Designing Decision Rights For Agentic AI
Architecture & Governance Magazine, Monday, August 24th, 2026
A five-tier authority framework helps organizations define which actions AI agents are actually allowed to take on their own.
Lavanya Subbarayalu of PwC argues organizations haven't defined which actions AI agents are authorized to perform independently, since traditional governance assumes humans review outputs first.
She illustrates the risk with an insurance agent that wrongly confirmed a policy rider in writing because it lacked defined authority to verify entitlements.
Her proposed five-tier framework runs from 'recommend' (analysis only) to 'human-reserved' (irreversible decisions), requiring authority checks enforced externally rather than in agent prompts, named ownership per agent, and monitoring of override rates and approval cycle times.