
Data errors are nothing new, but AI changes how far they travel and how convincing they sound. Consider a person in their early sixties who has recently moved to a new pension provider and is facing a series of consequential retirement decisions with no history of professional advice to draw on. Shortly before retiring, they move £30,000 from a maturing savings account into their current account, which their provider reads through Open Banking with their consent. This is completed through a series of transactions over a three-month period.Â
To a system reading raw transaction data, those payments being received look like regular income. As a result, they are placed in a cohort of customers with spare monthly income. Their provider’s app sends an AI-generated message suggesting they increase their pension contributions. In reality, they are running down their savings to bridge the gap until their pension starts. The suggestion is not correct for their circumstances, yet it arrives sounding personal and convincing.Â
The cause is a data error: savings were misread as income. What AI changes is everything that happens next: how far the error spreads, whether anyone checks it before it lands, and how convincing it sounds when it does. Â
According to the FCA, 23 million people are currently underserved by advice and guidance. Targeted support, an outcomes-focused regulatory framework, which went live on 6 April 2026, was introduced to remedy this gap. It allows firms to make suggestions to groups of customers who share a similar circumstance, the potential being millions more people can receive support more tailored to their circumstances.Â
The consequences of acting on false dataÂ
This is not a case of AI hallucination but a product of bad data. The model has reasoned correctly over something false.Â
The fallout and risk exposure from AI suggestions such as this are significant. Whereas a human adviser’s mistake might only reach one client, AI support can reach an entire cohort at once. One data error becomes the same wrong suggestion, repeated across every customer sorted the same way.Â
There may be no human checkpoint before the message reaches the customer. The error does not surface as one complaint from one unhappy customer, but sits quietly inside a cohort, affecting everyone who was sorted the same way, until something prompts an investigation. Monitoring for this kind of failure means watching not just whether the AI’s output is coherent, but whether the data feeding it was ever right in the first place, which is a much harder thing to test for at scale.Â
A further difference is how convincing the error looks. A rules-based tool that misplaced this customer would produce an output someone could check. An AI-generated message instead offers fluent, personal-sounding reasoning, which makes a wrong placement more persuasive to the customer and harder for the firm to catch, because the logic sits in prose rather than in a calculation that anyone can audit.Â
This challenge of bad data categorisation is not confined to targeted support. It runs through other obligations built on the same underlying data. Under Consumer Duty, firms are expected to identify customers in vulnerable circumstances or financial difficulty. Affordability assessments in lending rest on similar judgements. Those judgements depend on exactly the kind of transaction-level inference that misread this customer’s savings as income. If AI then acts on those misjudgements, it repeats them at scale.Â
According to PS25/22, the FCA states its rules don’t prevent firms from verifying the accuracy of the data they hold, and that doing so would be a reasonable step toward meeting its requirements. The same document sets out record-keeping expectations on segment design, customer alignment, and outcome monitoring, giving firms a concrete way to evidence the outcomes.Â
The role of Open Banking and Open FinanceÂ
For AI-generated suggestions in financial services, the fix for the data problem begins with better data categorisation, supported by Open Banking today and, in time, Open Finance.Â
This means that rather than relying on a single provider’s partial view of a customer, built from whatever transactions happen to pass through one account, Open Finance is intended to extend the access Open Banking already provides to payment accounts to savings, investments and pensions, building a much fuller financial picture over time. A more complete profile of the customer can then be created.Â
That breadth matters because the kind of misreading this customer experienced is, at its root, a visibility problem. A system working from one account sees £30,000 arrive and has no way of knowing it came from a savings account and so produces an incorrect suggestion. With a consolidated view, the transfer is given more context, so more appropriate support can be given. Better categorisation of the transactions themselves is possible today; seeing the savings account the money came from generally depends on Open Finance extending access beyond payment accounts.Â
Moneyhub’s Categorisation and Enrichment Engine is built to address the categorisation problem, applying enrichment to raw transaction data so it can be reliably sorted rather than taken at face value. It is this layer, as much as the breadth of data itself, that determines whether a firm ends up with the right picture of a customer or simply a bigger version of the wrong one. So far, it has categorised transactions with 98% accuracy.Â
Let us return once more to the customer and consider a different outcome. Instead of receiving the wrong suggestions, their transactions are correctly categorised and recognised as savings moving between accounts. They are then placed in a cohort of people who are drawing down their savings, and the support they receive reflects that.Â
There is far less risk of this customer making the wrong financial decision. Over time, as additional transactions are made, Open Banking, and in time Open Finance, helps build a more comprehensive profile of this customer, reducing the risk of bad data delivering the wrong financial support. Monitoring of both the data and the AI’s output will still matter. Bad data is an old problem, but AI raises the stakes, which is why the data must be right before the AI acts on it. Â



