FCA AI guidance: Your models are only as good as your customer data

Written by: Stuart Murgatroyd

01 November, 2026

AI can spot patterns humans miss. But what happens when the information it’s working with is already out of date?

The FCA’s latest AI guidance has plenty to say about fairness, governance, and customer outcomes. And rightly so. Yet there’s a problem that deserves more attention. A customer can appear financially stable in your records while their circumstances are getting worse.

They haven’t missed a payment. Their credit file looks fine. Meanwhile, they’re borrowing to cover the weekly food shop or applying for another short-term loan.

Alternative data can help firms pick up on these changes sooner. And with AI playing a growing role in financial decisions, it’s worth asking how much of a customer’s current financial situation your models can actually see.

Key takeaways from this article

  • The FCA’s latest AI guidance puts greater focus on how firms manage AI risks and customer outcomes.
  • Traditional credit data can miss early warning signs of financial difficulty.
  • Alternative data helps AI models recognise changes in affordability and customer circumstances sooner.
  • KYVC provides additional vulnerability insights to support earlier intervention and better-informed decisions.

 

That’s the TL;DR sorted. But there’s plenty more worth knowing. ↓

FCA AI Guidance Infographic

What does the FCA’s latest AI guidance say?

The FCA published its latest examples of good and poor AI practice in October 2026, following the Mills Review earlier in the year. There’s no new AI rulebook to get your head around, thankfully. But the regulator is giving firms a clearer idea of what it expects.

And some of the examples are worth a closer look.

The FCA wants firms to understand how their AI systems behave once they’re up and running. This includes checking whether models produce fair outcomes across different customer groups and spotting problems as circumstances change.

Data quality comes into this too. After all, an AI model can be thoroughly tested and still make questionable decisions if the information it’s using is incomplete or out of date.

Take a customer who’s been making every repayment on time. Your model might see someone who’s managing their finances perfectly well. What it may not see is that they’ve recently lost overtime income and started relying on short-term borrowing to cover everyday expenses.

That’s a fairly significant gap when you’re using AI to assess ongoing affordability.

And it raises an interesting point for firms investing in AI. How confident are you that the customer data feeding your models reflects what’s happening in their lives right now?

The customer looks affordable until they aren't

The customer who looks affordable until they aren’t

Here’s a situation most lenders will recognise.

A customer has been making their repayments on time for months. There’s nothing in their account history to suggest a problem, and their credit file isn’t raising any immediate concerns.

So far, so good.

But their working hours have recently been cut. They’re now using short-term credit to get through the month, and the gap between payday and their next loan application is getting smaller.

From the lender’s perspective, everything still looks fine. The customer hasn’t missed a payment, so there’s been little reason to reassess their affordability.

You can probably see where this is heading.

By the time an arrears flag appears, that customer may have been under financial pressure for weeks. And if an AI model is working with the same limited information, it’s unlikely to spot the problem any sooner.

This is where more current customer data becomes useful. Changes in borrowing behaviour and financial circumstances can give models additional signals to work with, potentially helping firms recognise affordability concerns while there’s still time to offer support.

What can alternative data reveal before a missed payment?

So, what could your AI models pick up if they had access to more current customer information?

Quite a lot, potentially.

Traditional credit bureau data gives lenders valuable insight into repayment history and existing commitments. But financial circumstances can change quickly, and some of the earliest warning signs appear in places a credit file doesn’t always capture.

That’s where alternative data comes in.

For example, a customer might start applying for high-cost credit more frequently. Perhaps they’ve recently changed jobs or their income has dropped. Individually, these changes don’t necessarily mean someone is struggling. But combined, they could give an AI model good reason to take a closer look.

Here’s how different data points can help:

Alternative data  What it could tell your AI model
Increased short-term loan applications A customer may be relying more heavily on borrowing
Borrowing linked to essential living costs Everyday expenses could be putting pressure on affordability
Changes in employment or income Previous affordability assessments may no longer reflect current circumstances
Repeat borrowing over shorter periods Financial pressure may be increasing
Significant life events A customer’s financial situation or support needs may have changed

Of course, the data need context. Applying for a loan doesn’t automatically make someone vulnerable, and a change in employment isn’t always bad news.

But when you’re monitoring thousands of existing customers, having access to this information can help your models identify patterns worth investigating.

This is where KYVC from Data On Demand fits in. It adds data on borrowing behaviour and changing customer circumstances, helping firms build a more current understanding of potential financial vulnerability.

