What is dual control?

Dual control is a safeguard that requires two authorised people to act together to complete a sensitive task, such as releasing a large payment, changing a supplier's bank details or granting privileged system access. Each holds part of the authority and neither has enough alone, so one person's mistake or dishonesty cannot complete the task on its own.

Why does dual control matter once AI agents act?

Because it raises a question every bank will face: can an agent be one of the two? Our answer is no, not as a substitute for an accountable person. Dual control works because two people with separate judgement and separate accountability must both agree. An agent can prepare the task and check it against rules, which is useful. It cannot carry accountability, and a supervisor will ask who answered for the decision.

Two agents do not make dual control either. If both run on the same model with similar briefs, they share failure modes, and a trick that fools one will often fool the other. Independence is the point, and that is the sense in which EU AI Act human oversight works as a control.

Where is dual control usually applied?

• Releasing payments above a set amount.
• Adding or changing a payee, or a supplier's bank details.
• Changing limits, fees or interest rates on accounts.
• Granting or raising privileged access to systems.
• Changing an agent's own bounds, through change control for AI agents.

Access to these actions should follow least privilege as well, so only the people who need to act can. The requirements are in DORA access control and least privilege for AI agents.

What does that look like in an example?

A corporate treasury team requires two approvers for any payment above 50,000. Its agent prepares payments from the approved schedule and checks each against the purchase records. For a payment of 80,000, the agent's verdict is held, naming two approvers. The system refuses to let the same person approve twice, and refuses approval from anyone who edited the payee's details in the past month. Both approvals, and the agent's preparation, are filed with the payment.

The agent saved most of the work. The decision still took two people.

How is dual control different from maker-checker?

Maker-checker splits a task into two roles in sequence: one person prepares, a second checks and releases. Dual control requires two people to exercise authority jointly, often at the moment of action, and neither role is subordinate. In practice the terms overlap and many firms use them loosely. The useful distinction is this: with maker-checker, an agent can be the maker and a person the checker. With dual control, both authorising parties should be accountable people.

Which related terms should you know?

• Segregation of duties: no single person controls a process end to end.
• Least privilege: each person or agent has only the access its role needs.

How these fit into the wider question of how you control what an AI agent is allowed to do is covered separately, and the bank-specific limits are in AI agents in banking.

Where Pryme fits

In Pryme, an agent preparing an action that needs dual control receives a held verdict naming the people whose approval is required. The verdict and both approvals are filed in an append-only record, so it is clear later who authorised the action and what the agent contributed.