflow itself.

• Information is not passed back from the Workflow to the AI Agent. The AI Agent does not “wait” for the Workflow’s result and cannot react to its individual steps.

• When testing an AI Agent, you’ll see a Workflow {workflow name} Started event in the Test AI Agent panel when a Workflow is triggered. Unpublished Workflows will not produce events in the test panel.

Handle Calls

Plan eligibility: All users can test call behavior in the AI Agent panel, but only Workspaces on the Advanced plan and above can publish and use call handling in live conversations.

What this action does:

Let your AI Agent automatically answer and handle inbound calls for you. When this action is enabled, the AI Agent speaks naturally with callers — greeting them, responding to questions, and ending the call politely once the conversation is complete.

Language support
- Voice AI Agent supports 32 languages via the ElevenLabs Flash v2.5 model.
- By default, the agent speaks English.
- During a call, the AI Agent can automatically detect and switch to the caller’s spoken language, provided the language is one of the supported 32.

Voice behaviour
- All voices are technically multilingual.
- Languages listed under a voice indicate optimization for accent, pronunciation, and naturalness, not hard language limits.
- Using a voice outside its optimized languages is supported, but may result in pronunciation differences, accent variation, and reduced naturalness.

When to use:

Use this action when you want your AI Agent to handle simple inbound calls — such as greeting customers, answering basic questions, or collecting key details before follow-up by your team.

This is especially useful for after-hours or overflow coverage when human agents aren’t available.

Advanced settings:

• First greeting message: Write a custom greeting, or include consent if you record calls. If you leave this field empty, the system will use a default greeting message.

• Choose AI Agent voice: Pick from a curated list of natural-sounding voices (e.g., Engliah, Spanish, Arabic).

• Allow AI Agents to record calls: Record calls based on your Workspace’s call recording setting.

Click Test AI Agent and use the phone icon to simulate a call before publishing.

Best practices:

• Keep instructions short and clear — your AI Agent’s responses should sound natural and conversational.

• Include consent language in your greeting if you plan to record calls. For example:

  - Hi, thanks for calling. I’d like to record this call to better assist you — is that okay?

• Test before publishing to ensure the chosen voice and greeting sound natural.

• For brand consistency, match your AI Agent’s voice tone with how your business interacts via chat.

• Keep call interactions simple — avoid asking multiple questions in a row or long scripted flows.

Follow the recommended instruction structure

For best results, write your call instructions using the same structure applied by the Optimize button. This ensures your AI Agent clearly understands the call’s purpose, stays within supported capabilities, and responds consistently during live voice conversations.

Use the template below as a reference and adapt the content to your use case.

# CALL CONTEXT
You are handling an inbound voice call.
This call is informational only and does not support bookings or confirmations.

# AGENT ROLE
You help answer general questions about cars, models, and features.
You do not handle scheduling, reservations, or confirmations.

# COMMUNICATION STYLE
- Speak in a friendly and helpful tone.
- Respond naturally and conversationally.
- Keep the conversation focused on information, not actions.

# WHAT YOU MUST DO
- Answer general questions about models, specifications, availability status, and features.
- Respond to preferences the contact naturally shares, such as budget, colour, or style.
- Explain the process of how a booking can be made, such as booking through the website, without performing any booking tasks yourself.
- Continue the conversation in a helpful and informative manner.

# WHAT YOU MUST NOT DO
- Do not ask for phone numbers for booking or confirmation.
- Do not ask for dates or times for viewings or appointments.
- Do not collect details for scheduling, reservations, or confirmations.
- Do not suggest available slots, next openings, or viewing times.
- Do not say or imply that a booking is confirmed or reserved.
- Do not promise that someone will contact the caller or follow up.
- Do not assign the conversation or escalate the call.
- Do not lead the contact into booking steps.

# HOW TO HANDLE COMMON REQUESTS
If the contact asks to book, confirm, or check viewing dates or times:
Say, “I’m not able to make or confirm bookings on this call, but I can help answer any questions about our cars.”

Limitations

• No call transfer or routing (coming soon). The AI Agent cannot transfer calls or route them to human agents or other teams, and once it answers a call, it handles the conversation until the call ends.

• Inbound calls only. Outbound calling is not supported, and the AI Agent can only answer incoming calls.

