What Happens When You Connect Your Bank to ChatGPT

I did not link my own bank to test this. I went looking for someone who had. The most thorough review I could find came from a US personal finance site that actually did the experiment and wrote up what worked, what failed and what worried them. I am going to walk you through what they found, what it means for someone in the UK, and the one thing you should do this month before the launch reaches you.

The reason I am writing this now is that the Mills Review, published by the FCA in July 2026, confirmed more than 10 million UK adults could imagine letting an AI manage at least part of their money. That number tells me the demand is real. The product being marketed to satisfy that demand is not yet built for the UK. That gap is the story.

What it does well, even across the Atlantic

Three things come up in every honest review of Finances in ChatGPT. The first is spending anomaly detection. Asking "has anything changed in my spending recently?" pulls a clean narrative across every linked account. The second is subscription audits. A 32 year old software engineer who connected her accounts within hours of the launch had ChatGPT flag four overlapping fitness memberships that were costing her £329 a monthin total. She cancelled three the next morning. The third is cross-account queries that a spreadsheet cannot do in your head, like "what would need to change each month for me to buy a house in five years."

For a UK household at the early stage of sorting their money, those wins are real. They are the kind of thing people spend three hours on a Sunday night trying to figure out from a stack of statements. The chatbot can do it in a sentence.

Where it falls down, even before you reach the UK gaps

The same engineer, two weeks later, asked ChatGPT for the 2026 Roth IRA contribution limit. It gave her a number that was a year out of date. She acted on it. The IRS hit her with a $360 excise tax penalty. The chatbot scored 79 out of 100 on OpenAI's own personal finance benchmark. That sounds high until you realise a 79 means roughly one in five finance answers is wrong. A 47 year old Coast FIRE planner who modelled his retirement date against the chatbot found the maths was off by four years when he cross checked it against a deterministic calculator. Same numbers, four year difference.

The pattern is the same every time. The model is confident. The output looks plausible. It is also wrong in a way that costs you money, and you do not know which kind of wrong until much later.

There is a deeper limitation that does not depend on accuracy. The feature has no category rules, no zero based budgeting, no joint household view, no auto cancel, and a memory model where the only way to scrub a balance from its memory is deleting chats one at a time. It is a conversational layer over your finances. Useful for ad hoc questions. Insufficient as a primary money management system.

The UK gaps that turn the picture around

This is where the launch stops being something you should consider and becomes something you should wait for. ChatGPT has no awareness of the ISA wrapper, no idea that the 2026/27 dividend allowance has dropped to £500, and no model for the interaction between a workplace pension, salary sacrifice, and the annual allowance. Ask it whether you should pay down your mortgage or top up your SIPP and it will give you a generic answer that ignores the UK tax treatment on both sides of the choice.

It has no CGT awareness at a level that survives contact with your actual situation. No inheritance tax planning. No understanding of the UK state pension, the qualifying year rules, or the contracted out differences from 2015 to 2024 that affect what people in their forties are actually going to receive. If you have multiple old workplace pensions from previous jobs and ask the chatbot to model your retirement, it will treat them as a single undifferentiated pile and miss the protected payment amounts, the GMP underpin differences, and the scheme specific tax free cash entitlements that change the answer materially.

None of this is a criticism of the engineers at OpenAI. It is the simple truth that a model trained on the open internet does not know the rules of one specific country until it has been told, and the rules in the UK change every year in the Budget. A US user gets a US-aware assistant. A UK user right now would get a generic assistant with a US shaped data layer.

What I would actually do this month if I were you

The launch is not in the UK yet, which gives you a real advantage. You can do the parts that matter today, on tools that already exist, and ignore the chatbot hype for the next six to twelve months. Three steps in order.

1. Get your complete picture in one place. Link every current account, savings account, credit card, workplace pension and ISA through Open Banking. If you do not have a tool that holds all of it, take the Delphina clarity test first. The whole point of an AI is to reason over the right data. You need the right data first.
2. Decide your actual question. Not "what should I do with my money." A specific one. Am I on track to retire at 60. Should I overpay the mortgage or top up my SIPP this tax year. What do I do with the £18,000 sitting in my current account. The smaller the question, the better the answer you will get from any AI, including this one when it lands.
3. Audit your Plaid exposure today. Most adults in the UK already have four or more stale Plaid connections leaking data to apps they stopped using two years ago. Revoke anything you no longer recognise. Your real exposure surface for any future AI breach is not the new feature. It is the old ones you forgot about.

If you are the person who reads the launch headlines and thinks "should I be doing this," the answer for the UK right now is no, and you are not behind. The product has not launched here. The FCA is still writing the rules for agentic AI in finance. Your job is to make sure the underlying picture is clean and accurate so that when a UK ready version does appear, you can give it the best possible starting point.

If you want the wider context, our guide on what AI can and cannot do with your money sets out the four questions to ask before any app acts on your behalf, and our explanation of guidance versus advice shows where the regulatory line sits.