The FCA has brought Anthropic into its Supercharged Sandbox and told the Prime Minister it will not add AI-specific regulation. For claims and legal operations, the signal matters more than the headline.
On 21 July 2026, the FCA confirmed that Anthropic will support the second cohort of its Supercharged Sandbox. The sandbox is a controlled environment where firms can experiment with advanced AI safely, built on the Digital Sandbox infrastructure from NayaOne and backed by NVIDIA compute. Anthropic is providing cohort participants with access to Claude, including Claude Code and Claude Cowork, to speed up development.
It is easy to read this as a vendor story. We think that misses the point. The more useful read is what the regulator is signalling about how it wants AI adopted in financial services, and that signal applies well beyond the 21 organisations in the room.
What the FCA actually announced
The second cohort has 21 organisations, including Scottish Widows, the Money Advice Trust and TrueLayer. Demand was strong. The FCA received 199 applications, up 51 percent on the 132 it received for the first cohort.
The cohort will work across five test themes: safer agent-led payments and commerce, detecting fraud and economic crime, strengthening AI governance and accountability, widening access to financial services for vulnerable and underserved consumers, and streamlining compliance and business automation.
Alongside the sandbox, the FCA launched an Agentic Academy, a 10-week specialist AI programme delivered with the Centre for Finance, Technology and Entrepreneurship. Jessica Rusu, the FCA's chief data, intelligence and information officer, framed the initiative as helping firms move from experimentation to safe, responsible adoption at pace.
The signal beneath the announcement
Here is the part worth drawing out. In its response to the Prime Minister's letter on growth, the FCA said it would avoid additional AI-specific regulation and instead rely on the frameworks that already exist.
That is a deliberate position, and it tells you how the regulator expects AI to be adopted. Not through a new rulebook written specifically for AI, and not through a pause while everyone waits for one. Through safe, controlled experimentation, under the rules that already apply, with governance and accountability treated as first-order questions rather than an afterthought.
Notice that governance and accountability is not a footnote here. It is one of the five explicit test themes the cohort will work on. The regulator is not treating AI oversight as a compliance tax to be minimised. It is treating it as a core capability to be built.
What this means for claims and legal operations
If you run a claims desk, a credit hire recovery function, or a claimant law firm, none of this requires you to be in the sandbox. The practical read is the opposite. The discipline the FCA is modelling is available to any operator willing to adopt it.
It comes down to a few plain principles.
Run controlled pilots, not open deployments. Pick a defined task, set the boundaries, measure the output, and expand only when the results hold up. This is what the sandbox is for at national scale. You can run the same pattern on a single workflow.
Keep a named human accountable. The FCA is not asking firms to prove the technology is infallible. It is asking who is responsible when it is used. Every AI-assisted output in a claims or legal file should have an identifiable person who reviews and signs it off. Accountability is not a setting you switch on later. It is designed in from the first pilot. The Pinsent Masons AI ruling shows what happens when that accountability is assumed rather than designed in.
Ground the work in your own knowledge and keep an audit trail. Existing frameworks already expect you to show your reasoning, cite your sources, and reconstruct how a decision was reached. AI that answers from a defined knowledge base, with the source visible and the trail recorded, meets that expectation. AI that answers from a general model, with no trace of where the answer came from, does not.
Do not wait for AI-specific regulation. The FCA has said plainly that it is not coming. Firms holding off on adoption until a dedicated AI rulebook arrives are waiting for something the regulator has told them not to expect. The rules you need are the rules you already work under.
A measured note, because the framing matters
Two points of balance are worth keeping in view, so this reads as adoption with discipline rather than a green light.
Sandbox participation is controlled experimentation. It is not a signal that unsupervised, live deployment is now acceptable. The whole design of the sandbox is that firms test in a bounded environment before anything touches real customers. The lesson to carry across is the bounded testing, not a shortcut past it.
And governance and accountability being an explicit test theme cuts both ways. It confirms the regulator sees AI oversight as unfinished work, even for sophisticated firms with resources to spend on it. If the organisations in the cohort are still building this muscle, no operator should assume they already have it. That is a reason to be deliberate, not a reason to hesitate.
Where CaseFlow Automation fits
We build specialist, compliance-first AI workflows for claims and legal teams. The design principles the FCA is signalling are the ones our platform is built around. Work is grounded in a defined knowledge base rather than a general model, every output carries an audit trail, and a named human signs off before anything leaves the file.
That is not a reaction to this announcement. It is the posture we have taken from the start, because it is the posture the existing rules already imply. If your firm wants to adopt AI in claims or legal operations with controlled pilots, clear accountability, and an audit trail that stands up to scrutiny, that is precisely the conversation we are set up to have.
We are happy to talk it through, either as part of an existing CaseFlow engagement or as a standalone discussion.
Craig Budsworth is the legal strategy lead across CaseFlow Automation's products. He reviews the case law and regulatory developments that shape how AI is adopted in claims and legal operations, and works with claims and legal teams on the discipline behind compliance-first AI.
Frequently Asked Questions
- Do we need to be in the FCA sandbox to adopt AI safely?
- No. The sandbox is a controlled environment for a small cohort, but the discipline it models is available to any firm. Run controlled pilots, keep a named human accountable, ground the work in your own knowledge, and maintain an audit trail. None of that requires cohort membership.
- Is the FCA about to introduce AI-specific regulation we should wait for?
- The FCA has said it will avoid additional AI-specific regulation and rely on existing frameworks. Waiting for a dedicated AI rulebook means waiting for something the regulator has told firms not to expect. The rules that apply to AI-assisted work are the rules you already operate under.
- What does governance and accountability mean in practice for a claims or legal file?
- It means a named person is responsible for every AI-assisted output, reviews it before it is used, and signs it off. It also means you can show where an answer came from and how a decision was reached. Accountability is designed in, not added later.
- Does the sandbox mean unsupervised AI in live claims work is now acceptable?
- No. Sandbox participation is controlled experimentation in a bounded environment. It is not a signal that live, unsupervised deployment is approved. The transferable lesson is the controlled testing before anything reaches a real customer.
- How should a claimant law firm or credit hire operator start?
- Choose one repetitive, rules-based task. Set clear boundaries, ground the AI in your own knowledge base, keep the audit trail, and put a named reviewer on the output. Measure the result over a defined period, then expand only if it holds up.
Related articles
Instant Claims Has a Trust Problem. Where Automation Actually Belongs in Motor Claims
UK insurers want instant motor claims, but AI-built fraud is rising. Where straight-through processing works, where it breaks, and what to automate first.
Where AI Has Actually Moved UK Insurance Distribution in 2026 (and Where It Hasn't Yet)
A calm read on what genuinely shifted in the first half of 2026: the Aviva and Compare the Market ChatGPT launches, the 70 percent consumer expectation signal, and the parts of the business that have not moved despite the headlines.
What Credit Hire Teams Should Take From the Pinsent Masons AI Ruling
A senior judge has said a junior solicitor 'almost entirely outsourced the thinking process' to AI. Here is what the Pinsent Masons ruling means for credit hire operations and how to design against the three failure modes it exposes.