The FCA's Mills Review Puts Public AI Into the Financial-Advice Conversation

Short answer: On 6 July 2026 the FCA published the Mills Review, "AI and the future of retail financial services", an FCA Board-commissioned review led by FCA executive director Sheldon Mills. It is not FCA policy and it is not an independent or external report: it is a review the Board asked for, and the FCA's Board and Executive will now consider its recommendations. Among its seven recommendations, one reaches a surface that most AI-governance conversations have skated past: the general-purpose AI tools people already use to ask about money. The Review recommends the FCA consider how AI models sitting outside the regulatory perimeter influence retail financial decisions. Elsewhere it names the general-purpose tools "ChatGPT, Claude or Gemini" as examples of what consumers are turning to. Nothing has been decided; the recommendation is that the FCA look at the question. But the question itself is the shift worth reading precisely, because it puts public AI answers into the financial-advice conversation at a prominent regulatory level.

So a review commissioned by a regulator's board has, in effect, named the surface: the AI a customer consults about you that you neither operate nor can inspect internally. This piece sets out what the Review actually recommends, why public AI now matters in a financial decision journey, and, carefully, what remains unresolved.

What the FCA Mills Review actually recommends

Start with what the document is, because it deserves an accurate reading before anyone builds a conclusion on top of it.

The Mills Review was commissioned by the FCA Board and led by FCA executive director Sheldon Mills, and published on 6 July 2026 under the title "AI and the future of retail financial services." That provenance matters. It is not an independent think-tank paper and it is not, on publication, FCA policy. It is a Board-commissioned review, and the next step the FCA has set out is that its Board and Executive will consider its recommendations. Everything that follows should be read through that frame: these are recommendations to the regulator, not decisions by it.

The Review makes seven recommendations. The one that bears on the surface we work on is "Secure and adapt the regulatory perimeter." Under it, the Review recommends that the FCA consider how AI models sitting outside the regulatory perimeter influence retail financial decisions. That is the whole of the ask at this stage: consider the question. The Review does not decide the answer, and it is careful not to pre-commit one. Any outcome is open, from guidance, to recommendations to government, to no change at all. What the recommendation establishes is not a rule but a question the regulator has been asked to look at: what happens when the tools shaping a customer's financial decision are ones whose influence may sit outside the FCA's activity-based perimeter.

A question raised, not a power taken. That framing is the defining property of the recommendation, and the one most commentary will overstate.

Why public AI now sits in the financial decision journey

Here is why that question is not academic, and why it landed in a Board-commissioned review rather than a footnote.

The Review names the general-purpose tools directly, "ChatGPT, Claude or Gemini", as examples of what people increasingly reach for, including for questions about money. That is a notable thing for a regulator's review to record: the assistant a customer consults before opening a product, disputing a charge, or choosing a policy is, more and more, a general-purpose tool. The Review identifies these general-purpose tools, operating outside or potentially outside the financial-services perimeter, as a question requiring examination: the perimeter is activity-based, and where its boundaries fall when such a tool shapes a financial decision is part of what the Review asks the FCA to look at.

The Review pairs that observation with evidence about what consumers understand while they do it. Drawing on a nationally representative survey of 5,026 UK adults holding a day-to-day bank account, commissioned for the Review, it records that when general-purpose tools such as ChatGPT, Claude or Gemini are used for financial advice, only around two in five people correctly identify the level of protection available. Read that carefully: it is a statement about awareness of protection, not about accuracy. But it points at the same seam. A growing share of consumers turn to these tools for financial advice, and many may not realise that the formal recourse routes they would assume, the ones that attach to a regulated firm's advice, will not apply in the same way to an answer from a general-purpose assistant. The tool is fluent, specific, and available at the moment of the decision. The protections a customer imagines standing behind it may simply not be there.

That is the journey the Review has put on the table: a customer, a high-intent question about money, and a confident answer from a general-purpose tool that may sit outside the perimeter, with the customer's understanding of their own protection running behind the technology they're using. It's the reason AI answer accuracy is becoming a governance issue reaches beyond any single firm's internal controls: the answer forms where those controls don't extend.

What the Review raises, and what it leaves unresolved

The reflex, on reading a recommendation like this, is to jump to the conclusion: the regulator is coming for the AI companies. It is worth being precise, because that reading is not what the document says.

The Review recommends the FCA consider how out-of-perimeter models influence retail financial decisions. It does not say the FCA will regulate model providers, it does not say the FCA is taking powers over any general-purpose tool, and it does not decide anything. The Board and Executive have yet to consider it. The outcome is genuinely open. So the honest reading is that the Review has raised a question and framed a surface, not closed either. What it has usefully done is make the external-AI influence and the consumer-protection and recourse question legible at a regulatory level: consumers are acting on answers from general-purpose tools that may sit outside the perimeter, and their grasp of the protection those answers carry lags behind their use of them.

That is a question about perimeter, protection, and recourse. It is the FCA's to work through, not ours, and not one this piece argues any particular way. We won't pretend to resolve it, because the Review didn't, and because it isn't the dimension we work on. What we can speak to is a narrower, adjacent, and measurable slice of the same surface: not whether a customer is protected, and not whether an answer counts as advice, but simply whether what a public AI tool says about a specific institution's facts is true. Is the fee it quotes the current one? Is the policy it describes the live one? Is the process the one that still applies? That is contextual accuracy: accuracy about your specific case, not the general category. It is one concrete, checkable thread inside the much larger question the Review opened.

We're deliberately modest about where that thread runs. Measuring whether a public answer about you is accurate does not solve the perimeter question, does not settle whether something is regulated advice, does not restore a recourse route, and is not a compliance or legal-duty guarantee. Those are the Review's questions and the FCA's to consider. Accuracy is simply the part a firm can see, check against the truth it controls, and keep a dated record of, while the larger questions are worked out by the people whose job that is. If the method is what you're after, how Lawnise verifies AI answers against official sources sets out exactly how that comparison is drawn.

A UK review with an audience beyond the UK

One last thing worth holding, lightly, about scope.

The Mills Review is a UK document, commissioned by the FCA Board, and its recommendations are for the FCA to consider in a UK context. It asserts nothing about any other market, and neither do we. The underlying question, though, is relevant wherever consumers use general-purpose AI in financial decisions. Consumers reach for the same general-purpose tools everywhere, and the gap between a fluent answer and an accurate one about a specific institution does not stop at a border. No rule follows from that observation, and we're not implying one. It simply means the surface the Review has named is one that firms in the markets we work in have reason to understand early, whatever any regulator eventually decides to do about it.

We measure that surface in the open. Our public barometers take genuine customer-style questions about defined scopes, put them to public AI assistants, and compare the answers against the live published truth. The point to carry away isn't a number. It's the pattern the barometers make visible: a general-purpose tool can answer a customer's money question fluently and specifically, and still be wrong about the institution's actual facts, on a surface the institution does not own or directly control. The Mills Review has now asked the regulator to consider what that means for perimeter and protection. That's the FCA's question. The narrower one, whether the answer about you is even accurate, is one a firm can start looking at today. If you'd like to see where the accuracy line falls for your own scope, we're glad to talk.