AI in collections helps prioritise accounts, automate outreach, and predict who’s likely to pay. Yet results are inconsistent. Some teams see improvement. Others invest heavily and see very little change.
It’s worth stepping back and asking: why?
Nearly half of UK adults now show signs of financial vulnerability. At the same time, borrowing behaviour is changing, income is less predictable, and customer circumstances can shift quickly.
But most AI models are still trained on data that only updates after those changes have already happened.
That’s where things start to break.
| What this really comes down to
AI in collections isn’t underperforming because the models aren’t good enough. It comes back to the data they rely on. Two things tend to limit what those models can do:
That’s where results start to plateau. Most teams focus on improving the model. The constraint is usually earlier than that. |
A lot of the value in AI sits in early-stage engagement. Starting conversations, handling simple queries, moving cases forward before an agent needs to step in.
But that only works if the customer can be reached.
If there’s no valid mobile number or email address, nothing happens.
This is where many setups struggle:
You might still identify the right accounts. But without contact, it doesn’t translate into activity.
Once contact data improves, the difference is immediate. More customers can be reached. More conversations happen earlier. More of the process can be handled without manual effort. That’s when the model starts to contribute to outcomes.
Customer behaviour today looks very different to what most collections strategies were built around.
We’ve talked about this a lot. Borrowing is increasingly being used to cover essentials like utilities and household bills, not discretionary spend. In fact, two-thirds of people take out payday loans to cover food, utilities, and transport to work.
At the same time, income is less predictable. Hours change, contracts change, and people move in and out of financial pressure much faster than before.
But the data feeding most AI models hasn’t kept pace.
Bureau data updates on a cycle. Internal systems reflect interactions that have already happened. Neither captures the speed or context of what’s changing in between.
And even when the data is there, it’s often incomplete. If a customer hasn’t explained their situation, or hasn’t been reached in the first place, that change never makes it into the model.
AI is being asked to learn from a version of the customer that no longer reflects how people actually behave. That mismatch feeds directly into how models prioritise accounts, how outreach is timed, and how risk is assessed.
And once that gap opens up, performance starts to level off.
Early signs of financial pressure show up as changes in income, increased reliance on short-term credit, and spending moving toward essential bills.
But those changes don’t appear in traditional datasets straight away.
Bureau data reflects confirmed events. Internal data captures what’s already been handled. By the time either updates, the situation has already moved on.
So the model is always reacting.
It flags risk after it’s visible. It prioritises accounts after the window to act early has narrowed. It supports decisions that would have been far more effective weeks earlier.
That means decisions rely on what’s already been recorded, not what’s currently happening.
And right now, most models are working a step behind.
Even setting timing aside, there’s another issue that’s harder to ignore. The data most models rely on isn’t always consistent or complete. Take bureau data. It’s widely used and trusted, but even the FCA has highlighted differences between agencies on the same individual.
That creates a problem straight away. If the inputs don’t match, the model has to reconcile conflicting information before it can even make a decision.
Then there’s contact data. People change phone numbers, email addresses, and where they live far more often than most databases reflect. That leads to failed contact attempts, wasted effort, and poor engagement. In some cases, each failed attempt can cost between £15 and £50.
So even if the model gets the prioritisation right, the outcome still falls short because the connection isn’t there.
And beyond contact, there are gaps in what the data actually captures. If a customer hasn’t shared a change in circumstances, or hasn’t been engaged at all, that context never makes it into the system.
This is where things start to compound. Incomplete records, inconsistent sources, and outdated contact details all feed into the same system. The model processes it, learns from it, and scales it.
And once that happens, those gaps carry through into every decision the model supports.
So instead, it’s worth getting clear on what AI needs to work properly in a collections environment.
At a minimum, models need reliable ways to identify and reach the right customer. That starts with strong match rates and accurate contact data. If the customer can’t be reached, the model can’t act. If those aren’t in place, everything downstream becomes harder.
The key point is that data needs to reflect what’s happening now, not what was true a month ago. Changes in employment, income, or circumstances have a direct impact on how accounts should be handled.
Context is crucial here. Without visibility of financial pressure or vulnerability, including signals that haven’t been explicitly disclosed, models can’t prioritise appropriately or support the right outcomes.
When contact and context are in place, decisions are based on what’s actually happening, not just what’s been recorded.
That’s when prioritisation, timing, and engagement begin to improve.
When you introduce more current and complete data, the same models begin to behave very differently.
Contact improves because there are more usable channels. More customers can be reached in the first place. Prioritisation improves because decisions are based on what’s happening now, not what happened weeks ago. Customers are approached in a way that reflects their actual situation.
You can see this in the results.
Adding alternative data can improve contact coverage where traditional sources fall short, and increase match rates by a meaningful margin.
That has a direct impact on performance.
More customers are reached. Effort is focused in the right place. Engagement improves because timing and context are more accurate, including signals that haven’t been explicitly disclosed.
And so, decisions are based on a fuller, more current view of the customer.
This is exactly where FinTrace comes into play.
It strengthens the areas AI depends on most, starting with match rates. More customers are correctly identified, which means more records are actually usable from the outset.
From there, contactability improves. Up-to-date phone numbers, email addresses, and address data increase the chances of reaching the customer on a channel that works. And that’s what allows AI to actually engage, rather than just prioritise.
FinTrace focuses on making customers contactable and reachable at scale. Without that, none of the downstream activity happens. What’s more, additional data can then be layered in to provide context around each account, including indicators of financial pressure or vulnerability that aren’t always captured through customer disclosure.
[Alongside this, KYVC (Know Your Vulnerable Customer) introduces indicators of financial pressure and vulnerability that customers don’t always disclose, giving a clearer view of what’s happening behind each account.
As a result, models can prioritise more effectively, outreach becomes more relevant, and decisions reflect what’s happening now rather than what happened previously.
The model hasn’t changed. The inputs have.
There’s a limit to what AI can deliver if the inputs aren’t right. If you want to see where your data is holding AI back, test it.
Run a proof of value using your own portfolio. Measure match rates, contact success, and coverage before and after, including how many customers can be reached and how much more context is available around each account.
The difference is usually obvious within weeks.
Talk to us. We’ll show you what improves when contact data is strengthened through FinTrace and additional context is introduced through KYVC.
Learn about how our products could benefit you.
Get in touch and find out about our solutions for Marketing, Verification and Trace.
© Copyright 2024 Data OD Ltd