The Two Clocks of Quantum Fraud Detection

By Luca Collacciani, Chief Growth Officer at Cleafy

Quantum computing has arrived in banking in two forms. The first breaks your cryptography: it comes with a regulatory deadline, a named owner, a steering committee and a migration plan that has been tested for years. The second might reduce fraud losses. It comes with a slide. Anyone who has spent part of a career selling the first of those knows the second is the more interesting problem.

Almost every large bank with a quantum programme lists fraud detection as a use case. Very few have anything close to in-production deployment. HSBC has run a dedicated quantum team for more than three years and has been piloting quantum-inspired algorithms with a focus on fraud, on the reasoning that even a couple of percentage points of accuracy is worth a great deal at that volume. Its own quantum lead has drawn the distinction plainly: defending cryptography is a must-do activity, while using quantum computing is a sensible-to-do one. Intesa Sanpaolo is exploring quantum machine learning for fraud with IBM. The pattern repeats across the sector; fraud sits on every quantum roadmap and in nobody's authorisation path.

The usual explanation is that the hardware is not ready; the better explanation is that fraud detection was never short of computing power.

A payment decision has a budget measured in tens of milliseconds, inside a system that must stay available continuously. Current quantum machines are cloud-queued, noisy and batch-oriented, and anyone offering a faster scoring engine is answering a question the industry did not ask.

What fraud detection is genuinely short of is examples. The attacks that matter most are the newest ones, and the newest ones arrive with almost no labelled history. Victims report days later. Some, particularly in impersonation scams, are too embarrassed to report at all. Generative AI has made this considerably worse, because a scam script can now be varied endlessly at close to zero cost, so the population of attack patterns refreshes faster than the population of labels describing it.

That is where the early quantum results actually sit. A trial published in late 2025 encoded more than five hundred features onto 128 qubits to identify subtle financial anomalies. The researchers explicitly declined to claim quantum advantage, describing the result instead as the clearest empirical signal so far that quantum-enabled machine learning could become useful. Learning from a few examples in very high dimensions, and optimisation over combinatorial structures such as mule account networks, are the two shapes where quantum methods currently look plausible. Neither of them is a transaction score.

So the honest way to plan the next decade is around two clocks.

The millisecond clock runs at the point of payment and stays classical. Behavioural signals, session context, device intelligence, models retrained continuously because the attack population changes weekly. Nothing about quantum computing helps here, and it is unlikely within the working lifetime of anyone reading this.

The structural clock runs upstream and much slower. It answers a different question: not whether this payment is wrong, but what the network around it looks like. Which accounts belong to the same laundering structure. Which recipients are newly created nodes in an existing ring. Which mule chains reassemble after being dismantled. These are combinatorial graph problems, the kind classical methods currently approximate rather than solve, and they are where quantum optimisation has a genuine claim.

This produces an unglamorous conclusion for anyone writing a quantum strategy this year. The binding constraint is not access to hardware, which three cloud providers will rent you this afternoon. It is the data asset. Optimisation over a mule network is worthless if you cannot see the network, and no single institution can see it, because mule chains are deliberately built across banks and across borders. That visibility requires cross-institution intelligence sharing and a legal basis for it, and in Europe both take years to negotiate rather than quarters.

So the limiting factor on quantum-assisted fraud detection in 2035 will not be qubit counts. It will be an intelligence-sharing agreement that somebody either signed or failed to sign in 2027. That is a decision available to every institution in this issue's readership right now, and it does not require a single physicist.

Meanwhile, the asymmetry compounds. Attacker capability at the human layer improves with every model release, on a cycle measured in months. Defender capability from quantum improves on a hardware roadmap measured in a decade.

The cryptographic half of the quantum question came with a date, so it will get done. The fraud half has no date, which means it only gets done by institutions that choose to set one. The years between those two curves are not a gap in capability. They are the business plan of every fraud network operating in Europe today.