Is the FCA ignoring an obvious breach of the regulatory perimeter?

The Financial Conduct Authority has acknowledged that consumers are increasingly using general-purpose artificial intelligence systems to make investment and other financial decisions. It has also warned that these systems may produce misleading, inaccurate or outdated information. Nevertheless, the FCA’s published position is that users of general-purpose large language models are not receiving regulated financial advice. That conclusion deserves serious scrutiny.

We decided to test that proposition. We asked a selection of widely available AI models for discrete financial advice—not merely general information or financial education. The questions sought specific recommendations based upon stated circumstances. The exercise demonstrated how readily a conversation can move from generic guidance to an apparently tailored recommendation capable of influencing a consumer’s financial decision.

This is important because the regulatory boundary should ordinarily depend upon the substance of an activity, not the technology used to deliver it or the provider’s description of its service. Article 53 of the Financial Services and Markets Act 2000 (Regulated Activities) Order 2001 captures advice on the merits of buying, selling or dealing in particular investments. The FCA’s own guidance recognises that a personal recommendation may arise where a recommendation is presented as suitable for a person or based upon that person’s circumstances.

An AI system can plainly produce such an answer. A consumer can provide details of their age, income, savings, pension arrangements, objectives and attitude to risk. The system may then recommend a particular fund, share, pension transaction or investment allocation. Functionally, this may be difficult to distinguish from a recommendation produced by a human adviser or regulated automated-advice service.

Our testing brings that issue out of the realm of theory. The relevant question is not simply whether an AI model is marketed as a financial adviser. It is whether, in response to an individual consumer’s circumstances and request, it actually provides an opinion on the merits of a particular financial course of action. Where it does so, calling the output “information” does not necessarily alter its substance.

The FCA nevertheless draws a distinction between a general-purpose LLM and an LLM specifically deployed to provide financial advice. In its March 2026 Perimeter Report, it says that consumers using general-purpose systems such as ChatGPT or Claude are not receiving regulated advice. Conversely, it accepts that an LLM specifically deployed to provide financial advice would likely fall within the perimeter.

That distinction is open to challenge. Article 53 is concerned with the activity performed, rather than whether the person performing it undertakes other activities. A solicitor does not cease to give legal advice because the solicitor also undertakes administration. Why, then, should an AI provider escape scrutiny merely because its system can discuss holidays, recipes and history as well as investments?

The “general-purpose” label describes the product but does not necessarily answer the statutory question. If the system gives a specific recommendation on the merits of a particular investment, the fact that it can also answer unrelated questions should not automatically change the character of that recommendation.

There are genuine legal complications. A regulated activity must ordinarily be undertaken “by way of business”. Providers may argue that they supply access to a general information tool rather than operate an investment-advice business. They may also point to warnings that responses are informational and should not be treated as professional advice.

Neither point is necessarily conclusive. Major AI platforms are commercial services. Some charge subscriptions, retain customers and promote the usefulness of their systems across a wide range of professional and personal decisions. Moreover, a disclaimer cannot automatically change the substance of what occurs. An unauthorised human adviser could not necessarily avoid regulation by placing “not financial advice” beneath an otherwise clear and personalised recommendation.

The FCA has demonstrated that it understands the risks. Its earlier work on automated investment services identified weaknesses involving suitability, risk profiling, disclosure and governance. Its 2026 Mills Review recognises that AI systems will increasingly recommend actions, initiate transactions and execute financial decisions. Yet there appears to have been no publicly reported FCA prosecution, injunction or perimeter case addressing financial advice generated by a general-purpose AI system.

This contrasts with the FCA’s approach to human and online publishers. The regulator has prosecuted individuals for unauthorised advice and pursued subscription services providing investment recommendations and trading alerts. It has also taken increasingly visible action against unlawful online financial promotions. The apparent message is troubling: a human charging for a trading recommendation may face investigation, while a global technology company producing a comparable recommendation through an algorithm may be characterised as outside the perimeter.

The FCA might respond that enforcement requires evidence, jurisdiction, an identifiable UK activity and a realistic prospect of establishing every element of an offence. Those are legitimate considerations. It is also possible that relevant investigative work is taking place confidentially. However, those considerations do not justify assuming that general-purpose AI responses are unregulated regardless of their content.

Nor is consumer education an adequate substitute for testing the perimeter. The FCA says its present focus is on ensuring consumers understand that general-purpose LLMs carry no Financial Ombudsman Service or Financial Services Compensation Scheme protection. That warning is useful, but it places the burden on consumers while leaving the underlying legal question unresolved.

The result is potential regulatory arbitrage. Regulated advisers must establish suitability, disclose risks, maintain records, hold appropriate permissions and accept responsibility when advice causes loss. An AI provider may generate comparable recommendations without observing those safeguards, while relying upon the breadth of its system and contractual disclaimers to remain outside regulation.

Our tests provide an obvious starting point for regulatory examination. The FCA could replicate them, preserve the prompts and outputs and analyse each exchange against the statutory tests. It should publish a detailed explanation of when an AI response becomes regulated advice and, where the evidence supports it, consider declaratory or injunctive proceedings against the responsible operator.

Until that happens, the FCA is vulnerable to the criticism that it is tolerating an obvious inconsistency: the same substantive recommendation may be regulated when delivered by a human or specialist robo-adviser but treated as unregulated when generated by a more powerful general-purpose AI platform. Technology neutrality requires a more convincing answer.