Many of us already use artificial intelligence to help us make decisions.
We ask it to compare products, explain unfamiliar terms, summarise documents or identify questions we should ask a professional.
The next stage is considerably more consequential.
Instead of merely asking an AI assistant to find a bank account, people may authorise it to open one. Instead of asking it to review an investment portfolio, they may permit it to buy and sell investments on their behalf.
At that point, AI is no longer simply providing information.
It is acting.
Writing in the Financial Times, Schroders chief financial officer Meagen Burnett examines how regulators should respond to this emerging world of “agentic finance”. Her argument is that simply subjecting general-purpose AI tools to the Financial Conduct Authority’s existing rulebook could deprive consumers of valuable assistance without adequately addressing the broader problem.
I think she is right that this requires more than forcing a new technology into an old regulatory category.
But there must also be a clear threshold.
Once an AI system is authorised to move someone’s money, enter an agreement or alter their financial position, “check the answer” is no longer an adequate form of consumer protection.
Advice and action are not the same thing
There is a meaningful difference between an AI assistant saying:
You may wish to consider moving some of your savings into an account with a higher interest rate.
And that same assistant transferring the money itself.
The first may influence a decision. The second executes it.
That distinction matters because financial decisions do not always have a convenient undo button. An automated system could sell an investment during a temporary market decline, move money into an unsuitable product, misunderstand a person’s tax position or accept contractual terms that the consumer never properly examined.
The question therefore cannot be only whether AI is engaged in a traditionally regulated activity.
We must also ask:
- What authority has the person delegated?
- How significant are the possible consequences?
- Can the action be reversed?
- Was the person warned before it occurred?
- Who is responsible if something goes wrong?
- What route does the consumer have to challenge the decision and recover their money?
An AI system that explains a mortgage is not identical to one that applies for the mortgage, selects the lender and accepts the terms.
The regulatory response should recognise that difference.
Convenience can disguise delegation
Part of the attraction of AI agents will be that they remove friction.
Instead of comparing dozens of savings accounts, insurance policies or investment products, a person could tell an assistant what they want and allow it to handle the rest.
That sounds useful because it is useful.
Financial services are often complicated, expensive and filled with language that ordinary consumers are expected to understand despite encountering it only occasionally. AI could help people ask better questions, identify unreasonable fees and gain access to information that was once practically reserved for those who could afford professional assistance.
We should not regulate away those benefits.
However, removing friction can also remove the moment in which a person realises the significance of what they are doing.
A consumer may believe they have asked an assistant to “find a better investment” without appreciating that they have authorised it to sell existing assets, trigger tax consequences or expose their savings to greater risk.
The easier an action becomes, the more important it is that the system makes the consequences visible.
Convenience must not be confused with informed consent.
Who exactly made the decision?
When a human financial adviser gives unsuitable advice, there is at least a recognisable person or regulated firm whose conduct can be examined.
Agentic AI could make responsibility much more difficult to locate.
The consumer may have instructed the AI assistant. The assistant may have relied on incomplete personal information. Its underlying model may have been developed by one company, incorporated into an application by another and connected to a financial institution through a third-party service.
The bank may say it merely executed an authorised instruction.
The AI provider may say the consumer approved the action.
The consumer may say they never understood what the system was going to do.
Everyone may have participated in the transaction while nobody accepts responsibility for the decision.
That cannot become the default arrangement.
A person should not lose meaningful protection simply because the chain between intention and action now contains an algorithm.
If a company designs, supplies or permits an AI agent to execute financial transactions, responsibility should be allocated before harm occurs—not improvised afterwards while the consumer is passed among providers.
The answer cannot be a disclaimer
Companies may be tempted to manage this risk through warnings:
AI-generated information may be inaccurate. You remain responsible for all decisions.
That may be reasonable when someone uses a general-purpose tool to explore an idea.
It becomes far less convincing when the system has been deliberately designed and authorised to act.
