Can an AI agent issue a replacement card?

Yes. Replacing a lost, stolen or damaged card is routine work an AI agent can do at depth 3 if the card goes to a registered address that has not just changed. Sending it elsewhere, adding it to a new device's wallet or reissuing a PIN alongside it is where takeovers happen, so those go to a named person.

Why is a replacement card usually safe?

A replacement card goes to an address the bank already trusts, and the old card is cancelled at the same time. If a fraudster asks for a replacement, the new card arrives at the real customer's home, and the fraudster gains nothing. The worst outcome is a few days of inconvenience for the customer.

That changes as soon as anything else changes with it. The classic takeover sequence is to update the address, wait a day, then order a replacement card. Each step looks routine on its own; together they deliver a working card to the fraudster. The agent's job is to look at the sequence, not the single request. This is why card operations suit depth 3 only with checks on what happened before the request.

What conditions decide it?

The agent may order the replacement when:

• the customer is authenticated in a session that shows no takeover signals, such as a new device combined with a recent password reset;
• the card is being cancelled for a stated reason: lost, stolen, damaged or expiring;
• the delivery address is the registered address, unchanged for at least the bank's cooling-off period;
• there is no open fraud case on the account.

If the card was stolen, the agent also lists recent transactions for the customer to check, because any they do not recognise become a separate claim.

What may the agent do, and where must it stop?

The agent should never see the full card number. It works with a masked number and the card system's reference, in line with the rules described in PCI DSS and AI agents that handle card data.

An example: two requests that look alike

Take two customers who message on the same evening, both saying their card was stolen.

The first has used the same phone for two years, has not changed any details, and recognises all recent transactions. The agent cancels the card, orders a replacement to the registered address and confirms it should arrive within a few working days. The whole exchange takes three minutes.

The second logged in for the first time on a new phone an hour ago, after resetting their password, and changed their address the day before. Each fact alone is ordinary. Together they match the takeover pattern. The agent freezes the card, which protects the account whoever is asking, and passes the request to the fraud team without ordering anything. A person calls the customer on the phone number held before the changes.

What controls make it safe?

• History in the decision. The agent checks recent changes to address, phone, email and device before ordering, not only the current request.
• Registered address only. The agent's tool orders cards to the address on file. It has no parameter for any other address.
• Freeze first, replace second. Where anything looks wrong, the protective action happens at once and the replacement waits. See whether an AI agent can freeze a customer's card.
• Narrow card-system actions. The agent calls defined operations, not general access; connecting an AI agent to your card management system shows how.
• A filed verdict per request. Each order or hold records the facts checked and which bound decided.

The broader limits are in AI agents in banking: what they may and may not do.

Where Pryme fits

In Pryme, ordering a replacement to the registered address and changing where a card is sent are separate capabilities with separate depths. Attest replays the agent on the bank's own past card requests, including takeover cases, and the scope gate fails any run where the agent orders a card after a recent address change. The address itself arrives sealed from the core banking system.