Yes, though not alone. Fast scoring models flag risky payments in milliseconds; an AI agent works in the seconds and minutes around them. It can hold a flagged payment, question the customer, read the context and prepare the case. It must not conclude that fraud happened, release a hold on its own judgement or go beyond containment.
What does "real time" mean in fraud?
Fraud detection runs at different speeds. A card authorisation has to be approved or declined within a fraction of a second. That is the job of purpose-built scoring models, which compare each transaction with the customer's pattern and known fraud signals. An agent that reasons through a case is not built for that speed.
The next layer runs in seconds to minutes. A payment has been flagged and held, and someone needs to find out whether it is genuine. That is where an agent adds the most. It can contact the customer at once, ask the right questions, read the transaction history and the payee details, and assemble what a specialist needs.
Speed matters most where payments settle almost instantly. Once money has moved through a real-time system, recovery is hard. For the settlement side, see what real-time gross settlement is.
What may the agent do with the customer's money and data?
| Task | The agent | Where it stops |
|---|---|---|
| Responding to a model alert | Holds the payment or freezes the card under approved rules | Holds beyond the policy period |
| Talking to the customer | Asks scripted questions about the payment | Deciding the answers justify release for higher-risk cases |
| Reading context | Reviews recent transactions, payees and device changes | Data unrelated to the case |
| Spotting scam patterns | Matches what the customer says against known scam types | Concluding the customer is a victim or a suspect |
| Investigating alerts | Prepares a case summary with evidence | The outcome |
| Escalating | Passes the case to a specialist with a recommended priority | Nothing |
| Releasing a hold | Prepares the release | A person decides |
The investigation step is covered in how to review card fraud alerts with an AI agent. The logic of the whole table is set out in in fraud, an agent should contain, not conclude.
Where does the agent add something a model cannot?
Scoring models see transactions. They cannot see the conversation behind them. Many scams involve the customer making the payment willingly, under pressure from a scammer. The payment can look normal: the customer's own device, a plausible amount, a payee they added themselves.
An agent can ask. "Has someone asked you to move money urgently?" "Did someone tell you not to tell the bank the real reason for this payment?" "How did you first hear about this investment?" The answers often reveal a scam that the transaction data alone could not. That is detection in real time in the sense that matters to the customer: before the money leaves.
A worked example
Take a bank whose scoring model flags a 4,500 transfer from a customer's savings to a new payee, made from the customer's usual phone, at an ordinary time of day. The model score is moderate.
The agent holds the payment and asks the customer the scripted questions in the app. The customer says the payment is for a deposit on a rental flat they found online. They have not seen the flat; the landlord is overseas and asked for the deposit before a viewing.
The agent recognises a rental scam pattern. It keeps the hold, explains plainly why the request matches a known scam, and suggests the customer does not pay until they have viewed the property. It passes the case to a fraud specialist with the conversation and the payee details. The specialist calls the customer, who cancels the payment.
What controls make it safe?
- Holds under approved rules. The agent acts on model alerts and bank rules, not its own suspicion.
- Scripted questions, tested. The questions are written by the fraud team and tested on past cases before go-live.
- Release held for a person. Releasing a hold is outside the agent's bounds.
- Testing without exposing real data. Fraud scenarios can be tested with realistic synthetic cases where real data cannot be used; see how to certify an AI agent without real data.
- A record per case. The alert, the hold, the conversation and the outcome are filed together.
The wider limits are set out in AI agents in banking: what they may and may not do.
Where Pryme fits
In Pryme, placing a hold and releasing one are separate capabilities. The agent can be certified to hold and question at depth 3; release is held for a named person. The Pryme Context Protocol marks what the customer says as declared and what a third party claims as external, and an external claim cannot authorise a payment.
Questions people ask
Can the agent replace the bank's fraud scoring models?
No. Scoring models do a different job, at a speed an agent cannot match. The agent works alongside them, handling what happens after a flag.
What if a scammer is coaching the customer during the conversation?
That is common. Scripted questions are designed to surface it, and the agent should hand over to a person when answers seem rehearsed or contradictory.
Can the agent detect fraud by bank staff?
It can flag unusual patterns in internal activity for a person to review. Investigations into staff are sensitive and should be led by the bank's investigations team.