And there’s another reason to pay attention to these trends. Data On Demand’s own research into UK borrowing behaviour gives us a closer look at how financial pressure can develop before it reaches an existing lender’s arrears reports.

What does research say about borrower behaviour

What does Data On Demand’s research tell us about borrowing behaviour?

We’ve talked about the warning signs. But how often do they actually show up?

Data On Demand analysed more than 24 million subprime loan applications, and the findings give us an interesting picture of how people use credit when finances are under pressure.

Almost one in four applications (23%) came from repeat borrowers.

That’s worth thinking about. Someone applying for credit repeatedly may be managing their repayments perfectly well. But if they’re returning to short-term borrowing more frequently, it could be a sign that their income isn’t stretching far enough.

The research also found differences in borrowing behaviour across age groups and locations, with essential household costs featuring in the wider picture of financial pressure.

Now consider what an AI affordability model could do with that information.

If a lender can see changes in borrowing frequency alongside existing account performance, it has more context to assess whether a customer’s circumstances are changing. That could help identify people who might benefit from an affordability review before their repayments become a problem.

Of course, repeat borrowing alone doesn’t establish financial difficulty. The useful insight comes from understanding how behaviour changes over time.

And that brings us to something the FCA is particularly interested in. Adding more data to an AI model creates opportunities, but firms also need to be confident they’re using it fairly.

More data gives AI models more to work with. So what could go wrong?

Quite a bit, if firms aren’t careful.

We’ve seen how alternative data can help AI models recognise changes in customer circumstances earlier. But adding new information also means thinking carefully about where it comes from, how reliable it is, and what your models do with it.

The FCA’s latest AI guidance makes fairness a particular concern. Firms need to understand whether their models could produce poorer outcomes for certain customer groups, including people with characteristics of vulnerability.

And this is where things can get complicated.

Take someone who applies for several short-term loans in quick succession. An AI model might identify that as a potential affordability concern. Fair enough. But what happens next?

If the information prompts a review or an offer of support, it could help the customer. If it leads to an automatic decision that disadvantages them without considering their circumstances, the outcome could be very different.

There’s also the question of accuracy. Employment information changes, borrowing patterns fluctuate, and an indicator that was relevant three months ago might tell you very little about someone’s situation today.

That’s why firms need to test how additional data affects their models, check for unintended bias, and keep monitoring the decisions those models produce.

Of course, this isn’t a reason to avoid alternative data. It means being thoughtful about how you use it.

And before adding another data source to your AI systems, there are a few practical questions worth asking.

5 questions about data feeding AI models

Five questions to ask about the data feeding your AI models

Before adding more data to your AI models, it’s worth checking how well the information you’re already using reflects your customers’ circumstances.

Here are five questions to put to your data and risk teams.

  1. How often is your customer data updated?

A customer’s financial situation can change considerably between credit bureau updates. Check whether your models have access to more recent information when assessing ongoing affordability.

  1. Can your models spot financial pressure before arrears?

Look at the signals your models currently use. Are they picking up changes in borrowing behaviour, or does a customer need to miss a payment before concerns appear?

  1. Which customers might your existing data be missing?

Think about people with limited credit histories or those whose circumstances have recently changed. Alternative data could help fill some of those gaps.

  1. Have you tested how additional data affects customer outcomes?

It’s worth checking whether new data improves identification without introducing unfair bias. Pay particular attention to customers with characteristics of vulnerability.

  1. Can you explain why your model flagged a customer?

If a model identifies someone as potentially struggling, your team needs to understand which signals contributed to that assessment. That makes it easier to review the information and decide what support, if any, is appropriate.

You may already have good answers to some of these questions. Others might reveal gaps worth investigating.

And that’s a useful starting point for looking at what additional customer data could contribute to your existing AI models.

What would your AI models see with more current customer data?

By now, you might be wondering how much financial pressure your existing models are missing. A customer can be keeping up with repayments while their borrowing behaviour tells a very different story. If that information isn’t reaching your models, there’s only so much they can pick up.

That’s where KYVC from Data On Demand can help. By adding more current insights into borrowing behaviour and changing circumstances, firms can give their AI models a better chance of identifying potential affordability concerns earlier.

You don’t have to guess what difference that could make, either.

A KYVC Proof of Value lets you test the data against your own customer portfolio, see what additional insights it reveals, and assess how those insights could support your existing processes.

Want to find out what your current data might be missing? Talk to Data On Demand about a KYVC Proof of Value.

Discover what Data On Demand can do for your business

Learn about how our products could benefit you.
Get in touch and find out about our solutions for Marketing, Verification and Trace.