An AI money assistant changes when it can move the money

A chatbot can explain an overdraft. A money agent could move funds to avoid it. Those actions may appear next to each other on a screen, but they do not carry the same authority, access or consequences.

That distinction sits at the centre of the UK Financial Conduct Authority’s Mills Review, published on 6 July. The review looks towards 2030 and beyond, when artificial intelligence might not simply summarise financial information but recommend actions, initiate transactions and execute decisions inside agreed parameters. It is a forward-looking regulatory review, not an announcement that autonomous money managers are generally available, safe or approved by the FCA.

The present is less automated than the forecast. Research commissioned for the review surveyed 5,026 UK retail financial-services consumers in April 2026. It found that 16% were already using AI for at least one personal-finance task. Among those users, the dominant activities were explaining, simplifying and suggesting. Fully delegated decisions were not the norm.

The research then tested hypothetical services at three levels. At the first, AI reviewed information and made recommendations. At the second, it could prepare or take an action after asking permission each time. At the third, it acted autonomously within instructions set in advance and reported what it had done. Willingness to use the concepts fell as autonomy rose: 36% said they would probably or definitely use the assistive version, 30% the version requiring approval for each action, and 20% the autonomous version.

Those figures measure responses to described scenarios, not adoption of products in the market. They still expose the important boundary. “Help me understand” gives a system information. “Do this for me” gives it authority.

In the review’s examples, that authority could mean moving money between accounts, switching a financial product, renewing insurance, adjusting repayments or changing savings and investments within preset rules. The action might be small and useful. It might also arrive at the wrong time, rely on incomplete data, trigger a charge or choose between products whose differences are difficult to reduce to one goal.

Permission therefore needs more shape than a bright consent button. There is a difference between approving one transfer, allowing a tool to act repeatedly under a standing instruction and giving it continuous access to financial data. The survey found that 13% of all respondents would be willing to grant AI real-time access to banking and financial information. Among people already using AI for personal finance, 24% said they uploaded personal data or documents, while 14% gave ongoing access to data, accounts or software.

The data question is not a side issue. A system that watches balances, bills, borrowing and insurance can be more useful because it sees more of a household’s financial life. The same visibility raises the stakes of a breach, an incorrect match, a manipulated instruction or an account connection that stays open longer than expected.

Consumers in the study noticed that trade-off. Fifty-five per cent identified at least one possible benefit from AI in day-to-day money management. Better-value products, more accessible financial guidance and personalised recommendations all attracted support. Yet 68% were highly concerned about misuse of personal or financial data, and 67% about being left without protection if something went wrong.

Recourse is where a polished assistant can become especially confusing. In one survey scenario, respondents were asked about acting on investment information or advice from a general-purpose AI tool. Only 40% correctly recognised that none of the listed formal complaint routes automatically applied. Some believed they could complain to the tool as if it were a regulated firm or take the matter to the Financial Ombudsman Service.

That finding should not be stretched into a universal legal answer. Responsibility can depend on who provides the service, what it does, how it is marketed and which regulated firm sits behind it. It does show why the provider’s identity matters as much as the assistant’s tone. A general-purpose chatbot, a feature inside a bank’s app and an authorised adviser using AI are not the same arrangement simply because all three produce a fluent response.

The FCA’s current approach also stops short of a new AI rulebook. In June, it said it planned to rely on existing frameworks, including the Consumer Duty, the Senior Managers and Certification Regime, and expectations around governance and controls. Its Consumer Duty says regulated firms should avoid foreseeable harm, communicate clearly, provide fair value and offer appropriate support. The regulator is separately studying how firms govern, test, monitor and explain AI-driven outcomes.

The Mills Review makes seven recommendations for the FCA Board to consider, including monitoring the shift towards autonomous models, strengthening coordination, expanding the AI Lab and building foundations for “agentic finance”. Recommendations for consideration are not final consumer protections. The difficult work lies in translating broad duties into permissions, records, limits, human support and complaint routes that remain understandable when software acts quickly.

Five labels will matter in any future money-agent journey: who provides the system; whether it only explains or can act; what data it can see; whether permission applies once or continuously; and where an error can be challenged. None decides whether a product is suitable for an individual. Together, they reveal whether an apparently simple assistant is still a guide or has been handed the keys.

That is the useful reading of the FCA review. The near future of financial AI is not merely a better conversation about money. It is a negotiation over the gate between advice and action, and over who remains accountable after the gate opens.