Do AI sales agents need human approval before they send?
Last updated
Not for everything, and approving everything is how approval stops meaning anything. Approval belongs to a short list of moments — a deal outside its usual shape, a line nobody stands behind, a promise a person must make — and the rest should go out with a record instead.
A person approving every message reads the first fifty carefully and the next five hundred not at all. The queue clears, nobody is reading, and the approvals are now a signature on things nobody saw. That is worse than no approval, because it looks like oversight.
Where approval earns its place
| send with a record | ask first |
| a line from your register that is current | a line that is overdue, disputed or unowned |
| a price inside a range somebody owns | a number outside it |
| a follow-up inside the limits you set | a promise with a date in it |
| a standard term | a variation on one |
The split is the point. Most companies control who signs and leave the shape of the deal open. Approval should sit on the shape.
Approvers run out
Every ask spends somebody's attention, and attention does not grow because more agents are asking. A register full of gaps sends more to approvers, so the fastest way to need fewer approvals is fewer gaps: owners on lines, dates kept current.
And approval never disappears entirely. As an agent earns room, less of what it does is reviewed. Some of it always is, because a share nobody looks at is how drift goes unseen.
What to decide, and write down
- Which moments need a person, in terms of the deal and the line, not the agent.
- Who that person is, by role, read from your own directory.
- What silence means. An ask nobody answers is not a yes, and the agent should say I'm checking rather than wait for ever.
What this does not do
Reynforce records each ask and who it went to. Recording the approver's answer is not built yet, and a page saying otherwise would be the first untrue thing on it.
/roll-out-ai-agents-in-sales-without-legal-blocking-it.md covers the sign-off that comes before any of this. /what-should-an-ai-agent-be-allowed-to-promise.md lists the permissions.