Synthetic Data Testing: What The FCA's Quantum Experimentation Group Shows About Trying AI Safely

Synthetic data testing means trying a new tool on artificially generated data that looks like the real thing but contains no real clients. The FCA used that approach in an announcement on 5 November 2026, and it is a useful model for any firm testing AI in operations.

What did the FCA announce on 5 November?

The FCA said it will bring together financial services firms, quantum technology providers and national research laboratories to explore how quantum computing could help tackle financial crime. The Quantum Experimentation Group will initially explore how quantum machine learning could help detect money laundering and fraud, and will launch in the first quarter of 2027.

Participants will use an FCA synthetic dataset, which the FCA describes as artificially generated data, to help overcome a key barrier to testing: access to suitable data. Jessica Rusu, the FCA's chief data, information and intelligence officer, said: “We want to help firms explore where quantum could make a meaningful difference, and early experimentation is essential to understanding its potential in financial services.”

Financial services firms, quantum technology providers and national research laboratories can register their interest in joining the group by 31 December 2026.

Why does synthetic data testing matter for AI in operations?

• **No client data at risk.** Early trials cannot leak what they do not hold.
• **Known answers.** Synthetic cases can be built with the right answer known, so you can measure what a tool catches and misses.
• **Faster start.** Data-sharing approvals are often the slowest part of a trial.
• **Rare cases on demand.** For financial crime work, synthetic data can include unusual patterns, such as chains of transfers, that appear only a handful of times in real records.
• **Repeatable.** The same cases can be rerun after every model or prompt change, so you can see if anything got worse.

Synthetic data has limits. It only contains the patterns someone thought to put in, so a tool that does well on it still needs a careful run on real records, with a person checking every result. Our article on AI model testing covers how to test before switching versions.

How should firms test AI agents?

• **Step one, synthetic or sample data.** Prove the workflow end to end with nothing real in it.
• **Step two, a parallel run.** The agent prepares the work on real records alongside your team, and a person compares results.
• **Step three, live with approval.** The agent prepares, a person approves every decision, and every step is logged.

Illustrative staged testing for an AI agent.

Stage	Data used	What a person does
Synthetic or sample run	No real client data	Checks results against the known answers
Parallel run	Real records, no actions taken	Compares the agent's work with the team's
Live with approval	Real records	Approves every decision before it takes effect

The demo in our Desk Twin uses illustrative records for the same reason. Our client onboarding files workflow follows the three steps above, and decisions on taking on a client stay entirely with your people.

Where should a firm start?

Pick one workflow and build twenty sample cases with known answers, including a few awkward ones, before any real data is involved. Our AI Developers London team can build those cases with you, and a 14-day Proof Run can take it from there to your own records. This article summarises an FCA announcement and is not advice.