AI in the debt sector: how can we get the underlying exchange of data right?

12th April 2024

I don’t know about you, but whenever I hear someone talk about AI or Artificial Intelligence, I feel somewhat sceptical, even scathing. When I was quite a bit younger, I remember watching Tomorrow’s World on the BBC, learning all about something new called an ‘Information Superhighway’ that had been invented and was going to change our lives. I seem to remember there also being forecasts that robots would in the future do everything for us, our houses and clothes would all look incredibly futuristic, and we’d be travelling in cars hovering above the road or flying through the air… To say I’m disappointed might be overstating it a bit as things have largely stayed reassuringly similar, albeit the ‘Internet’ as we now know it, has delivered a lot more than forecast and revolutionised the way we communicate and share information and views to say the least.

So, when I hear companies hailing the advent of AI in financial services, I’m still expecting great things, the sorts of which I was promised by Raymond Baxter and Judith Hann all those years ago. In reality, I don’t think we are yet seeing any ‘artificial intelligence’ of the type we were promised back then. It feels more like process automation at the very best, rather than machines actually learning anything – not saying automating processes can’t be really useful, of course.

Bear with me for a moment if I sound like I’m going off on a tangent here, but I was similarly under-impressed by all the fuss when ChatGPT was first launched if I’m honest – whilst I don’t understand whether it really is AI or not, I wasn’t in the slightest impressed with the results it churned out. The problem is that whilst ChatGPT may for all I know, be incredibly clever, not all the information it was looking at is… well, terribly accurate. Most of it has been put there by people who are possibly not qualified to do so and in the main, without any validation.

In talking to a couple of my developer colleagues at Trustfolio, both much cleverer than me, they tell me that ChatGPT is actually really good at churning out high-quality computer programming code when it is asked to do so. So why is it really good at that and yet not when I asked it to write some text describing how a specific debt solution works for an article I was writing, or when I tried to get it to write a business policy for me? To start with the results from those queries looked impressive, but once I started reading in detail, they were both quite useless.

Discussing this with my clever colleagues, we decided that it’s all down to the data available. It’s likely, they think, that the code ChatGPT can produce, is drawing on good quality data placed on the internet by people who know what they are doing, whereas a lot of general so-called ‘facts’ are written by a wide variety of people, some good and some not quite so good. I believe the latest version of ChatGPT can pass a bar exam at law school. Maybe it’s learned from a better source of information than the debt-solutions descriptions I was looking for earlier.

But that all made me think. If we want AI to work in debt collection, don’t we have to get the underlying data right first?

What worries me, especially in the debt solution and debt collection space, is the lack of quality, meaningful data being shared between creditors, advisers, debt collectors and the like. The way data is exchanged is a problem too with, believe it or not, still paper being physically posted between organisations, or at best, electronic exchange of word documents, PDFs, excel spreadsheets and in the case of one or two advice organisations, electronic portals of varying descriptions and levels of sophistication.

When I joined Trustfolio I spoke with a lot of debt advisers, collectors, buyers and creditors about the way IVA proposals and debt management plan proposals were being exchanged between organisations.  What I learned then was that in most organisations, there was still a ‘cottage industry’ approach to making all this work – lots of people dealing with a wide variety of data transmission methods and data sets, not much consistency, and lots of room for error and delays.

What I’ve learned since, is that the way forward is to digitise as much of this as possible, taking and receiving quality data using APIs – data straight from systems, not spreadsheets – ideally via one independent route – and as soon as we can. It’ll increase accuracy, improve efficiency, and allow us to use the tools that are available to us – even AI. But it’s not a quick fix. There’s a lot of work to do in standardising the data that is collected and exchanged, never mind how the data is actually exchanged. The question is whether there is a desire to change this quickly enough.

One thing that might speed the process up could be the FCA’s Consumer Duty – are authorised firms able to demonstrate effectively enough, that they are delivering the right outcomes to customers? The FCA places a lot of weight on management information and analysis of data to prove the right outcomes are being achieved. Can authorised firms do that with the data they hold, now? If they’re not even ready to do that, they’re not in my opinion, ready to use AI.

Peter Wallwork, Trustfolio