Designing the approval screen for an agent: the same action shown as a narrative vs as its effect

Many agent setups add an approval step and then find that the approvals get a glance, not a read. One factor I think gets overlooked is what the reviewer is shown. Here’s a made-up example of the same action on two different screens. Version 1: the narrative The agent wants to issue a credit to Acme Corp to compensate for the billing mix-up reported in the ticket. [Approve] [Reject] It reads reasonably, and a well-written summary of a bad action reads reasonably too. The reviewer has to trust the explanation or go look things up. Version 2: the effect Action: issue_credit Account: Acme Corp (id 4821) Amount: $40.00 Changes: 1 record (account balance 0.00 -> 40.00) Last credit: none in 90 days Policy check: above the $25 auto-approve limit -> needs review Reversible: no (the customer sees it immediately) Source: ticket #1093 [Approve] [Reject] [Ask the agent why] Why the second one holds up better: The reviewer can judge it in seconds from facts, not from prose. It states reversibility and the policy trigger, which are the reasons a human is involved at all. Rejecting is as easy as approving. Where to put the gate in the first place: The variables I’d weigh: how hard the step is to undo, whether a mistake would announce itself, how expensive the review is, and how many of these happen per day. A drafted reply is cheap to discard. A refund or a deleted record is not. Spending review on the second kind, and covering the first with logs, rules and sampling, tends to keep the gate meaningful. Limits of this: A better screen doesn’t fix a volume problem. If one person sees thousands of items a day, the gate will degrade whatever it looks like. If an automated rule shares a blind spot with the agent, the check passes while the mistake repeats. Some domains require a named human sign-off regardless of risk, and then the work is making that sign-off real. I’d track the rejection rate along with sampled audits. A gate that never rejects could mean a good agent or a review that isn’t catching anything, and approval rates alone can’t tell you which. What do your approval screens show, and has anyone measured whether reviewers actually read them?