Yes. Holding a payment that looks suspicious is protective and reversible, and an AI agent can do it within set rules. It can pause the transaction, contact the customer and gather what an investigator needs. Releasing a payment held for fraud, concluding that fraud happened, closing accounts and reporting to authorities stay with people.
Why is blocking safer than releasing?
The two directions carry different risks. A hold delays a payment. If the payment was genuine, the customer is inconvenienced, sometimes badly, but the money is still there and can be sent once checked. A release lets the money go. If the payment was a scam, it may never come back.
That asymmetry gives the agent room on one side and not the other. It can place a hold when the signals are present, because the cost of a wrong hold is delay. It should not release a hold on its own judgement, because the cost of a wrong release is loss. The agent contains; people conclude. The same split applies when a customer reports a lost card, as in blocking and replacing lost or stolen cards with an AI agent.
What are the conditions for a hold?
The agent may hold a transaction when:
- the transaction meets a rule or model threshold the bank has approved;
- the hold is temporary and has a defined review time;
- the customer is told the payment is held and how to resolve it, unless telling them would tip off a suspect in a money laundering case;
- the hold does not cut the customer off from money for essentials for longer than the bank's policy allows.
A hold on a payment the customer is making to a scammer is very different from a hold on funds the bank suspects are criminal proceeds. The first protects the customer. The second engages anti-money laundering duties, where what the customer is told is tightly controlled.
What may the agent do, and where must it stop?
| Action | The agent | Where it stops |
|---|---|---|
| Holding an outbound payment | Holds it under approved rules, at depth 3 | Holds beyond the policy period |
| Freezing a card | Freezes it on a fraud signal or customer request | Unfreezing after a fraud-triggered freeze |
| Contacting the customer | Asks scripted questions about the payment | Deciding the answers are satisfactory for a large payment |
| Reviewing history | Reads recent transactions for patterns | Using data beyond the case |
| Releasing the payment | Prepares the case for release | A person releases it |
| Concluding fraud | Never | An investigator |
| Reporting suspicion | Drafts the report | A trained person decides and files |
Reading the customer's transactions is part of the job, within limits set out in whether an AI agent can read a customer's transaction history. Card freezes follow the pattern in how to let an AI agent block and unblock cards.
A worked example
Take a retail bank. A customer who usually makes small card payments sets up a transfer of 9,000 to a new payee described as an investment platform, at eleven at night.
The bank's fraud rules flag it. The agent holds the payment and messages the customer through the app: the payment is paused for a check. It asks the scripted questions. How did they find the investment? Has anyone told them to move money quickly or keep it secret? The customer says a contact from a social media group recommended it and promised high returns within a month.
The agent explains that this matches a known scam pattern and keeps the hold. It writes up the case with the customer's answers and passes it to a fraud specialist, who calls the customer the next morning. The customer decides not to send the money.
What controls make it safe?
- Holds under approved rules. The agent places holds when the bank's rules or models trigger them, not on its own suspicion.
- Time limits. Every hold has a review deadline. A hold that drifts becomes a harm of its own.
- Release held for a person. Releasing a fraud hold is outside the agent's scope at every depth.
- Tipping-off controls. Where a case involves suspected money laundering, the agent's customer messages follow a script reviewed by the financial crime team.
- Reporting stays with people. Under the FATF Recommendations, suspicious transaction reporting is a core duty. The agent drafts and a person files, as set out in FATF suspicious transaction reporting: the agent drafts, a person files.
The wider limits for agents in banks are set out in AI agents in banking: what they may and may not do.
Questions people ask
Can the agent release a small held payment if the customer confirms it?
Only if the bank has a written rule allowing release below a set amount after defined checks, and the capability has been proven on past cases. Without that, release stays with a person.
What if the customer is angry about the hold?
The agent explains the reason in plain words and the expected review time, and offers to pass them to a person. It should not release the payment to end the complaint.
Can the agent block all payments on an account?
Blocking an account is a larger step that may stop the customer paying for essentials. It should be a person's decision, unless the customer asks for it themselves.