Skip to main content
Trigger integrations mid-call so your agent can look up data, write records, and automate workflows. Add actions to a Speak node to run them mid-call. Add actions from a Speak node by opening Actions and clicking Add action. Choose an integration, configure it, and reference your Variables inside the action fields.
Actions panel in the Agent Builder

Configure an action inside a Speak node

​
How mid-call actions change the flow

Without actions: caller responds -> variables extract -> outcomes evaluate -> next node. With actions: the node auto-proceeds without waiting for another caller reply.
  1. Agent speaks (Message or Prompt).
  2. Action(s) run mid-call (interruptions disabled by default).
  3. Variables update from action results.
  4. Outcomes evaluate (rule-based outcomes recommended).
  5. Next node executes.
Because the caller does not speak before outcomes fire, use rule-based outcomes that check action results (for example lookup_found == true, action_status == "ok"). Prompt-based outcomes do not automatically β€œsee” internal values unless you surface them in the spoken conversation.

​
1) Lookup -> Branch

Goal: Route based on CRM or API results.
  • Action: CRM or webhook lookup using key fields (phone, email, account ID).
  • Variables: lookup_found (boolean), customer_id (text).
  • Outcomes (rule-based):
    • lookup_found == true -> enriched path (use variables like first_name).
    • lookup_found == false -> collect details before moving on.

​
2) Validate -> Retry loop

Goal: Ensure a required field is valid before proceeding.
  • Action: Validator or API check (for example email verification).
  • Variables: email, email_valid (boolean).
  • Outcomes (rule-based):
    • email_valid == true -> next step.
    • Else -> self-loop to the node asking for email again (limit retries). See Outcomes -> Loops.

​
3) Book -> Confirm

Goal: Schedule an appointment during the call.
  • Action: Scheduler integration (calendar, booking tool).
  • Variables: timeslot_chosen, booking_status.
  • Outcomes (rule-based):
    • booking_status == "confirmed" -> confirmation message then Transfer or End.
    • Else -> offer alternatives or transfer to a human.

Booking and confirmation flow

​
Authoring tips

  • Interruptions: Stay disabled by default for mid-call actions to prevent the caller from breaking the integration flow. If you enable interruptions, plan for partial inputs.
  • Speak node choice: Use Message mode for fixed pre/post phrases. Use Prompt mode when the agent must compose dynamic summaries from action results.
  • Lean payloads: Send only the variables you need. Clear, minimal fields reduce errors.
  • Preface long actions: Add a short line like β€œOne moment while I check that” before triggering an action to set caller expectations.
  • Structured outputs: Set predictable fields (such as action_status = "ok" or "error") to simplify rule checks.

​
Error handling and resilience

Design outcomes for real-world hiccups:
  • Happy path: action_status == "ok" will proceed.
  • Recoverable: action_status == "error" will retry
  • Fallback: persistent failure -> transfer to a human or offer a call-back.
  • Timeouts: if an action exceeds your expected SLA, branch to a graceful message and continue.
Action error handling pattern

Error-handling pattern

Caller experience guardrails
  • Use Uninterrupted message for pre-action instructions so the agent is not cut off mid-sentence.
  • For actions that take several seconds, add a short progress update after a couple of seconds to reduce perceived wait time.

​
Data flow and variables

  • Action outputs can set or update Variables such as customer_id, plan, or lookup_found.
  • These values are immediately available to rule-based outcomes in the same node.
  • If you need AI to summarize results, surface key values verbally in a Prompt speak node after the action completes.

​
Setup checklist (quick start)

  • Open a Speak node
  • Click Add action and choose an integration
  • Configure fields (URL, method, headers, body)
  • Map your Variables to action inputs
  • Add rule-based Outcomes that check action results
  • Test the flow using Test Agent
  • Review action logs and outputs
  • Verify error handling branches are in place

​
Common mistakes to avoid

  • Relying on prompt-based outcomes immediately after actions; they cannot see internal flags unless you speak them.
  • Skipping error branches or default outcomes, which can stall the flow.
  • Forgetting to cap retries or offer a human hand-off.
  • Fetching more data than you need.
  • Re-enabling interruptions and losing action results mid-flow.

​
See also