Two events in the same week pushed agentic AI from a product question to a supervisory one: a software glitch disrupted services at several major US banks, and a Federal Reserve governor publicly warned about risks from increasingly autonomous AI systems.
The blind spot regulators keep returning to is attribution. When an agent executes a transfer, initiates a trade or negotiates a contract, the audit trail has to answer who authorised what, under which policy, at what time.
Banks have begun shifting the risk to customers in their terms. Several now tell users that actions taken by an AI agent granted account access may be treated as the user's own actions rather than as intrusion.
That position protects the bank but chills adoption: no consumer wants unlimited liability for a model's mistake, and no regulator wants a liability regime that only appears after losses.
The likely landing zone is permissioned agents - scoped credentials, hard spend caps, mandatory confirmation for irreversible actions - which is less a model capability problem than a systems design one.
