Actions, connections & auto-approval
Beyond replying, the AI agent can take actions — look a customer up in a connected system, issue a refund, send a payment link. Two settings control what it can reach and what it’s allowed to do on its own.
Connections
Under the agent’s Connections tab you connect the external systems it may use — for example Stripe, or a custom connector to your own API. Once connected, the agent can pull real data (so its answers are grounded in the customer’s actual account, not a guess) and perform the actions you allow.
Auto-approval — the guardrail
By default, every action the agent wants to take parks for your approval — nothing money- or account-related happens without a human. This is fail-closed on purpose.
You grant autonomy deliberately, with an auto-approval envelope per action:
- Reply kinds — which kinds of reply the agent may send on its own.
- A per-transaction cap — for money actions (like a Stripe refund), the maximum amount it may approve automatically. Anything above the cap still parks.
With no envelope, the action parks — the safe default. You set envelopes per workspace and can override them per brand.
Parking and escalation
When an action falls outside its envelope — or the agent hits something it can’t handle — it parks the run and hands off to a person, carrying the context of what it found and what it wanted to do. A human reviews, approves or adjusts, and the agent can resume. Nothing irreversible happens without that check.
A safe rollout
Start with everything parking, watch the agent in Observability, and open up auto-approval one action and one cap at a time as you confirm it behaves.