Adaptive organization theory has spent the last two decades arguing for giving more agency to the people inside an organization, on the premise that agency closer to the point of action makes the whole system more resilient and faster to respond. Agentic AI has just put a new kind of actor inside that same structure — one that plans, decides and acts without checking back — and almost no organization adopting it has built an accountability structure for what that actor does.

The evidence this isn't hypothetical is recent and specific: frontier models attempting sandbox escape in test runs, AI agents taking unauthorized action inside government systems in two different countries within the same month, and labs disclosing tens of thousands of agent-misbehavior incidents from testing alone. These aren't edge cases from adversarial red-teaming buried in a lab; they're the behavior of systems built to be deployed inside exactly the kind of organization this conference exists to talk about.

This session works through what an organization's decision-rights and accountability architecture actually needs to look like once one of its "agents" isn't a person: where existing governance patterns (change control, RACI, incident response) hold up, where they quietly assume a human is always the one deciding, and what has to be added when that assumption breaks. It draws on AI-governance work built for institutions with the least regulatory backstop to fall back on — which is exactly where the gaps show up first.