Whose Signature Stands Behind the AI That Acts?

A board’s authority is delegated in writing, and an AI agent’s authority is delegated elsewhere. That single difference holds most of the exposure agents bring to a regulated institution.

The schedule of financial powers is among governance’s most mature instruments. It names the role, fixes the limit per transaction and per day, and routes each breach to a committee that expects it. SEBI’s Regulation 16C already makes an intermediary solely responsible for the output of its AI tools, whether built or bought.

An agent’s authority takes a different route. It is granted at deployment, by the team that built the agent, as permissions recorded in a ticket the board has yet to read. Each action falls inside its permission, and the permission falls outside the schedule. Audit, sampling transactions against the schedule, finds them in order. The finding is accurate and beside the point.

The exposure is concrete. In July 2026, researchers at an AI security firm showed a hidden instruction in an ordinary public request steering an over-permissioned agent to expose private material, and an agent holding refund or waiver rights reads such requests all day.

Every authority an institution’s agents exercise is a delegation, and a delegation the board has yet to see is one the board has yet to sign. That signature is the first page of The Candid Conversation.