AI in Credit Underwriting: Who Gets Left Out, and What Gets Missed
Everyone talks about AI taking away jobs or AI chatbots. But something quieter is happening too, and it might matter even more. AI is changing who gets a loan from a bank.
Right now, banks look at your money history to decide if you can get a loan. About 1.4 billion people in the world don't even have a bank account. Many others have some history, just not enough for a bank to trust them, so the bank says no, even if that person would actually pay it back (World Economic Forum, 2025). I used real studies, government papers, and company reports for this. If something came from a company selling a product, I'll say so, since that's not the same as a real study.
Some companies now check new kinds of information to decide who gets a loan, stuff a normal bank would never look at. Tala and Branch give loans in East Africa and the Philippines by looking at how someone uses their phone instead of checking a bank record. Nubank and Konfío do something similar in Latin America, looking at things like electric bill payments (Penser, 2020; World Economic Forum, 2025). One report said women pay back loans as well as men, sometimes better, once this new checking removes old unfair habits. Promising, but it's one report that nobody else has checked yet. Might be true. Can't be sure.
It all looks nice at first glance, but not anymore. In Kenya, people who received those mobile loans have suffered greatly. The interest rate charged by such lenders was extremely high. Moreover, some firms were harassing people through their relatives and friends forcing them to repay their debts (Kariuki, 2025; Boston Review, 2026). Finally, the government has taken action on these companies. It has prohibited such firms as well as others to report borrowers to credit offices, while a new regulation has been put in place requiring licensing, fixing interest rate limits and shutting down lenders for non-compliance by 2021 (Boston Review, 2026). Even Tala admitted that it had some issues with its debt collecting system and hired external companies to fix them.
They use peculiar types of data and are thus difficult to test. A model may become discriminatory to particular categories of people without being biased on the surface (Anderson et al., 2026; Partida, 2026). In America, there is an obligation to provide an adequate explanation if someone was denied something based on an algorithm's decision. "It is the computer's decision" does not work (Lexology, 2026; Openlayer, 2026). There is a new legislation in Europe treating the loan approval algorithms as high-risk and requiring good documentation and a human oversight of the computer's decision-making process even though the deadline keeps moving. India has its own unique way of sharing financial data with the lenders – perhaps the largest one worldwide. It is a separate subject, and I will leave it at that.
Another approach is federated learning, through which banks can train a common model without revealing customers' sensitive information. An experiment involving many banks concluded that the resulting common model was more efficient and unbiased (Kaarat, 2026). Interesting approach, should we follow it up. The only thing is that it is more difficult for regulators to evaluate a model which has been trained across multiple banks simultaneously, hence more privacy means less control. Another issue is the model drift, when a model created in a certain economic environment becomes ineffective after a change of conditions. A different issue from bias but still requires constant verification. Buy-now-pay-later apps use the exact same fast approval process as credit lending apps but for purchases instead of credits..
Estimates of where the AI lending market will be in the future vary from $110 billion to over $2 trillion in the mid-2030s (TIMVERO, 2026; GlobeNewswire, 2026). This broad spectrum of estimates results from the lack of a consistent approach to assessing the size of this industry (TIMVERO, 2026; GlobeNewswire, 2026), so no attempts to sum those up should be made. The notion of increasing loan availability and ensuring equitable access to loans not necessarily contradicting each other is based on two researches yet to be validated, so it is nothing more than a promising hypothesis that needs to be tested. If true, then banks are likely to have increased transparency in the process of lending, rather than reduced. The only thing known for sure is that fair treatment of clients and transparency in the relationship with borrowers are not something that comes automatically as a result of increased lending.
Predictive modeling is already quite an advanced tool in terms of identifying potential delinquency of borrowers. What is missing is somebody taking responsibility for mistakes.
Bibliography
Comments