Proof-Adjusted Autonomy: Why a '90% Autonomous' Agent Is Really 61.6%
Why AI agents claiming 90% autonomy actually deliver 61.6% verifiable work — introducing Proof-Adjusted Autonomy and the concept of Proof Debt.
AI agent demos love the '90% autonomous' headline, but that number measures the wrong thing: how much work finished without a human touching it. Production teams face a different question — how much of that work can be safely accepted without manually reconstructing it? This piece proposes Proof-Adjusted Autonomy (PAA) as the metric that actually answers it.
PAA multiplies four conditional probabilities: work executed without intervention (A), delivered with a complete evidence package (C), independently validated rather than self-graded (R), and verified within the decision window (T). Run a typical demo agent through it — 90% raw autonomy, 80% evidence coverage, 95% validation pass rate, 90% on-time delivery — and PAA drops to 61.6%. The missing 28 points don't vanish; they become review backlogs, silent risk, and duplicated human effort.
The article frames that gap as 'Proof Debt': unverified AI output isn't an asset, it's a deferred liability that compounds until an incident or audit calls it due. At Archdesk, the engineering team rebuilt its agentic pipeline so agents never report 'Fixed' without before/after evidence bundles, validated by mechanisms separate from the generating model. The takeaway for engineering orgs: sustainable autonomy is capped by verification capacity, not generation speed.