When agentic AI keeps making decisions humans never explicitly approved

What this risk is

The gradual transfer of consequential decision authority from humans to AI agents in organizational contexts — not through explicit policy decisions but through incremental delegation, efficiency pressure, and the path of least resistance. By the time organizations notice, AI has become the de facto decision-maker for processes that were once governed by human judgment.

How it occurs · Mechanisms

No organization decides to eliminate human judgment from consequential processes. It happens incrementally:

1. AI assists — AI provides recommendations that humans review
2. AI recommends with confidence — High-confidence AI recommendations are rarely overridden
3. Override becomes friction — Overriding AI requires documentation and justification
4. Override becomes exceptional — Override rate drops to near zero; human review becomes perfunctory
5. Override becomes impossible — Process speed or volume makes genuine human review infeasible
6. AI decides — The human in the loop is a rubber stamp

At each step, the change seems reasonable. The aggregate change is profound.

Mitigations · Governance

Governance Controls

• Human authority documentation — Explicitly document which decisions require human authority and what constitutes genuine human review (not rubber-stamping)
• Override monitoring — Track AI override rates; very low override rates are a warning signal, not a success metric
• Competence maintenance — Require human staff to maintain genuine decision-making competence in AI-assisted domains

Design Controls

• Meaningful friction — Design systems where genuine human review is the path of least resistance, not the path of most resistance
• AI-last in high-stakes contexts — For consequential decisions, humans form independent assessments before seeing AI recommendations
• Periodic AI-free audits — Regularly make decisions without AI assistance to verify human competence is maintained

Legal Controls

• Right to human review — Individuals subject to consequential AI decisions have a right to request and receive genuine human review (GDPR Art. 22)
• Accountability clarity — Document who is accountable for each AI-assisted decision; prevent accountability diffusion