A disclaimer cannot transform an autonomous transaction into a purely personal decision. Nor should a company be permitted to benefit from presenting its agent as capable, personalised and intelligent while describing it as an unreliable tool the moment something goes wrong.
The more control the system exercises, the less credible it becomes to place the entire burden on the consumer.
Responsibility should follow control.
Financial regulation cannot solve the whole problem
Burnett raises an important challenge: general-purpose AI does not remain neatly within the borders of financial services.
The same assistant could be asked about an investment in the morning, a health symptom in the afternoon and a legal dispute in the evening. Regulating it as a financial institution whenever it discusses money may produce a fragmented system in which different sectors impose conflicting obligations on the same underlying technology.
There is also a risk that providers simply block useful questions.
If financial queries become too legally difficult, AI companies may prevent their systems from explaining investments, insurance, credit, taxation or state benefits. That could particularly disadvantage people who cannot afford personalised professional advice.
The broader answer may therefore require regulation that follows capability and consequence across sectors.
A system that merely supplies general information might face duties concerning accuracy, transparency and the identification of uncertainty.
A system that recommends personalised action would require stronger safeguards.
A system authorised to execute consequential decisions—whether financial, legal or medical—should face the highest level of oversight, record-keeping, confirmation and accountability.
The relevant question is not simply:
What industry is this AI talking about?
It is:
What power have we allowed it to exercise over a person’s life?
Human confirmation must mean more than clicking “approve”
Companies may respond by requiring users to confirm an action before it is completed.
That safeguard is only meaningful if the person understands what they are confirming.
We already click through terms, cookie notices and software agreements that few people realistically read. Agentic systems could reproduce the same ritual: a large “approve” button following an explanation too technical, too vague or too conveniently reassuring to support an informed decision.
Meaningful confirmation should tell the consumer, in plain language:
- What the system is about to do.
- Why it selected that action.
- What information it relied upon.
- What it does not know.
- What the likely costs and risks are.
- Whether the action can be reversed.
- Whether a regulated human has reviewed it.
- Who will be responsible if the transaction is disputed.
For especially consequential or unusual actions, the system should pause rather than proceed instantly.
Sometimes friction is not a design failure. It is a safety mechanism.
Access to a human still matters
AI may help close financial knowledge gaps, but it should not become a cheaper barrier placed between consumers and accountable human support.
People will need access to a person when a recommendation is unclear, a transaction is contested or their circumstances do not fit the system’s assumptions.
This is particularly important because financial vulnerability is not always visible in the data. Bereavement, cognitive decline, coercive control, financial abuse or sudden distress may affect a person’s decisions without appearing in an ordinary customer profile.
A well-designed system may help identify some of these risks. It should not be presumed capable of resolving all of them.
The option to reach an accountable human must remain real, understandable and reasonably accessible—not hidden behind an automated loop after the money has already moved.
The badge of regulation must still mean something
Burnett argues that good regulation could become a source of strength for the UK. I agree.
As AI-generated services become more common, consumers will need reliable ways to distinguish between general information, personalised recommendations and regulated financial assistance.
That distinction must be visible before a person acts.
People should know whether the system they are using is regulated, what protections apply and whether they have access to compensation or formal redress if something goes wrong.
Regulation should not exist merely to make innovation slower. Its purpose is to make trust possible where the consequences of failure are too significant to leave entirely to private promises.
AI has genuine potential to make financial information more accessible and to help people make better decisions.
But access and accountability are not opposites.
We should not force every general-purpose AI assistant into a regulatory structure designed for traditional financial institutions. Nor should we allow companies to escape responsibility by describing consequential financial actions as merely another feature of a general-purpose tool.
The line should be drawn around power.
If AI can explain, we need transparency.
If AI can recommend, we need safeguards.
If AI can act, we need clear responsibility.
Because once a machine can move someone’s money, the question is no longer simply whether its answer was helpful.
It is who answers for what it did.