How to Run a Quantum Pilot for Fraud Detection in Financial Services

Knowing how to run a quantum pilot used to be a question for research labs. After this week, it is a question for compliance and data teams at banks, payment firms and fintechs.

On November 18, 2026, the UK’s Financial Conduct Authority announced a Quantum Experimentation Group launching in early 2027. According to the FCA release via EuropaWire, the regulator will set the problem statements, supply a synthetic dataset and work alongside firms testing whether quantum machine learning can improve fraud and anti-money laundering detection. Registration closes on 31 December 2026.

That is six weeks. Firms that show up with a clear use case, a baseline model and a small internal team will learn far more than firms that show up curious. This guide walks you through a practical pilot you can start now, whether or not you join the FCA group.

What You Need Before You Start

  • A sponsor in financial crime or compliance: someone who owns transaction monitoring and can say what a real improvement would look like.
  • A small cross-functional team: one data scientist with classical ML experience, one engineer comfortable with Python and cloud APIs, and one analyst from the fraud or AML team.
  • Cloud quantum access: an account on IBM Quantum, Amazon Braket or Azure Quantum. Free and pay-per-use tiers are enough for a pilot.
  • Quantum software libraries: Qiskit or PennyLane, both of which run on classical simulators before you touch real hardware.
  • A modest budget: for most mid-sized firms, a three-month pilot costs less than one extra compliance analyst.

Step 1: Pick One Narrow, Measurable Problem

Do not try to “apply quantum to AML”. Pick something you can measure. Good candidates include reducing false positives on a single transaction monitoring rule, classifying alerts as likely genuine or likely noise, or detecting a specific typology like mule account networks.

Write the problem as one sentence with a number attached. For example: “Reduce false positives on our cross-border payments rule by 15% without lowering true positive detection.” That sentence will guide every decision after it, and it is exactly the kind of statement the FCA says it will define for participants.

Step 2: Build a Strong Classical Baseline First

This is the step most teams skip, and it is the most important. Before any quantum code runs, train the best classical model you reasonably can on the same problem: gradient boosted trees with XGBoost or LightGBM, a graph neural network for network typologies, or simply your current rules engine.

Record precision, recall, false positive rate and alert volume. Without this baseline, any quantum result is meaningless. To be fair, many quantum ML experiments in finance end with the classical model winning. That is still a useful result, because now you know where the boundary is.

Step 3: Prepare Synthetic or Privacy-Safe Data

Real transaction data is sensitive under GDPR and UK GDPR, and sending it to an external quantum cloud raises hard questions. The FCA itself named data access as a major blocker, which is why it is supplying a synthetic dataset.

For your internal pilot, generate synthetic transactions with tools like SDV (Synthetic Data Vault) or Gretel, or use public research datasets such as PaySim. Keep the same statistical shape as your real data: fraud rate, transaction sizes, time patterns. Run your classical baseline on the synthetic data too, so comparisons stay fair. Then involve your data protection officer early. A one-page data flow diagram saves weeks of back-and-forth later.

Step 4: Design a Small Hybrid Quantum Experiment

Current quantum hardware is noisy and small, so practical pilots are hybrid. A classical system does feature engineering and reduces the data to a handful of features, then a quantum model handles the classification step.

Two approaches are common in early finance work. Quantum kernel methods (like a quantum support vector classifier) map features into a quantum state space and compare them. Variational quantum classifiers train a parameterised circuit, similar to a tiny neural network. Start with 4 to 8 features and a few thousand samples. Run everything on a simulator first, then move the best configuration to real hardware on IBM Quantum or Amazon Braket and compare.

Step 5: Compare Results Honestly and Document Everything

Put the quantum results next to your classical baseline using the same metrics and the same test data. Also record what regulators and auditors will care about: runtime, cost per run, reproducibility (quantum results can vary between runs), and explainability. A model that is 2% better but impossible to explain to a supervisor is not production-ready.

Write a short internal report: problem, data, baseline, quantum approach, results, limitations, next steps. That report is your entry ticket to the FCA group and to any future regulatory conversation.

Step 6: Register for the FCA Group and Plan Your Security Track

If you operate in the UK or serve UK clients, register your interest with the FCA before 31 December 2026. The group is open to firms at every stage, from early exploration to advanced experimentation. Firms in the EU should watch closely too, since UK regulatory thinking often shapes conversations in Brussels and Frankfurt.

While you are at it, start a separate post-quantum cryptography inventory. The same machines that may one day help detect fraud could eventually break today’s encryption. The UK is also putting money into quantum commercialisation, as The Quantum Insider reported, so expect both opportunity and pressure to grow.

A Realistic 12-Week Timeline

Weeks 1 and 2: agree the problem statement, success metric and data approach with your sponsor and DPO. Weeks 3 to 5: build the synthetic dataset and the classical baseline. Weeks 6 to 9: run quantum experiments on a simulator, then on real hardware for the best two or three configurations. Weeks 10 to 12: compare, document and present to leadership. Keep a weekly 30-minute check-in with the fraud team so the work stays tied to real alert patterns.

Common Mistakes to Avoid

Chasing “quantum advantage” headlines. The goal of a 2026 or 2027 pilot is learning, not beating every classical model. Set expectations with leadership upfront so a neutral result is not treated as a failure.

Using real customer data too early. Sending raw transactions to an external quantum cloud without DPO sign-off creates a compliance problem bigger than the one you are trying to solve.

Running the pilot inside the innovation team only. If the fraud and AML analysts are not involved, the use case will drift toward what is technically interesting rather than what actually reduces alert backlogs.

Key Takeaways

  • Start narrow: one rule, one typology, one measurable target.
  • Baseline first: a strong classical model is what makes quantum results meaningful.
  • Protect the data: synthetic datasets remove the biggest privacy and access blocker.
  • Go hybrid: classical feature engineering plus a small quantum classifier is the realistic setup today.
  • Act on the deadline: FCA registration closes on 31 December 2026.

Need Expert Help?

If this feels like a lot to manage alone, TecniForge can handle the heavy lifting. Our team specializes in custom software development and AI integration. Get in touch with our experts.

Also read: FCA Quantum Computing Group: 6 Things Financial Firms Should Know Before 2027, our earlier coverage on why this matters today.

Your move: book a 30-minute session this week with your head of financial crime and pick the one alert type that wastes the most analyst hours. That is your pilot. Everything else follows from it.


Discover more from TecniForge

Subscribe to get the latest posts sent to your email.