Credit in developing countries has always been a struggle, and it’s not that lenders in these regions don’t know how to lend. In fact, some of the most resourceful lending innovations I’ve seen have come out of Bangladesh, Nigeria, Kenya, India, and Peru. But the problem is that the entire credit environment is stacked against good lending outcomes.
In most of the markets I’ve worked in, whether in sub-Saharan Africa, across large parts of Latin America, through South Asia, or even in the Middle East, there are structural weaknesses that make lending incredibly risky. The rule of law is often shaky or slow to respond. Data systems are fragmented or simply unavailable. And there is usually no consistent legal or reputational consequence when borrowers default.
The idea that a borrower will repay because they’re afraid of what will happen if they don’t simply does not hold. You can miss five loans with five different lenders and still walk into a sixth loan application with no record of the past five. Blacklists exist, but they rarely work the way they should. Courts may technically be an option, but they are expensive, time-consuming, and offer no guarantees. In most cases, lenders are left to eat the loss.
Because of this, we’ve had to rely on what I often call the “spirit of the Lord” to make lending decisions. While I would like that to be some silly joke I shipped in to lighten the mood, it is not. We lean on gut feeling, pattern recognition, and learned instincts. You try to make sense of behaviors and signals that may not be written down anywhere. You might use airtime top-ups, device type, social circles, or informal savings patterns. None of this exists in a formal credit policy document, but it becomes the basis for whether you disburse or not. There is no clean formula, and there is rarely any objective way to explain your decisions. You just go with what your experience tells you feels right.
Now, AI shows up with the promise of order, structure, and prediction. For the first time, it looks like we might be able to build systems that learn from massive volumes of data, understand hidden patterns, and make underwriting decisions that don’t rely on formal employment history or access to traditional banks. In regions where most people live and earn informally, where addresses are fluid and income is unstable, this kind of “tool” feels like a potential breakthrough. It means we could lend to people who have always been excluded, not because they are irresponsible but because they are invisible to the traditional system.
The idea sounds promising. And it truly is. But it is also dangerous. Because for every loan that AI helps you approve correctly, it can also give you a false sense of certainty about decisions you don’t fully understand. If you’ve ever seen what bad underwriting does at scale, you know it’s not something you want to take lightly. In India, for instance, the early years of digital lending saw a surge in default rates and borrower distress, prompting the Reserve Bank of India to issue stricter guidelines for algorithmic credit. In Kenya, digital lenders using behavioral data models were responsible for a spike in loan defaults and aggressive collections, which led to a major regulatory clampdown in 2022. In Brazil, the central bank had to step in with stronger data privacy laws after AI-based lenders were found to be targeting vulnerable consumers with exploitative credit offers.
What this shows is that while AI offers new tools, it doesn’t remove the underlying problems of credit. If anything, it amplifies them. A bad lending decision made manually affects a few borrowers. A bad decision made by a model affects hundreds of thousands, and it does so quickly, silently, and with almost no room to intervene once it’s in motion.
So yes, AI gives us scale and structure. But that structure is only as good as the assumptions it is built on. And in environments like ours, where those assumptions are hard to pin down, that scale can easily become a risk multiplier.
In the absence of law and data, AI becomes a proxy for judgment.
Let’s start with why AI even matters in our context. In countries where the legal system is functional and the credit infrastructure is well-established, lending decisions are backed by real enforcement. People repay loans not just because it is the right thing to do, but because there is a predictable chain of consequences if they do not. A missed repayment damages your credit score. A poor credit score makes your next loan more expensive, or unavailable altogether. That risk of losing access to financial services creates discipline. It builds a feedback loop that reinforces itself over time. This is how credit markets in the US, UK, and parts of Europe and East Asia manage to work at scale with relative stability.
But in most developing markets, the story is very different. There is no systemic punishment for default. A borrower in Ghana, for example, can default on one lender, switch mobile numbers, and immediately apply for a new loan from another lender who has no way of knowing what just happened. In Nigeria, several microlenders have seen borrowers stack five to six loans from different platforms within the same week. India has had similar issues in the past with shadow NBFCs extending loans to customers with no central record of their borrowing history. The lack of integration across systems means there is nothing stopping someone from gaming the market.
So what has kept lending going in these markets? Improvisation. We’ve made do with whatever signals we can find; Phone metadata, airtime recharge patterns, device models, transaction histories on mobile money and social data. None of these were ever built for underwriting, but we have been using them anyway because there has been no better alternative. What AI does is take these scattered, messy, hard-to-interpret signals and turn them into something more structured. It learns from past behavior and tries to predict future repayment capacity. In doing so, it starts to function as a stand-in for formal systems. If you cannot rely on law or data to tell you if someone is trustworthy, AI becomes the thing that fills the gap.
And that is where the hope comes from. If AI can make sense of behavioral data, especially the kind of data people in informal economies actually generate, then maybe we can finally lend to those who have always been excluded. Across markets, there is growing evidence that this might work. In Kenya, some fintechs have used AI-powered mobile data models to reach rural borrowers with no banking footprint. In India, lenders like CASHe and KreditBee have used machine learning to underwrite first-time borrowers based on social and digital signals. In Brazil, initiatives like the Cadastro Positivo have helped incorporate alternative data into national credit scoring efforts, creating new entry points for low-income borrowers. These are early signs, but they show that with the right models, it is possible to design credit systems for people who were never part of the formal banking sector to begin with.
But we should not get carried away. The entire logic behind this promise rests on a very fragile assumption: that AI can understand intent and behavior better than a human can, and that it will continue to do so even when the inputs are irregular, incomplete, or manipulated. This is not a small risk. Because when you remove enforcement and visibility from the equation, what you are left with is a machine that makes decisions in the dark. And that’s not something you can trust blindly.
