Frameworks · AI Don't Fix Stupidity

The Earned Autonomy Ladder

The Earned Autonomy ladder is one instrument with two views: an org-wide view for how far a whole organisation or department has come, and a per-capability view for how much a specific capability is trusted to do on its own.

No organisation earns trust by decree. The two views move independently: a Scaling department can still run a new capability on the most supervised rung. A capability climbs one rung at a time on evidence (golden-set evaluation, sustained agreement, complete audit trail, accepted residual risk) and can be demoted on the same evidence. Planning defaults: a golden set of roughly 150 to 200 cases per capability and a permanent sampled human-review floor of about 1 to 3%, both adjusted for the consequence class of the work. The riskier the tools in scope, the more evidence a capability needs before it climbs.