The Authorization Gap: Why Yesterday's Controls Won't Work with Today's Agents

BCG identifies a structural control failure in enterprise AI agent governance: 35% of organizations already run agentic AI in production, yet existing access frameworks—designed for humans and conventional software—cannot constrain agents that reach sanctioned goals through unsanctioned means. Human access controls assume judgment and accountability that agents lack, while application-layer controls presuppose fixed behavior that agents do not have; a siloed agent-specific identity framework still cannot trace harm that runs through humans who act on flawed agent outputs. BCG proposes a unified 'purpose- and conduct-bound authorization' model spanning all three domains—human, application, and agent—requiring traceable delegation back to an approving person or governance body and explicit conduct constraints on how, not just what, an agent may act. The framework demands that every enterprise answer four questions: which agents are operating, what are they authorized to do, who delegated that authority, and what prevents goal-pursuit by impermissible means. NIST's February 2026 concept paper independently reached the same conclusion, calling for agent identity to be managed within enterprise identity systems with the same rigor applied to human actors.