Regulation

Anthropic is now inside the FCA's AI sandbox, not just a vendor selling into it, as applications to test AI in UK financial services jump 51%

The FCA's second Supercharged Sandbox cohort runs 21 firms including Scottish Widows and TrueLayer through live AI testing from 13 July to 31 December 2026, with Anthropic providing Claude access rather than deploying its own product, after applications rose 51% year on year to 199 submissions.

· Source: FinTech Futures


The detail worth sitting with here is who is in the room. Anthropic is not one of the 21 firms testing an AI use case, it is the model provider supporting the firms that are, offering Claude Code and Claude Cowork so participants can build and ship faster during the sandbox window. That is a different kind of regulatory engagement to the usual pattern of a fintech pitching a finished product to a supervisor. The FCA has effectively pulled a foundation model lab into its own testing environment, which says something about where the regulator thinks the real technical risk now sits: not only in how firms apply AI, but in the underlying models they are applying.

The use cases in this cohort read like a checklist of where AI is quietly doing real work in regulated finance already: agent-led payments, fraud and economic crime detection, compliance automation, AI governance frameworks, and tools for vulnerable consumers. None of that is share trading directly, but the governance patterns being tested here, agentic systems making or assisting decisions inside a regulated environment, are exactly the patterns that matter once AI agents start touching execution, portfolio construction or client-facing investment tools. A sandbox result on agent-led payments today is a preview of the supervisory questions AI-driven trading tools will face tomorrow.

The 51% jump in applications, to 199 total, is the more mundane but more telling number. That is not a handful of curious firms, it is a market signal that UK financial services see enough commercial upside in AI deployment to want a supervised route to test it rather than build in the dark and hope. Six months of live testing against a real regulator, ending 31 December 2026, is a meaningfully longer runway than most sandbox cohorts get, suggesting the FCA expects these use cases to need real time to surface genuine failure modes rather than obvious ones.

What the sandbox does not resolve is the harder question sitting behind it: once a model provider like Anthropic is embedded in how regulated firms build AI tools, does supervisory scrutiny of the tool implicitly become scrutiny of the model underneath it, and is the FCA equipped to draw that line consistently across 21 very different use cases at once. Six months from now, the interesting outcome will not be whether the sandbox produced compliant products, but whether it produced a clearer answer to where model-level responsibility ends and deployer-level responsibility begins.


Read the original: FinTech Futures - Anthropic is now inside the FCA's AI sandbox, not just a vendor selling into it, as applications to test AI in UK financial services jump 51%. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.