Let AI move fast without losing human judgment on the big calls

Sits in front of the high stakes actions an AI workflow wants to take. You set what needs approval: an ARR cutoff, a segment size, the kind of action, or how confident the AI was.

A webhook other workflows call to submit an action. It carries the kind of action, a one line summary, the full payload, which workflow asked, and why this one needs approval.

Emails go to marketing, record updates to RevOps, lead routing to a sales manager, and other decisions by type. The wrong approver is what turns this into a bottleneck.

An interactive message in the approver's channel with the summary, the workflow that asked, why it needs approval, a preview such as the subject line and first 200 characters, and approve or reject buttons.

Execution pauses until someone responds, four hours by default. Shorten it for a big send, stretch it for routine updates. Silence counts as a rejection and posts a notice.

Approved runs the action. Rejected archives it with the reason, tells the originating workflow to stop, and posts the reasoning to the governance channel. A timeout behaves like a rejection and is logged as one.

A webhook back to the workflow that asked, carrying the approver, the timestamp and any comment, all written to the audit trail so every AI action is traceable.

Every rejection and its reason goes into a structured log, which over time shows which AI behaviours need tighter prompts or guardrails.

Validated and published, with approvals per week, the approval rate against an 80% target, the median time to a decision inside business hours against 30 minutes, and the timeout rate, which points at the wrong approver or a dead channel.