Skip to main content
How-toBy Lloyd Pilapil

When should an AI agent ask for human approval?

An AI agent should ask for human approval when an action is hard to undo, reaches a customer or moves money, or falls below the confidence the team agreed. Low-risk, reversible work can run on its own, with a log a person can review.

Approval is a design decision that has to be made per action. If the agent asks every time, it saves nobody any time. If it never asks, people stop trusting what it produces. Agentic experience (AX) design draws the line for each action, based on what happens when the agent is wrong.

Three levels of human control

LevelWhen to use itWhat the person does
Review a recommendationThe agent prepares a decision a person makes anywayReads the reason and decides, such as which lead to call
Approve an actionThe action is hard to undo or reaches a customerSays yes before it runs
Stop and reverseThe action runs on its own but could go wrong at scaleWatches a staged rollout and can stop it or roll it back

Signals that an action needs approval

  • It cannot be undone, or undoing it is expensive.
  • It changes what a customer sees or receives.
  • It moves money or commits the business to a price or a date.
  • The agent is less confident than the threshold the team set.
  • The workflow is new, so its first runs deserve a person watching.

How our products draw the line

  • Vector routes leads on its own and shows the reason. Operators can override a route, and a daily spot-check report lets the team review filtered leads.
  • Hive brings in a person when an agent is unsure, logs every handoff, and stages releases so a supervisor can stop them or roll them back.
  • Radar shows the evidence behind each finding and a suggested fix, and the owner chooses what to change first.

If the interface tells a person they will review an action before it runs, the application has to enforce it. A review step that only exists on screen gives a false sense of control.