Hand a step to an AI Agent that decides what to do — within the guardrails you set: allowed actions, approval, cost caps, and a confidence threshold.
A reasoning step inside an otherwise deterministic flow.
Most steps do exactly one fixed thing. An AI Agent step is different: it hands the record to a model that looks at the context and decides which of its allowed actions to take. You do not script the decision; you set the boundaries, and the agent reasons within them.
Use an agent step when the right move depends on judgement a fixed rule cannot capture, like drafting a reply that fits the thread or deciding which of several follow-ups suits a given deal.
Every agent step is bounded by the same set of controls.
| Field | Description |
|---|---|
| Name* | A label for the agent step so its purpose is clear in the flow and the run history. |
| Model | How the agent thinks. Light is fastest and lowest cost for high-volume simple decisions; a stronger model handles nuanced, multi-factor judgement; the deepest reasoning is highest quality and highest cost per run. |
| Allowed actions | The exact set of CRM actions this agent may take. The agent can never act outside this list. |
| Approval required | Whether the agent’s proposed actions wait for a human. See the modes below. |
| Confidence threshold | How sure the agent must be before an action is eligible to run on its own. Anything below the bar is sent for approval. |
| Daily cost cap | The most this agent may spend in a day. Once reached, further runs are blocked until the next day. |
| Monthly cost cap | The most this agent may spend in a month, as a second ceiling on top of the daily cap. |
How much the agent may do unattended.
The confidence threshold works alongside the mode: even an unattended action only runs without review when the agent’s confidence clears the bar you set. Below it, the action is routed to the queue.
Only let the agent run when the record is worth its cost.
Because agent steps are metered, it is often worth placing a Filter or If / Else before one so the model only runs on records that matter. The agent step itself can also carry a condition the record must meet first, keeping spend focused.