FCA’s AI Sandbox cohort two admits 21 organisations, including Scottish Widows, Money Advice Trust and TrueLayer, chosen from 199 applications — a 51% rise on the 132 applications the regulator fielded for cohort one. The programme now runs with Anthropic providing Claude, including Claude Code and Claude Cowork, alongside existing partners NayaOne and NVIDIA. But the categories the Financial Conduct Authority set out for this round are just as notable for what they leave out: neither autonomous underwriting nor automated claims decisioning appears among them.
The five use cases the FCA agreed to watch
The cohort’s mandate covers five specific problem areas: safer agent-led payments and commerce, sharper fraud and economic-crime detection, stronger AI governance and accountability, wider financial access for underserved consumers, and streamlined compliance automation. None of the five is an open licence to build a new product category from scratch — each maps to a supervisory concern the FCA has flagged before, from money-laundering exposure to advice-gap consumers falling outside mainstream credit and insurance.
FCA chief data, intelligence and information officer Jessica Rusu said the level of applicants “demonstrates the demand for trusted environments where firms can experiment safely and responsibly.” The sandbox itself is not new infrastructure built for this round: it builds on existing support from NayaOne and NVIDIA, the same technology partners that backed the original scheme, with Anthropic’s tools layered on top to speed up participants’ build cycles rather than replace the underlying testing environment.
The use cases missing from the cohort list
Notably absent from the five approved categories is anything resembling autonomous underwriting or automated claims decisioning — the two AI applications insurers most want regulatory clarity on. The FCA has not published a rationale for their absence, nor described the five categories as an exhaustive list of what it will consider. Fraud detection and compliance automation get a sanctioned sandbox; a model that prices a policy or denies a claim without human sign-off is not among the listed categories. What the published materials do state plainly is the limit of what participation confers — the FCA’s own disclaimer that taking part in the Supercharged Sandbox does not indicate its approval, endorsement, or authorisation of a product or service. Participation buys firms a testing ground, not a regulatory stamp — a distinction insurers pitching AI-driven pricing or claims tools to their boards should not blur.
The gap lines up with warnings raised elsewhere in UK financial regulation. The Mills Review has already flagged opaque AI-driven pricing models as a risk to fair value assessments in insurance — the same category of autonomous decisioning that does not appear in the cohort’s published use cases. On InsuraBeat’s reading, the two documents together describe a regulator whose published experiments so far involve AI that assists a human decision rather than AI that makes it; the FCA has not framed the distinction in those terms itself.
Nearly double the applicants, roughly the same number of seats
Cohort one drew 132 applications and the FCA selected 22 firms; it ran from September 2025 to January 2026. Cohort two launched on 13 July 2026 and runs until 31 December 2026, with applicant numbers up 51% but the accepted cohort size holding roughly flat at 21 organisations. That widening gap between demand and capacity is itself a signal: the sandbox gives participants access to high-performance, GPU-enabled infrastructure to develop and test models, a resource-intensive commitment that caps how far the FCA can scale acceptance even as interest grows. Firms also do not need to be FCA-regulated to take part, which widens the applicant pool to insurtechs and vendors well beyond the authorised-firm population the regulator directly supervises.
Two testing tracks, and a data gap that explains the caution
The Supercharged Sandbox is not the FCA’s only AI proving ground. Its separate AI Live Testing service picked eight new firms for its second cohort, including Barclays, Experian, Lloyds Banking Group’s Scottish Widows arm, and UBS, in a track aimed at regulated firms closer to live deployment rather than early-stage experimentation. Live Testing sits downstream of the sandbox model: firms there are validating AI systems nearer to production, which makes the absence of autonomous underwriting or claims automation from both published cohort descriptions all the more noticeable.
A joint Bank of England and FCA survey sets out the adoption picture behind the cohort. It found 75% of firms are already using AI, with a further 10% planning to within three years, yet foundation models, including large language models, account for only 17% of AI use cases industry-wide — most deployments remain narrower, rules-based systems rather than generative models. More tellingly, only 34% of firms report a complete understanding of the AI they use, against 46% with only partial understanding. That gap between adoption and comprehension is the backdrop against which any widening of the sandbox’s scope would be judged, and it echoes the same opacity concerns that sit behind rules such as the FCA’s broker professional indemnity consultation. It also tracks with debates playing out beyond the UK, where state insurance regulators are drafting their own AI governance bulletins rather than waiting for a single federal standard.
None of this is retrofitted caution. The Supercharged Sandbox’s original June 2025 launch gave firms access to NVIDIA accelerated computing and NVIDIA AI Enterprise Software, and the FCA has kept the same infrastructure partners in place for cohort two. The five approved use cases for this cohort look, in that light, like a continuation of a perimeter the regulator has been drawing for over a year. Whether that perimeter widens toward underwriting and claims automation is something the FCA has not yet signalled either way.
Mini-FAQ
Does joining the FCA’s Supercharged Sandbox mean a firm’s AI product is approved?
How many firms applied to join the second Supercharged Sandbox cohort?
Do firms need to be FCA-regulated to join the sandbox?
Sources used:
