Agentic AI and fraud prevention: how banks can close the asymmetry with attackers

On 1 October 2026 in Milan, Cleafy hosted a panel on Proactive defence and autonomous investigations: the new frontier of anti-fraud in the age of agentic AI. Moderated by Cleafy CTO, Nicolò Pastore, it brought together three fraud and security leaders.

The conversation was honest and open. Attackers have adopted AI faster than banks, and the gap is widening. But the panel also mapped where defenders still hold the advantage, and where agentic AI fraud prevention can help them use it. Here is what we learned.

AI has improved fraud on two fronts: volume and quality. A scam SMS aimed at Italian customers once had to be written in good Italian, which limited who could run it. Today, generative AI produces fluent Italian voice and text, and volumes have risen sharply.

The operation behind the message has industrialised too. Panellists described "silent calls" in which no one speaks: automated systems check that a number is live before a callback, using customer lists bought on the dark web. Fraud now runs as a structured industry, and agentic AI can connect criminal groups that once worked separately.

Better language has not ended deliberate crudeness. Some criminals still send poorly written texts on purpose, to filter out sceptical recipients and reach only the most vulnerable. AI simply gives them the choice. One bank even runs a team that plays along with fraudsters to gather intelligence and waste their time.

New fraud types keep emerging. Tax-payment fraud has become a significant source of losses, something few would have predicted three years ago. Fraud teams are also tracking deepfake videos injected into digital onboarding.

Most scams start far from the bank. At one large bank, around 85% of reported scams originate on social networks or messaging platforms. A victim might see a deepfaked executive promoting a fake investment, then receive a call from a spoofed bank number.

Measured by time, the bank is involved in only about 2% of that journey: the moment of payment. It then has around 250 milliseconds to decide whether a payment is fraudulent. The first payment is often small, around €80, and looks genuine.

By the time the bank calls, the fraudster has built trust. One bank cut the average call needed to talk a customer out of a scam payment from 20 to 25 minutes down to 11. Those calls have since lengthened again to around 45 minutes. However hard the bank works, many customers still prefer to believe the fraudster.

A benchmark cited on the panel holds that an excellent fraud team analyses at least 20% of the alerts it generates. That assumes time to prioritise. At one large bank, around 55% of payments are now instant, so every case is urgent, and stolen funds can leave Europe within four minutes.

With headcount fixed and volumes rising, there are two options: generate fewer alerts, or analyse them differently. The panel was clear on agentic AI. An agent should not review a sample of cases; it should review all of them. It closes the simple cases and hands analysts a prepared file for the rest. The human share of the work becomes smaller but genuinely qualitative.

The same logic applies beyond alerts. On one team, only about 40% of time goes on fraud analysis itself. The rest covers complaint and pre-litigation reports, regression testing of new software releases, and documentation. That work needs fraud knowledge, but it does not all need to be done by people.

Panellists pointed to warnings from European supervisors about a growing asymmetry between attack and defence, and stressed that the gap is not only technological. Criminals have no regulation, no data protection officer and no procurement process. Banks do.

The effect is sharpest at mid-sized banks. Bringing an AI agent into a workflow means engaging the DPO and AI Act owners from the start, which can take two to six months per project. Yet smaller banks face the same AI-powered attacks as large ones. At one regional bank, the contact centre's share of fraud workload has grown from around 10% to as much as 40%.

There is an upside. Regulation pushes banks to address new technology earlier than other sectors, and to do it properly. The challenge is speed without shortcuts.

Make it routine. A new AI agent should go through the same process as any new vendor, software or product, with a single end-to-end assessment of data, purpose, model handling and record-keeping. Banks built that muscle with the EBA outsourcing guidelines and with DORA, and can apply it again. Shadow IT became shadow AI; the response is to intercept it early.

Break silos too. In one organisation, fraud, AML and credit sit in one department, and that integration is real firepower. Fraudsters run as one industry, and defenders benefit from working as one. Throughout, the panel stressed that people must stay at the centre, governing the machine and protecting customers.

Blocking alone no longer works. A home renovation and an investment scam can look identical in the data, so heavy blocking hurts genuine customers. Instead, one bank initiates an automated dialogue when a payment appears anomalous within the 250-millisecond window, prompting the customer to pause. The bank reports a 90% reduction in false positives.

The next step is a dynamic dialogue powered by generative AI. Questions would adapt to each customer's history and digital literacy, clearing genuine payments quickly and escalating to a human when needed.

This works because scams exploit cognitive shortcuts, especially urgency and authority. Generic training is hard to measure and fades under pressure. Guidance at the moment of greatest vulnerability works far better: stop, call someone you trust, then decide. Panellists argued for bringing behavioural scientists into fraud teams.

AI can work for customers too. In one recent case, a customer targeted by fake Police officers uploaded the forged document to an AI assistant, which flagged it as a scam. Banks have an opportunity to give customers trusted tools that do the same.

As one panellist put it, even if banks are not ahead of the game, they need to play a different one.

The panel described one pattern from three angles:
1. AI scaled fraud to industrial volume.
2. Fraud operations, still bound to human speed and sampling, did not keep pace.
3. Faster tools alone will not close that gap; the operating model has to change.

That is the idea behind Autonomous Fraud Operations. Specialised agents are designed to process every signal as it appears, reconstruct coordinated campaigns and continuously re-evaluate past activity, so coverage no longer depends on sampling. Analysts move up the stack to the decisions, campaigns and policies where human judgement matters most.

Autonomy also depends on the data underneath it. Agents bolted onto fragmented telemetry produce confident guesses rather than reliable outcomes. Cleafy's Digital Narrative correlates users, devices, sessions and transactions as they happen, giving agents the full context they need to reason. Each step is designed to be logged and replayable, supporting the governance the panel called for under DORA and the AI Act.

Discover how Nyx brings Autonomous Fraud Operations to your bank.