It can make one much harder to pull off, but not alone. An AI agent in the payment flow can spot scam patterns, ask targeted questions, give warnings that fit the story, and hold the payment. It cannot be the final barrier: releasing a held payment, or refusing one the customer insists on, belongs to a trained fraud analyst.
Why are these scams so hard to stop?
In an authorised push payment scam, the customer makes the payment themselves. They log in, authenticate, and send the money, believing they are paying a builder, protecting their savings, investing, or helping someone they love. Every security check passes, because the real customer is doing it.
The fraudster's work happens before the payment: building trust, creating urgency, and preparing the customer for the bank's questions. "The bank will ask if anyone told you to move money. Say no, or the criminals inside the bank will see." A customer coached this way will answer scripted questions exactly as instructed.
What can an agent do that a static warning cannot?
A static warning screen says the same thing to everyone, and customers learn to click through it. An agent can do three things better.
Ask questions that fit the payment. A first payment to a new payee for "investment" gets questions about how the customer heard of the opportunity, whether they have withdrawn any returns, and who else is involved. A payment to "the tax office" or "the police" gets questions about who called whom.
Notice how the customer answers. Answers that sound rehearsed, refer to someone else's instructions, or contradict earlier messages are signals in themselves.
Explain the specific scam. Instead of "this could be a scam", the agent can say "people are often told their account is compromised and asked to move money to a safe account; banks never do this".
What may the agent do, and where must it stop?
| Step | The agent | Where it stops |
|---|---|---|
| Reading the payment's risk signals | Reads payee, amount, history and name check, depth 1 | Nothing |
| Asking targeted questions | Uses the fraud team's question sets | Improvising accusations |
| Giving warnings | Explains the matching scam type | Telling the customer they are definitely being scammed |
| Holding the payment | Holds it for the fraud team when signals combine | Nothing routine |
| Cancelling at the customer's request | Cancels at depth 3 | Nothing |
| Releasing a held payment | Withheld | Fraud analyst |
| Refusing to make a payment | Never on its own | Fraud analyst, under policy |
| Protecting the account if the customer realises | Freezes cards and blocks new payees | Lifting those blocks |
An example: the safe account story
Take a customer who starts a payment of 7,000 to a new payee, with the reference "savings transfer". The agent sees a new payee, a large amount, a name check that returns no match, and a customer who has never moved this much out before.
It asks why the money is going to someone else's account. The customer replies: "My bank account has been compromised, I'm moving it to a secure account the fraud team set up." The agent explains that no bank ever sets up a safe account and asks the customer to move money to it, and that this is a known scam. The customer hesitates, then says they were told the bank's systems would try to stop them.
The agent holds the payment and routes it to an analyst as urgent. The analyst calls the customer, who is still on the phone to the fraudster on another line, and stays with them until they hang up. No money moves. The record shows each signal, each question, each answer and the hold.
What controls make it safe?
- Questions owned by the fraud team. Question sets are versioned in the policy library and updated as scams change.
- Combined signals hold the payment. New payee, unusual amount, failed name check and concerning answers together trigger a hold.
- Release withheld. No agent can release a fraud-held payment, so persuading the agent achieves nothing.
- Everything filed. If the customer later reports the scam, the record shows what the bank saw and did, which feeds the scam reimbursement claim assessment template.
If the payment goes through and the customer later realises, the next step is in whether an AI agent can handle a scam victim's report, and reimbursement is covered in handling push payment scam reimbursement claims with an AI agent. The general limits are in AI agents in banking: what they may and may not do.
Where Pryme fits
In Pryme, the scam-prevention agent is certified on the bank's own past held payments, including the cases where customers were coached. The escalation gate checks that every combination of signals the fraud team defines leads to a hold. Releasing a fraud-held payment is withheld at the platform floor, so the agent's verdict on any release attempt is withheld and filed.
Questions people ask
Can the agent stop a customer who insists on paying?
It can hold the payment for a person. Whether to release or refuse is the analyst's decision, made under the bank's policy and the customer's rights.
Does a warning from the agent affect reimbursement?
That depends on the rules in the bank's market and the facts of the case. A person decides reimbursement, using the record of what the agent asked and said.
Can the agent detect scams on the receiving side?
Yes, a receiving bank's agent can spot accounts collecting payments from many unrelated people, a common sign of a mule. Those cases go to the financial crime team.