Call Actions

Call Analysis decides what is extracted from a call. Call Actions decides what happens next because of it.

A caller agrees to a follow-up and a task appears on the right person's board. They book a slot and it lands in the calendar. They ask for a quote and a payment link is drafted — and held for your approval before it reaches them. That is the post-call action engine.

In this guide

• How the engine works
• Getting to the page
• The action types
• Internal vs outward actions, and why the default modes differ
• Execution modes: automatic and approval
• Confidence thresholds
• Guards — firing only when a condition is met
• Parameters — what each action needs to work
• The approvals queue
• Who can set what
• Troubleshooting

---

1. How the engine works

1. A call completes and is analysed.
2. The analysis produces signals — the outcome, the sentiment, action items, and every custom field you defined.
3. For each action type you have enabled, the engine checks the confidence of the signal and evaluates any guards you set.
4. If the action passes, it is either run immediately or queued for approval, depending on its mode.
5. Every run is logged with its result: done, skipped, or failed.

---

2. Getting to the page

1. Open the Voice AI hub.
2. Click Call Actions in the left navigation (the lightning icon).
3. The page opens at …/voice/call-actions.

The header carries three live figures — Awaiting approval, Completed and Failed — and two tabs:

Call Actions is available from Launch upward. It is driven entirely by analysis signals, so it shares the Call Analysis gate — without analysis there is nothing for the engine to act on.

---

3. The action types

Update contact has no switch. Caller details are already written to your CRM as each call is analysed, by the CRM projection. Routing it through the action engine as well would double-write the same caller, so it shows as handled automatically rather than as a toggle that does nothing.

---

4. Internal vs outward

The single most important distinction on this page is whether an action reaches the caller.

Internal actions — create task, set appointment, update contact — change your own records. If one fires wrongly, someone deletes a task. They default to running automatically.

Outward actions — send email, payment link, send form, share document, ask for a review — put something in front of your customer. If one fires wrongly, your customer receives it. They default to approval.

"Ask for a review" counts as outward even though it is delayed: it queues a message that reaches the caller's inbox or phone a couple of hours later.

---

5. Execution modes

Switching an outward action to Automatic shows a warning: nothing will be reviewed before it reaches a caller.

Client accounts cannot set an outward action to run unattended. The choice is not offered in the UI, and the mode is pinned server-side regardless of what is submitted — a rule enforced only in a form is decoration, not a control. Internal actions can be automatic on any account.

---

6. Confidence thresholds

Every enabled action carries a minimum confidence, defaulting to 0.7.

Analysis attaches a confidence to the signal behind each action — how strongly the call supports doing this. Below your threshold, the action is skipped and logged as skipped rather than run.

Raise it when an action is firing on weak evidence. Lower it when clear cases are being missed. Changing the threshold does not affect actions already queued.

---

7. Guards

A guard is a condition that must hold before the action runs. Guards are evaluated against the analysis result, including your custom fields.

Each guard has three parts: a field (a dotted path into the analysis), an operator, and a value.

Guards are written as JSON in the action's disclosure panel, together with its parameters:

{
  "guards": [
    { "field": "sentiment", "op": "eq", "value": "positive" },
    { "field": "custom_analysis.lead_quality", "op": "gt", "value": 7 }
  ],
  "params": {}
}

A guard pointing at a custom analysis field resolves through its stored shape automatically, so custom_analysis.lead_quality reads the value rather than the wrapper around it.

The JSON is checked as you type for immediate feedback, and validated again server-side — the server's check is the authoritative one.

Guards are how you stop an action being a blunt instrument. "Ask for a review" with no guard asks everybody, including the caller who complained. sentiment eq positive fixes that in one line.

---

8. Parameters

Three actions need to be told what to send before they can do anything useful. The page warns you when a required parameter is missing rather than letting you wait for a run that quietly does nothing.

---

9. The approvals queue

The Approvals tab lists every action waiting on a person, newest first. Each row shows the action type, the caller, the call it came from, and what would be sent.

From a row you can:

• Open the call — read the transcript before deciding.
• Approve — the action runs now.
• Reject — the action is dismissed and logged.

The three counters at the top of the page — awaiting approval, completed, failed — are loaded when the page opens, not when you first click the tab, so a queue that is quietly accumulating work is visible immediately.

A queue nobody looks at is worse than no automation. If you enable outward actions in approval mode, put a daily check on someone's list.

---

10. Who can set what

Both read and write go through the same API, which re-derives the caller's role server-side rather than trusting the page.

---

11. Troubleshooting

---

Related guides

• Call Analysis — the signals actions fire on
• Callbacks — the callback path, which is separate from this engine
• Conversations & Recordings — the calls behind each action
• Payments · Forms · Vault — what the outward actions send