Banca d’Italia deputy director calls for consumer control over agentic AI

Sergio Nicoletti Altimari, deputy director general of Banca d’Italia, has set out consumer protection principles for financial services using agentic artificial intelligence. Speaking at the 13th Joint European Supervisory Authorities Consumer Protection Day in Rome on 4 November, Altimari said consumers must retain control when AI systems plan or execute actions on their behalf. His opening speech sets out the regulator's expectations for firms developing or deploying such systems.

“Delegating tasks to AI should not mean losing control over financial decisions,” Altimari said. He called for protection to be built into services from design through deployment, and said innovation should ultimately be judged by the outcomes it delivers for consumers rather than by technological sophistication.

The controls Altimari wants

Consumers should be able to understand and define an agent's tasks and limits, monitor and challenge its actions, stop an operation and revoke an instruction. For financial service providers, Altimari said safeguards must include:

- checks on identity, consent and authorisation for material decisions;
- traceable and auditable agent actions;
- human oversight proportionate to the potential effect of automated decisions on consumers' financial well-being; and
- clearly allocated responsibilities.

The principle applies even where a system acts within apparently valid permissions. Altimari said responsibility towards consumers must remain with people and institutions: “Tasks can be delegated to AI, but responsibility towards consumers must remain firmly with people and institutions.”

The speech also links these controls to transparency, understandable explanations, challenge rights and redress. Consumers should be able to establish how a decision affecting them was made and contest it when something goes wrong.

For compliance teams, this moves the focus beyond model accuracy or cybersecurity. A firm will need to be able to evidence what authority a consumer granted, the limits attached to it, which checks were performed and what the agent actually did. In Atlas's view, the practical implication is that delegated authority should be designed as a controlled, reversible permission rather than as a broad transfer of decision-making power.

Payments and agent-to-agent commerce

Altimari used a payment example to expose the accountability problem. If an agent makes a transfer and the consumer disputes it, firms and supervisors need to determine whether the agent acted within the authority granted and who bears ultimate responsibility.

The issue becomes more difficult when a consumer's agent interacts with an agent representing a firm. Altimari referred to the rise of “agentic commerce”, where automated systems could produce unintended outcomes or facilitate more sophisticated fraud and manipulation. A transaction may use valid credentials and remain within delegated limits, yet still fail to reflect the consumer's actual intention.

Altimari said AI could reduce search and transaction costs, help consumers compare and select products, manage their finances and interact with providers. It could also improve fraud detection and complaint handling. But those benefits depend on reliable data, sound models and effective safeguards. Consumers, he said, should not be expected to carry the protection burden through financial education and digital literacy alone.

Complaints as a supervisory input

Altimari assigned supervisors a role in testing whether firms' governance arrangements preserve consumer control and accountability in practice. Complaints and consumer reports should help identify fraud, manipulation and misalignment between an agent's actions and the user's actual intent, with particular attention to vulnerable consumers.

He also warned of possible perimeter gaps. Consumers may use AI tools supplied by entities outside financial regulation, while those tools influence financial decisions or interact with regulated firms. Cooperation between authorities responsible for implementing the EU AI Act could help identify and address those risks.

The speech gives no implementation date or formal compliance deadline. Its immediate value is as a control framework for firms assessing agentic use cases: they should be able to show who authorised an action, what limits applied, how the action was recorded, who remained responsible and how the consumer could stop or challenge it.