How to let an AI agent block and unblock cards

Treat blocking and unblocking as different capabilities with different risks. Let the agent place a temporary freeze at depth 3, because a freeze protects the customer and can be undone. Let it lift only freezes the customer placed, after fresh authentication. Blocks placed by the fraud team stay with a person, and permanent cancellation is depth 4.

What does card blocking involve, and who owns it?

A card can be stopped for several reasons, and each reason belongs to a different owner. The customer may freeze it while they look for it. The fraud team may block it after suspicious authorisations. Card operations may cancel it permanently when it is lost, stolen or compromised and order a replacement. A legal or regulatory hold may sit on top of all of these.

Most card platforms hold these as separate block layers or reason codes. That is the structure to copy when you set the agent's permissions. The agent should never see "block" as one action. It should see "place customer freeze", "lift customer freeze", "permanently block and reissue", and "lift fraud block", and hold a separate grant for each.

How do you set it up, step by step?

1. Person: maps the block layers. Card operations and fraud agree which reason codes exist, who owns each, and which the agent may touch.
2. Agent, depth 1: reads card status. Every block on the card, its reason code, when it was placed and by whom, plus recent authorisations.
3. Agent, depth 3: places a temporary freeze. On the customer's request, or when a defined fraud signal fires. The freeze stops new authorisations and nothing else.
4. Agent, depth 3: lifts a customer freeze. Only a freeze with the customer reason code, and only after the customer passes strong authentication in an authenticated channel.
5. Agent, depth 2: prepares fraud-block cases. If a fraud block sits on the card, the agent refuses to lift it and drafts a summary for the fraud team.
6. Agent, depth 4: permanently blocks and reissues. On a confirmed report of loss or theft through an authenticated channel, if certified at depth 4. Otherwise the action is held for a person.
7. Agent, depth 3: confirms to the customer. An approved message saying exactly which block changed and which remain.

Why are blocking and unblocking treated so differently?

A wrong freeze is an inconvenience. The customer is stopped at a till, and the freeze can be lifted in a minute. A wrong unblock reopens the card to whoever has it. The cost of error is lopsided, so the grants are too.

The same logic splits customer freezes from fraud blocks. A customer freeze says "I want my card paused". A fraud block says "the bank believes someone else may be using this card". Lifting the second needs a person who can weigh the evidence and, often, speak to the customer about specific transactions. The related questions of whether an agent can block a suspicious transaction and whether it can unblock a locked customer account follow the same pattern.

Permanent blocks sit at depth 4 for a different reason. They cannot be undone. The card number is dead, recurring payments will fail, and a replacement is on its way. That is the same irreversibility as moving money.

What is recorded for each action?

Every attempt, allowed or not, is filed as a verdict in an append-only record. For card blocks the useful fields are:

• the card, the action and the reason code;
• who asked, through which channel, and the authentication result returned by the bank's systems;
• the five bounds checked and the bound that decided;
• the verdict: allowed, held, refused or withheld;
• for held actions, the person who released or rejected it and when.

The authentication result matters most. An unblock must rest on a result the bank's own systems produced, a sealed fact, not on anything the caller says in the conversation. A caller who claims to be the customer is making a declared claim, and a declared claim cannot authorise an unblock.

What should you measure?

• Time from report to freeze. The main benefit of giving the agent depth 3 here.
• Unblocks followed by fraud within days. The most important safety signal; review every case.
• Refused attempts to lift fraud blocks. A spike can mean social engineering aimed at the agent.
• Held permanent blocks and how long they wait. Long waits leave a lost card only frozen.
• Customers told a card was unblocked while another block remained.

Where does it usually go wrong?

One block, many layers. The agent lifts the customer freeze, tells the customer the card works, and the card is declined because a fraud block remains. The confirmation must list every block still in place.

The wrong card. Customers with several cards say "my card". The agent must confirm the last four digits or card type before acting, and record which card it chose.

Talking past authentication. A persuasive caller asks the agent to "just unfreeze it, I'm at the till". The bound is the authentication result, so persuasion has nothing to work on.

Silent consequences of cancellation. Recurring payments fail after a permanent block. Tell the customer which merchants may need new details. The full replacement flow is covered in blocking and replacing lost or stolen cards.

The wider rules for bank agents are in AI agents in banking.