• No human takeover during a call. Human agents cannot join, interrupt, or take over a call while the AI Agent is handling it.

• AI Agent must be assigned before the call starts. If the AI Agent is assigned after an incoming call begins ringing, it will not pick up the call.

• Actions apply only after the call ends. The AI Agent can update Contact fields, Lifecycles, or trigger other actions only after the call has ended. Actions cannot run during an active call.

• Call duration is limited. Each call is capped at a maximum of 3 minutes.

• Testing differs from live usage. All users can test call behavior in the Test AI Agent panel, but publishing and handling live calls requires an Advanced plan or above.Upgrade now.

Writing Instructions to Trigger Actions

Good AI Agent instructions directly lead to clear, actionable responses. When writing instructions, keep these best practices in mind:

• Clearly specify conditions and desired outcomes. For example:

- If the customer asks about pricing, reply with our standard packages and assign the conversation to @Sales Team.

• Use straightforward, simple language. Avoid ambiguity.

• Anticipate common scenarios and build clear, simple instructions for each action.

• You can also define what to tell the customer when an action is used. For example:

- Once booking date and time are collected, reply to the Contact: "Thank you for your booking. We look forward to seeing you on $contact.bookingdate at $contact.bookingtime."

Timing & Action Behavior

• AI Agents always perform actions before generating a reply.

• Actions like “Close Conversation” or “Assign Conversation” take effect once a scenario is received. For example, if the Contact said there are no more questions, AI Agent will first close the conversation, and then generate a thank you note.

How actions work with audio messages

AI Agents can now understand audio messages shared by Contacts in real conversations.

When an audio file is received:

• It is automatically transcribed into text so the AI Agent can understand its content.

• The AI Agent processes the transcribed message and can perform relevant actions, such as assigning the conversation, updating Contact fields, or closing it, based on what was said.

• You can view transcripts of audio messages sent to AI Agents below the message bubble.

Important notes:

• Audio messages that are sent before a conversation is assigned to an AI Agent will not be taken into context and will not be displayed.

• Transcriptions may not show up instantly. Refresh your browser or switch to another conversation and back in to view transcript.

Supported audio and video file formats:

• Audio: MP3, WAV, FLAC, M4A, OGG/OGA, Opus, AAC, AIFF, WEBA

• Video: AVI, M4V, MKV, MOV, MP4, MPEG, MPG, WEBM, WMV, 3GPP

Chaining Actions Together

You can combine multiple actions in one scenario with common examples below.

For chained actions (For example, 1️⃣ Update Lifecycle → 2️⃣ Update Contact Field → 3️⃣ Close Conversation), make sure to place them in “Instructions” and ensure that the actions are turned on.

Examples:

Update Contact Field + Assign Conversation

- Collect and save Phone Number, then assign the conversation to a @Sales Team by round robin.

Update Lifecycle Stage + Close Conversation

- Update Lifecycle to "Issue Resolved", and close the converstion by summarising the customer issue and the solution provided.

Best practices:

• Keep it logical and goal-driven: Only chain actions if they naturally follow one another (e.g., qualify a lead → close the conversation → assign to Sales).

• Keep the order clear: Place actions in a sequence that reflects how a real conversation would flow — don’t assign a conversation before collecting the key info.

• Don’t overload one scenario: Stick to 2–3 actions max. Too many chained actions can make it harder for the AI Agent to predict and troubleshoot.

• Be explicit in your instructions: Clearly state each step, rather than relying on the AI Agent to “guess” what comes next.

What to avoid:

• Vague instructions

- If it feels like the issue is resolved, maybe update Lifecycle or something.

Problem: AI Agent won’t know which Lifecycle stage to use or when to trigger the action.

• Contradicting actions

- Assign to @Sales Team, then unassign if not needed.

Problem: This creates confusion for the AI Agent and leaves the conversation in an undefined state.

• Overloading one scenario

- Update Lifecycle to Qualified, collect phone number, update Contact Field with location, assign to @Sales Team, assign again to @Support Team, and close the conversation.

