The approval gate

Always gated. Mail and messages to a customer. Opportunity and deal stage changes. Renewal and close dates. Anything a customer will read or that finance will later report on. Can run unattended. Reversible internal writes: activity capture, internal notes, hygiene sweeps, an alert to your own Slack channel. Each still leaves a receipt. Never automated. Anything the customer has not scoped. An agent has no standing authority to invent new categories of action for itself.

Because the failure mode is asymmetric. A model that drafts a slightly wrong internal note costs a few seconds. A model that emails the wrong pricing to a customer, or silently moves a close date that a board deck is built on, costs something you cannot take back with a retry. Model quality reduces how often the bad case happens. It does not change what happens when it does. A gate does, which is why the gate is structural here rather than a toggle in a settings page.

It would, if approving were as expensive as doing. It is not. Reviewing a drafted action is reading one screen and deciding yes, edit, or no. Producing that same action by hand is opening four tools and retyping what you remember. The design target is a rep clearing the day's drafted actions in about ninety seconds, in one place, rather than context-switching for the rest of the afternoon.

The action, the systems it will touch, and the moment in the call it came from, quoted. That last part matters more than it sounds. An approval you cannot trace back to something the customer said is a guess with extra steps. Illustrative, not a customer record.

Nothing is written, and the rejection is recorded alongside the draft. That record is useful: a pattern of rejections on one kind of action is the clearest signal that the build behind it is wrong, and it is the input to the next iteration. Rejections are not failures of the system. They are the system working.