Stopping the bleeding and finding someone guilty are different jobs. In fraud, an agent should do the first: hold a suspicious payment, freeze a card, gather evidence, and put the case in front of an investigator quickly. It should not do the second. Concluding that a customer committed fraud brands a person, and that judgement stays with investigators.
Why split containment from conclusion?
The two halves of fraud work have opposite risk profiles.
Containment is mostly reversible and time-critical. A held payment can be released. A frozen card can be unfrozen. The cost of a false alarm is inconvenience and a phone call. The cost of acting slowly can be money gone for good. That is exactly the profile where agent speed helps and errors are cheap.
Conclusion is mostly irreversible and not time-critical. Deciding that a customer is a fraudster can lead to an account being closed, the customer being recorded on shared fraud databases, a claim being refused and a report being made to the authorities. Those outcomes follow the person for years. Getting them right matters far more than getting them fast.
An agent that does both blurs the line between "this looks suspicious" and "this person is guilty". Keeping them apart lets the agent act at speed where speed matters and leaves judgement where it belongs.
What does containment include?
| Containment action | Why it suits an agent | Bound that applies |
|---|---|---|
| Holding an outbound payment for review | Instant, reversible, and the window to stop it is short | The hold must reach a person within a set time |
| Freezing a card after suspicious transactions | Protective and quickly reversed | The customer is told and can reach a person |
| Blocking a new payee pending confirmation | Prevents loss from a likely scam | Release goes through stronger authentication |
| Gathering evidence across accounts and channels | Faster and more thorough than manual collection | Read access limited to what the case needs |
| Contacting the customer to confirm a transaction | Often resolves false alarms at once | Wording must not tip off a suspected perpetrator |
| Raising a case for an investigator | Puts the evidence in front of a person quickly | The case carries every fact with its source |
The guide on whether an AI agent can detect fraud in real time covers how agents fit alongside existing rules and models. The playbook on how to review card fraud alerts with an AI agent shows containment in practice for one common alert type.
Why is concluding so hard to get right?
Fraud cases are rarely clean. The same pattern can fit a fraudster, a victim and an innocent customer.
Take authorised push payment fraud. A customer sends money to a scammer believing the payment is genuine. From the transaction data, they look like the person making the payment, because they are. Whether they are a victim, careless, or complicit is a judgement about their circumstances and behaviour, often involving a conversation.
Money mule cases are similar. An account receiving and passing on stolen funds may belong to a knowing participant, a person who was deceived, or someone whose account was taken over. Each warrants a very different response.
Application fraud raises its own ambiguity: an inconsistency in a loan application may be deliberate, a mistake, or a sign of identity theft affecting the applicant. The playbook on reviewing loan applications for application fraud with an AI agent keeps the finding with a person for that reason.
An agent can surface these patterns well. It cannot hold the conversation, weigh the context and take responsibility for the finding in the way an investigator can.
What stays with investigators?
- Deciding that a customer committed, or knowingly participated in, fraud.
- Closing or restricting an account on fraud grounds.
- Recording a customer on internal or shared fraud databases.
- Refusing a reimbursement claim on the basis that the customer was complicit or grossly negligent.
- Deciding to make a suspicious activity report, which rests with the bank's nominated officer.
- Releasing a payment or lifting a block that was put in place for fraud reasons.
The last item completes the picture. Containment is the agent's to apply; lifting it after a fraud concern is a judgement, because it means deciding the concern has been resolved.
How should the runbook reflect this?
The agent's runbook should spell out both halves: which signals trigger which containment actions, what the agent tells the customer, how quickly a held case must reach a person, and what the agent must never say or do. The guide on what an AI agent runbook should contain covers the general structure. For fraud, the most important lines are the ones about tipping off and about the time limit on holds, because a hold that waits days becomes its own harm.
Where Pryme fits
In Pryme, containment actions such as holding a payment or freezing a card can earn depth 3 on your own cases. Concluding actions, such as closure on fraud grounds or lifting a fraud block, are held for a named investigator. Every containment carries the facts and their provenance into the investigator's case. See AI agents in banking: what they may and may not do.
Questions people ask
Can an agent close a mule account it has identified?
It can contain the account, for example by restricting outbound payments, and raise the case. Closure on fraud grounds is a conclusion and stays with an investigator.
Should the agent tell customers why their payment is held?
It should tell them that it is held for a check and how to reach someone. The detail of the concern is for a person to share, if appropriate, because of tipping-off risk.
What if the investigator queue is too long?
Then holds grow and customers suffer. The answer is staffing and prioritisation, with the agent ranking cases by urgency, not letting the agent conclude cases itself.