Problem: Too many steps in one chain make the flow unpredictable and hard to debug. Best to split them and follow natural order. For example: First, ask for phone number, then save it. Second, update Lifecycle to “Qualified", etc

Avoiding Workflow & Assignment Conflicts

When setting up actions, ensure they don’t conflict with your existing Workflows or manual team assignments:

• Check Workflow assignments: If you have Workflows assigning conversations automatically, ensure these don’t override your AI Agent’s assignments unintentionally.

• Default AI Agent considerations: If using a default AI Agent, ensure its actions align with any automated Workflows that trigger at conversation open.

Best practice: Regularly review your Workflows alongside AI Agent settings to avoid conflicts and unexpected behavior.

General Limitations

AI Agents are powerful, but there are some platform-level limitations to be aware of. Use the recommended workarounds below to ensure your Agent behaves consistently and reliably.

Conversation visibility

AI Agents have limited visibility into conversation metadata, history, and internal context.

Channel and source metadata:

AI Agents cannot see channel-specific details, such as:

• Which channel the Contact is connected to

• Where the Contact came from (e.g., Ads vs Comments)

• Which TikTok Ad the Contact came from

This means AI Agents should not rely on channel or source metadata when making decisions.

Recommended workaround:

If your use case depends on knowing where a Contact came from, use a Workflow to capture that information and store it in a Contact field.

For example:

• Create a Workflow with a Conversation Opened trigger

• Add conditions based on channel or source (e.g. TikTok Ads)

• Update a Contact field such as Lead Source or Campaign Name

You can then instruct your AI Agent to reference that Contact field in its instructions.

This approach allows AI Agents to use source-related context reliably, even though they can’t access channel metadata directly.

Assignment history

AI Agents cannot view assignment history or past assignment actions, including whether another AI Agent assigned the conversation to them.

This means:

• If AI Agent 1 assigns a conversation to AI Agent 2, AI Agent 2 does not know it has just been assigned.

• AI Agent 2 cannot distinguish between messages sent by AI Agent 1 vs. messages sent by a human agent.

• As a result, AI Agent 2 may start responding without knowing the conversation context or may repeat actions unnecessarily.

Recommended workaround:

If you are chaining multiple AI Agents, include a clear handoff message that signals the next Agent to start.

Example: In AI Agent 1’s instructions

Before assigning the conversation to AI Agent 2, send a message: ‘I’m transferring you to AI Agent 2 who will help you next. Then assign to AI Agent 2.

Example: In AI Agent 2’s instructions

Only start your top-level flow when you see the message: ‘I’m transferring you to AI Agent 2 who will help you next.’ When you see it, greet the user and begin your qualification/support flow.

This ensures:

• AI Agent 2 knows the handoff happened

• AI Agent 2 starts at the right moment

• Users get a smooth transition between agents

Internal context and message history

AI Agents cannot see internal comments and can only access the most recent 20 messages in a conversation.

Because of this, AI Agents should not rely on channel origin, internal notes, or older conversation context when making decisions.

Recommended workaround:

Store important context in Contact Fields or Lifecycle stages, and instruct the AI Agent to reference those instead.

For example:

• Use Workflows or human processes to keep fields like Plan Type, Customer Status, or Issue Category up to date

• Use Lifecycle stages (e.g., New Lead, Qualified Lead, Customer) to guide how the Agent responds

• In your AI Agent instructions, explicitly reference these fields (e.g., “If Lifecycle is Customer, prioritize support-related responses.”)

This ensures the AI Agent can make consistent decisions using reliable, structured data, even with limited access to older messages or internal comments.

External capabilities

AI Agents currently operate within respond.io and do not have access to external tools. As such, they can’t perform web searches.

Any logic that requires external data or API calls should be handled through Workflows or Integrations instead.

Note: Support for HTTP requests is coming soon.

Model configuration

AI Agents do not support manual temperature settings because newer AI models no longer expose temperature controls.

If you want to adjust the AI Agent’s “creativity” (similar to temperature), the best approach is to prompt for it directly in your Instructions. For example, you can guide the AI to behave more creatively or more strictly depending on your use case.

High creativity (higher “temperature”):

Respond warmly and creatively in the customer’s tone.

Low creativity (lower “temperature”):

Answer the customer’s question using only the provided company knowledge sources and the conversation history. Do not guess or add new information.

This gives you better control over tone and consistency without needing a dedicated temperature setting.

FAQ and Troubleshooting