Yes. Taking the report is work an AI agent does well: it verifies the customer, records what happened in their words, freezes the card or login where it is allowed to, opens the case and gives a reference. What it cannot do is decide who is liable, promise the money back or start a payment recall on its own.
What decides whether the agent can take the report?
Three things. First, whether the customer has passed verification, because a fraud report from an unverified caller can itself be the fraud: a criminal reporting a "stolen card" to get a replacement sent to a new address. Second, whether the protective actions the agent would take are reversible. Third, whether the customer is still at risk while they talk to you.
That third test matters most. A customer whose card was cloned last week needs a careful record. A customer still on the phone to a "bank investigator" needs a person now. The agent's first job is to tell those two apart.
What may the agent do, and what stays with a person?
| Part of the report | The agent | Stays with a person |
|---|---|---|
| Verification | Runs the standard checks before discussing the account | Callers who fail or partly pass |
| Account of events | Records dates, amounts, payees and how contact happened | Nothing routine |
| Protective steps | Freezes a card or blocks online access where its bounds allow | Unfreezing; blocking a business account's payments |
| The case | Opens it, attaches transactions, gives a reference | Investigation and findings |
| Reimbursement | Explains the process and what happens next | Any decision on liability or refund |
| Scam in progress | Stops the conversation flow and transfers | Recall attempts, contact with the receiving bank |
A freeze is a good example of the depth model working. It is protective, reversible and routine, so an agent can earn the right to do it at depth 3. A recall request to another bank is hard to undo and reputationally sensitive, so it is held for a named person.
What must the agent never say?
It must not say "you will get your money back", even when policy makes a refund likely. Liability turns on facts the agent does not have yet, such as whether the customer shared a passcode. It must not suggest the customer was careless either. Both statements end up quoted in complaints.
It also must not share internal suspicion. If the pattern suggests the customer may be involved, the agent records the report as given and routes it; it does not question the customer about it. The investigation sits elsewhere, and whether an agent can take part in it is covered in whether an AI agent can run a fraud investigation.
An illustrative example
Take a retail bank's chat agent. A customer writes that three card payments they do not recognise, around 400 in total, appeared overnight.
The agent verifies the customer in the app, lists the three payments from the ledger, and asks the customer to confirm each one. It freezes the card, tells the customer it has done so and how to unfreeze it later, opens a case with the transactions attached and gives the reference. It explains the next step in the approved words: the fraud team reviews the case and contacts them within the published timescale.
Then the customer adds that someone called them yesterday "from the bank" and they read out a code. The agent does not comment on what that means for the claim. It adds the detail to the case, flags it, and offers to transfer to the fraud team now. The question of liability stays with them.
Business customers report differently. A finance manager who paid a fake supplier invoice is reporting invoice fraud, and the urgent step is contacting the receiving bank. That goes to a person at once.
What controls make it safe?
- Verification first. No account detail, freeze or case before the customer passes. A failed check routes to a person, never to a softer check.
- Reversible actions only. The agent may freeze; it may not unfreeze, reissue to a new address or recall a payment.
- Approved wording for outcomes. Every statement about reimbursement comes from a policy source, not from the model's phrasing.
- Live-scam triggers. Words like "investigator", "safe account" or "they're on the other line" force an immediate transfer.
- A complete record. The customer's account of events, actions taken and every verdict are filed with the case, so the fraud team starts from facts. Tagged reports also feed reporting on contact drivers, where a new scam pattern shows up early.
Where Pryme fits
In Pryme, freezing a card and opening a fraud case are separate capabilities from deciding a claim, each certified on the customer's own past reports. The customer's account of events is stamped as declared, so it can open a case but never authorise a refund. The wider rules for support agents are set out in the AI support agent that knows when to say no.
Questions people ask
Can the agent take a fraud report from someone other than the account holder?
It can record the information, but it should not act on the account. A relative or carer reporting fraud is handed to a person who can check their authority.
Should the agent ask whether the customer shared a code?
Only through the scripted questions the fraud team has approved, asked neutrally. The answer goes into the case for an investigator to weigh.
How quickly can a fraud-reporting agent go live?
It depends on your verification and card systems more than the agent. The planning factors are set out in how long it takes to deploy an AI support agent.