The term AI (artificial intelligence) is often used randomly. It is presented as the holy grail, the solution that makes every product better. However many products don’t apply actual AI like OpenAI or Mistral. In reality, it is often just an algorithm, or a third party model. How does this relate to credit management? Is AI relevant and applicable for credit management, or does it belong in the same category as the big data hype?

The goal of AI is for a machine to mimic human intelligence. To achieve that, a lot of data is needed, and AI, as we know it from ChatGPT (OpenAI), works based on Large Language Models (LLMs).

Training an LLM requires a lot of computing power, and thus energy. The larger the LLM, the more it costs. Training ChatGPT 4 likely cost around $100 million. The successor, ChatGPT 5, will have many more parameters and thus be larger. Consequently, the costs are estimated to rise to between $1.7 and $2.5 billion. Such investments are unattainable for most organizations.

Many current AI applications developed by companies, therefore, use existing models from companies like OpenAI and Mistral or open-source models like Llama from Meta.

Advantages of open-source models are that you can run them on your own hardware and customize and train them for specific applications. The costs for training a customized model depend on the modifications and the model used. A pitfall in training is whether you have enough data to train the model. An additional advantage of hosting your own model is that you have full control over privacy.

When looking at credit management, the question is whether AI is the solution to an existing problem and if so, whether it is the best solution. Suppose a customer receives a reminder and cannot pay in full at once. You can include a link in that reminder so the customer can arrange a payment plan online. That link can lead to an AI chatbot or a form on a webpage. Let’s examine the differences in steps of the offered solution.

Option 1: AI Chatbot
The following is a hypothetical but realistic scenario for setting up a payment plan via an AI chatbot.

  1. The customer clicks on the link.
  2. A webpage with a chatbot opens.
  3. Chatbot: How can I help you?
  4. Customer: I would like to pay in installments.
  5. Chatbot: What is the reason you can’t pay?
  6. Customer: I have unexpected expenses.
  7. Chatbot: Do you have more debts?
  8. Customer: Yes.
  9. Chatbot: …

This conversation can continue for a while. The goal seems almost to get the customer to provide a lot of information instead of quickly setting up a payment plan.

Option 2: Web Form
With a web form, it works differently. You might see the following steps.

  1. The customer clicks on the link.
  2. A portal (webpage) opens with an overview of the invoices and several response options, such as setting up a payment plan.
  3. The customer selects the payment plan option and enters what they can pay per month.
  4. The portal assesses whether the payment amount falls within acceptable limits and provides one of the following responses:
    1. The payment plan is approved, and you will receive a confirmation by email.
    2. Your request has been received and is being reviewed. You will hear from us shortly.

The above takes less time for the customer, increasing the chances of success. With this option, an employee only needs to look at the real exceptions. You automate a large part of the tasks and make it easier for the debtor.

The assessment of the requested payment plan in the example is done based on a simple algorithm. This includes the following parameters:

  • Has the customer had a payment plan before and was it adhered to?
  • What is the number of installments?
  • Is the first installment paid within an acceptable period?

The parameters for which the algorithm approves something can vary by type of debtor and organization. We do not call this AI, though it can use historical information in the assessment.

Similarly, a system can determine the most appropriate action within the collection process based on location, age, historical payment behavior, and more. Or you can use it to determine the right form of communication we wrote about earlier. AI can also play a role here, for example, in drafting texts and responses to make the process more adaptive.

A payment plan has a limited number of parameters, and therefore we think a web form is more suitable than an AI chatbot. That was also the response of people we interviewed for this purpose. For more complex processes, it can be a solution. For example, Klarna successfully uses AI for customer service chats.

Besides Klarna, there are many other examples where AI is promising. For instance, in cancer screening, as you can read in this article.

Another area where AI already means a lot is programming. It is difficult for people to remember the details of a large number of programming languages and frameworks. For AI, that is not a problem. With the right instructions, it quickly delivers code, saving time. A programmer then checks the delivered code, and perhaps modifies it, and processes it further.

AI offers opportunities, but it is not the answer to everything, and not everything sold as AI is AI. And customizing and training open-source models is not feasible for most organizations. For now, we therefore conclude that AI in credit management for the vast majority of companies is a hype. Focus on simple algorithms that allow you to make relatively large strides quickly.

p.s. In a year, our conclusion may be completely different, as developments are moving at a rapid pace.