Citi has become the latest major bank to put limits on personal finance AI agents, joining institutions that now tell customers actions performed by an agent with account access may be treated as the customer's own actions.
The contractual logic is consistent with how banks treat authorised users: if you grant access, you own the consequences. It is also a blunt instrument for a category whose entire value proposition is delegation.
Several banks have gone further than warnings, blocking or throttling agent traffic that behaves like automated account access. From a fraud-detection standpoint, an agent's request pattern is nearly indistinguishable from a scripted attack.
The resolution probably runs through standards rather than terms of service: scoped credentials with hard limits, signed intent, and machine-readable audit logs the bank can verify.
Until those exist, agent builders face a distribution problem that no model improvement solves - the rails they need are controlled by risk departments.
