The FCA is handing its sandbox cohort Claude: what that signals for UK financial services

What changed

On 22 July 2026 the FCA announced that Anthropic is supporting the second cohort of its Supercharged Sandbox, providing participants with access to Claude, explicitly including Claude Code and Claude Cowork. The cohort is 21 organisations, selected from 199 applications, a 51 per cent increase on the 132 applications the first cohort drew; the FCA's named examples include Scottish Widows, Money Advice Trust and TrueLayer. Cohort 2 launched on 13 July 2026, runs on the NayaOne platform with NVIDIA infrastructure, and participants keep access until 31 December 2026, with a showcase demo day on 26 November 2026. The FCA's stated focus areas are safer agent-led payments and commerce, fraud and economic crime detection, AI governance and accountability, access for vulnerable and underserved consumers, and compliance and business automation. Jessica Rusu, the FCA's Chief Data, Intelligence and Information Officer, framed Anthropic's involvement as giving participants the technology they need to move faster.

Why it matters for UK business

The announcement is short. The signal inside it is not, and it comes in three layers.

The first is the tooling itself. The FCA did not provision a chat window. It provisioned Claude Code and Claude Cowork by name, which are agentic tools: software that takes actions, writes and runs code, and works across files and systems rather than answering questions. A regulator choosing to put agentic tooling in front of 21 firms, inside its own controlled environment, is a statement about where it believes the sector's real experimentation needs to happen. UK financial services firms have spent two years piloting chat assistants; the regulator has just skipped ahead to the harder question.

The second layer is the theme list, which is the closest thing the sector gets to a map of where supervisory comfort is forming. Safer agent-led payments sits first, which is striking given that agent-initiated payments are precisely the use case most compliance teams currently rule out. Fraud and economic crime detection, and access for vulnerable consumers, are the FCA's perennial priorities restated in AI form. And AI governance and accountability appears as a theme in its own right, which means the FCA expects the cohort to produce not just applications but evidence about how to govern them. When the showcase lands on 26 November, the governance outputs will be the ones worth reading most closely.

The third layer is the experiment's structure, because it is replicable and most firms have not noticed. The sandbox pattern is: a bounded environment, synthetic or governed data, defined themes, a fixed end date, and a demo day. None of that requires the FCA. A UK financial services firm evaluating Claude can run the same shape internally, and a firm that does so is not just borrowing good practice; it is aligning its evidence with the structure its regulator has just demonstrated it finds acceptable. That alignment has a quiet procurement value too. When a board or a risk committee asks whether experimenting with frontier models is defensible, the honest answer is now that the FCA is running the same experiment, on the same tooling, with 21 firms watching.

One number deserves its own sentence: 199 applications for 21 places. The demand side of UK financial services AI experimentation is oversubscribed nine to one at the regulator's door, which says the constraint in the sector is not appetite. It is safe structure.

What to do, and what not to do

Do:

• Map your own candidate use cases against the five cohort themes. Where a use case lands on the list, the sandbox gives it a regulator-visible reference point; where it does not, that is worth knowing before you commit budget.
• Copy the structure for internal pilots: bounded environment, synthetic or properly governed data, a named governance workstream, a fixed end date, and a demo. The structure is what makes results defensible, to a board now and potentially to a supervisor later.
• Diarise 26 November 2026. The showcase demo day is the dated event where the cohort's outputs, especially on governance and accountability, become visible.
• If agent-led payments or fraud detection are on your roadmap, treat the cohort themes as evidence that these are discussable with the regulator, and prepare the governance case accordingly.

Do not:

• Present sandbox participation, anyone's, as FCA approval of a vendor, a model, or a use case. It is a test environment, and the FCA's provision of Claude is tooling access, not endorsement.
• Plan around a third cohort. Applications for cohort 2 closed on 1 June 2026 and no further window had been published as at 11 August 2026.
• Read the theme list as permission to deploy. The distance between a sandbox experiment on synthetic data and a production system touching customer money is the entire discipline of financial services change control, and it has not shortened.

Where The AI Consultancy fits

Structuring a bounded Claude pilot with a governance workstream that would survive supervisory reading is what our Claude consulting engagements do, and for regulated firms the starting point is usually the assessment work in our AI readiness service. Our guide to Claude for UK financial services covers the sector's compliance posture in detail, and our briefing on the FCA's new Handbook API covers the other half of this month's FCA signal: the regulator making its own rulebook machine-readable.

Verified on 11 August 2026 against the FCA's press release of 22 July 2026 and its Supercharged Sandbox page. This briefing is general information, not regulatory advice; confirm your firm's position with your compliance function or a qualified adviser.