What AI Can and Cannot Do With Your Money

The honest split is not flattering to either side of the argument. AI is genuinely good at things most people find hard, and quietly poor at things most people assume it can do. Here is the line, with real numbers and the four questions to ask before any app gets to act on your behalf.

What AI genuinely does well

A 42 year old with a £185,000 pension pot, a £24,000 stocks and shares ISA, a £12,000 cash ISA and a £310,000 mortgage does not have a single number problem. They have a layered one. They want to know if they are on track for 58, whether overpaying the mortgage or topping up the pension is the better move this tax year, and what changes if they go part time at 50. Most people cannot run that model in their head. Most financial advisers will not run it for less than £3,000.

A modern AI can. It can stress test the same household at 3 per cent, 5 per cent and 7 per cent real returns, add a child going to university in three years, and show what each scenario does to the retirement date in seconds. It can compound correctly, remember the ISA and pension annual allowances, and label each assumption so you can challenge it. For pattern recognition and scenario modelling, AI is the cheapest modelling tool that has ever existed.

The Mills Review found that around 1 in 4 UK adults would be comfortable with an AI handling at least part of their day to day money. On the technical side, that trust is not misplaced. The maths is reliable. The integrations with UK banks via Open Banking are stable. The risk is not whether the model can count. It is what happens around the maths.

Where AI quietly fails

The same model that is excellent at compound projections cannot tell whether you are about to take a year out to care for a parent, whether a bonus you received last month is a one off or the start of a pattern, or whether the spending category labelled "entertainment" is actually a quiet sign of stress. Those are behavioural and life questions. The pattern is in the data, but the meaning is in the person.

Tax is the second failure mode. UK tax is rule heavy, frequently changed, and full of edge cases that live in the interaction between two allowances, a marriage and a property. AI trained on the internet will not know that the 2026/27 dividend allowance dropped to £500, that the CGT exemption on share transfers between spouses still applies, or that a defined benefit pension transfer can remove your right to a survivor's pension. Ask ChatGPT for a CGT answer today and you will often get a number that is out of date by a year.

Debt is the third. An AI can model a balance transfer and a consolidation loan, but it cannot tell whether the debt is a one off medical bill or a symptom of a pattern that will repeat the month the loan clears. A human conversation is usually the only thing that surfaces the difference. Without that, the AI optimises the wrong thing.

Each of these failures has a common shape. The model is confident. The output looks plausible. It is also wrong in a way that costs you money, and you will not know which kind of wrong until much later.

The accountability gap

This is the part the marketing pages skip. If an AI moves £14,000 from your ISA into a fund that turns out to be inappropriate for your tax band, who is responsible? The model provider will tell you the model is a tool. Your bank will tell you the model is a third party. The FCA will tell you the rules for agentic AI are still being written. The Mills Review, published in July 2026, is the first attempt to close that gap, and it is not yet binding.

Compare that with a financial adviser. They carry personal regulatory accountability. The advice is theirs. If it is wrong, you can complain to the Financial Ombudsman Service and usually recover the loss. The asymmetry is the point. AI that explains your options is different from AI that moves your money. Until the first kind is held to the same standard as the second, the risk stays with you.

Four questions to ask before any app acts on your money

Before you give any financial app permission to do anything, ask four questions. They take under two minutes and they catch most of what goes wrong.

1. Can I see why it made the decision? If the model produces a number without showing the inputs, the number is a fortune teller, not a forecast. A good AI shows the assumptions. A bad one hides them.
2. Can I stop or reverse the action? Anything that moves money should have a pause, an undo and a human override. If the button only says "confirm", the system is not on your side.
3. Who is accountable if it gets it wrong? If the answer is "no one", the system is not ready for your money. Real accountability has a name, a regulator and a complaint route.
4. Is a human involved when the stakes are high? Showing you the consequences of a choice is the AI's job. Making the choice for you, on a pension transfer or an inheritance tax decision, should be a human conversation, not a button.

The verdict for this month

AI is a tool. A good one. The most useful financial tool in a generation for people who can think clearly but do not have the time to model a household by hand. It is not a replacement for the conversation, the life decision or the tax nuance. It is the bit in the middle that turns a pile of accounts into a picture you can act on.

Delphina sits in that middle. The AI does the compound maths, the scenario modelling and the assumption labelling. The human conversation stays with you. The result is clarity, not autonomy. You are still the one making the call.

If you want the wider context on how the FCA is thinking about agentic AI in finance, our guide to guidance versus advice sets out the regulatory split. The aim is the same: know where you are, and choose what to do next with your eyes open.