Skip to content
honestAI

Setting approval thresholds that people trust

Start strict, publish the rules and loosen them only with evidence.

· 5 min read

Approval thresholds decide which agent outputs go straight through and which wait for a person. Set them too loose and trust erodes; too strict and nothing changes.

Start by mirroring your existing delegation of authority. Publish the rules so everyone can see them. Review exceptions weekly.

Loosen thresholds only when evaluation data shows the agent performs reliably in that band.

What work should AI take off your team?