In the press · 1 October 2026
Can lenders’ chats answer their customers? Only twelve of 31 can even be asked
In the September issue of Kreditnytt, the newsletter of the Swedish Credit Association (Svenska Kreditföreningen), Billuminate founder and CPTO Tim Bisander presents our study of Swedish consumer lenders’ customer chats. Only twelve of 31 lenders could be asked a question while logged out, four questions in ten went unanswered, and roughly one answer in five was wrong.
Source: Kreditnytt, newsletter September 2026, pp. 22–24 (in Swedish). Read the original article in Kreditnytt (PDF).
Tim Bisander
Founder & CPTO, Billuminate
How did we test?
During summer 2026 we went through every company offering consumer loans in Sweden and put questions to their chats. We did what most customers wanting a quick answer would do: logged out, on the website, without calling.
The questions were the kind ordinary borrowers ask, such as whether student finance counts as income, what happens to a co-borrower if payments stop, and how long a payment default stays on record. Every chat got the same questions, and each answer was checked against the company’s own published information.
We started with consumer loans because it is the fastest credit process on the market. Applying takes minutes and the decision a fraction of a second, which makes it all the more important that a customer can get a correct answer before signing.
What did we find?
| What we measured | Result |
|---|---|
| Lenders whose chat can be asked a question logged out, at any time | 12 of 31 (14 have no chat at all) |
| Questions left unanswered | About 4 in 10 |
| Answers containing incorrect information | About 1 in 5 |
| Chats stating that they are an AI | 8 of 12 |
| Chats that also warn answers may be wrong | 4 of 12 |
| Chats answering in Swedish only | 5 of 12 |
The chats did better on questions about the company’s own products than on questions about co-borrower liability, affordability assessments or what happens when a customer cannot pay. The most common error was an answer that contradicted the company’s own terms. Wrong answers were delivered with the same confidence as correct ones, so the customer has no way of telling them apart.
The EU AI Act’s transparency requirement has applied since 2 August 2026. We measured three weeks later. We report what we observed, not who may be in breach, since the regulation exempts cases where it is obvious that the customer is talking to an AI.
Why do the answers go wrong?
According to Tim Bisander, the technology itself is not the problem. When a language model has no reliable content to draw an answer from, it fills the gap with a rough idea of how things usually work, without knowing that it is guessing. Most companies’ content covers the products they sell, not the questions customers actually ask. Many have chosen to let the chat decline to answer instead, which is not a sustainable solution.
Language is a business question too. People who move to Sweden for skilled jobs are often exactly the borrowers a lender wants, yet they meet chats that understand the question but do not answer it in their language.
What should lenders do?
Billuminate’s view is that the combination works best: AI handles the simple, recurring questions, or prepares draft answers for a case handler to review, while complex matters go to a person.
01
Only answer when there's content to answer from
Let the chat hand over to a person when it does not have reliable material, rather than filling the gap with a guess dressed up as an answer.
02
Write for the questions customers ask
Not just for the products you sell. Co-borrower liability, affordability assessments and what happens when a customer can't pay are the questions that actually arrive.
03
Measure against your own terms
Continuously, against your own terms and current rules, not just how satisfied the customer said they were.
What happens next?
Sweden’s new Consumer Credit Act takes effect on 20 November 2026, replacing several of the rules the chats already struggle to describe. We will repeat the measurement after it takes effect and extend it to loan brokers, mortgages and savings.
The full report, without company names, is public on our website.
About this article
This piece is based on the article “Stort test av bankernas kundchattar – få av dem funkar!” (“Large test of the banks’ customer chats – few of them work!”) by Tim Bisander, published in Kreditnytt, the newsletter of the Swedish Credit Association, September 2026 (pp. 22–24, in Swedish). Read the original in Kreditnytt (PDF) · kreditforeningen.se
Tim Bisander, founder and CPTO at Billuminate
Tim has worked with credit models and credit risk since 2000.
LinkedIn →