Human approval patterns for AI systems that act

“Human in the loop” is easy to promise and hard to do well. If reviewers see hundreds of near-identical approvals a day, they stop reading. If they see too little context, they cannot judge. Good approval design puts the right decisions in front of the right people with the evidence they need.

Human oversight is also a regulatory theme. The EU AI Act, for example, requires high-risk AI systems to be designed so that they can be effectively overseen by people, and NIST’s AI Risk Management Framework treats human oversight as part of managing AI risk.

Decide what needs approval

Classify actions by consequence and reversibility. Reading data rarely needs approval. Drafting is low risk because a person sends the result. Writing to systems of record, contacting customers, spending money and changing access are consequential and often irreversible. Approval effort should follow that ranking.

Five patterns

1. Draft and send: AI prepares, a person edits and sends. Best for customer communication.
2. Propose and confirm: AI proposes a specific system change, shown as a diff, which a person confirms.
3. Threshold routing: actions above a value, risk score or uncertainty go to review; others proceed.
4. Sampling review: low-risk automated actions are reviewed on a statistical sample to monitor quality.
5. Dual control: high-impact actions need two independent approvals, mirroring existing financial controls.

Design the review screen

• Show the proposed action in plain language and as the exact change.
• Show the evidence the AI used, with links to sources.
• Make rejecting as easy as approving, and capture the reason.
• Show the AI’s uncertainty honestly instead of hiding it.

Microsoft’s research guidelines for human-AI interaction remain a useful checklist here: make clear what the system can do and how well, support efficient correction and dismissal, and explain why the system did what it did.

Watch for automation bias

People tend to over-trust automated suggestions, especially under time pressure. Track approval rates and time spent per review. An approval rate close to 100% with a few seconds per item suggests that review has become a formality. Inject known-bad test cases occasionally to check that reviewers catch them.