AI models don’t just reflect data. They reflect bias, too.
This is where the risk really begins to show. A lot of people still assume AI is inherently objective, that once you train a model on enough data, it becomes a pure decision-making tool that is free from human interference. But that’s just not how it works. Every model starts with a set of decisions made by people. We decide which features to engineer, which data sources to prioritize, how to label the training data, what outcomes to optimize for, and which variables to ignore. All of those decisions come from our own perspective. And that perspective is never neutral.
We’re not talking about deliberate discrimination here. No one wakes up and says, “Let’s design a biased model today.” At Lendsqr, we’ve had to build our own credit models, and I’ll be the first to admit that the features we select are based on our understanding of lending patterns. That understanding is based on our experiences in Africa, or in the lending markets we’re most familiar with. We build based on what we think is relevant. And more often than not, that relevance comes from exposure. If you’ve only worked with urban borrowers who use mobile wallets every day, you’ll probably miss the signals from a rural borrower who earns in cash and barely interacts with digital platforms.
This is the quiet way bias seeps in. The model is not making bad decisions because the math is wrong. It’s making bad decisions because the foundation it was built on is incomplete. If your training data is drawn mostly from salaried, male, urban customers aged 25 to 40, then the model will learn to optimize for those types of customers. It will penalize the ones who fall outside that pattern, even if they are just as creditworthy. In India, studies of algorithmic lending platforms have shown that certain groups, particularly women, first-time borrowers, and people from rural districts, were systematically receiving fewer approvals or worse loan terms simply because the training data did not adequately represent their behavior. Similar patterns have been observed in South Africa and Brazil, where informal workers were either excluded outright or assessed using variables that did not reflect their actual repayment behavior.
The hard part is that this kind of bias is invisible until something breaks. You do not notice it when the model is working within familiar territory. But the moment you try to scale or expand to new customer segments, the gaps begin to show. Someone who looks like a good borrower in the real world gets rejected by the model. Another person who should have been flagged as risky ends up getting approved. These mistakes are not always easy to detect, and they are rarely caught early. Meanwhile, real people are affected. They get denied access to credit, mispriced on risk, or quietly pushed to the margins of the system, all without any explanation they can act on.
This is why we have to be skeptical of the language around fairness in AI. Whenever someone says their model is fair or unbiased, the right response is to ask: according to whom? What data was used to train it? Whose behavior shaped the outcomes? What assumptions were made about what risk looks like? Because fairness is not a default setting. It is a moving target, and it depends entirely on whose reality the model is built around.
The moment it works, we rush to scale, and that’s when it breaks
One of the most recurring issues we face in this space is the temptation to scale too fast. It is almost a pattern at this point. You build an AI model, train it on a well-selected batch of borrowers, maybe a few thousand loans. The results look great. Default rates are low. The predictions match what your risk team hoped for. Everything seems stable. There is a quiet confidence that maybe you have cracked something. And then, without fail, someone suggests that it is time to take it national. Maybe even regional.
That is usually where things start to go sideways. Because a model that works on one population size, geography, or behavioral pattern does not necessarily carry over cleanly to a wider and more complex borrower base. The original sample might have shared invisible traits like device type, app literacy, income sources, cultural context. Those are often not factored explicitly, but the model learns to rely on those patterns anyway. And sometimes, the issue is not even about moving across regions. Models that work on a tribe of people in a geography may not even work for a different tribe within the same geography as well. Even within the same city or state, different communities have different ways of living, spending, earning, and using digital tools. So when you begin lending beyond the test segment, even if everything looks the same on the surface, the behavioral assumptions underneath no longer hold.
Even more concerning is how quickly borrowers learn to adapt. In most developing markets, especially those where regulation is still playing catch-up, customers are not passive participants. They observe, learn and test systems. Once they figure out that a certain behavior improves their chances of approval, like maintaining a regular airtime top-up schedule, or ensuring consistent mobile data usage, they will start replicating that behavior, not because it reflects who they are financially, but because it gets them what they want. And when that happens at scale, the entire model becomes distorted. You start lending based on manipulated patterns, not real ones.
We have already seen this play out in multiple markets. In India, several lenders had to quietly reduce exposure in rural states after realizing that their AI models were handing out credit too loosely. The model had been trained on urban digital borrowers with some stability, but rural borrowers had different usage habits and far less capacity to repay. In Kenya, digital credit platforms that relied heavily on SMS and call metadata began seeing spikes in delinquency once customers realized how to mimic the traits the model favored. The signal became noise.
All of this points to the same conclusion. AI models are not static, they can be tricked. They are sensitive to shifts in context and behavior. So while they can be powerful tools for scaling credit, they require constant observation. You cannot simply switch them on and walk away. You need teams watching the data, testing the assumptions, and checking whether the original model is still valid under new conditions. And when the model starts to fail, you need the courage to pause, retrain, or pull back, even if it means slowing down the growth you were hoping to achieve. Otherwise, the consequences show up in the form of mass defaults, broken trust, and reputational damage, things that are far harder to recover from than the time it would have taken to test carefully.
We don’t even fully understand how this thing works.
AI, especially the kind we now use in credit underwriting, is still very much a black box. We feed it inputs, we observe the outputs, and sometimes we can draw a straight line between the two. But when you go beyond simple models into deep learning architectures or complex ensemble decision trees, that clear line becomes difficult, if not impossible, to trace. You might see patterns in the data. You might even see consistent results. But the exact reasoning behind a specific decision is often buried somewhere inside thousands of mathematical layers, and not even the data science team can confidently tell you what the model was “thinking.”
I have sat in meetings where we were reviewing cases flagged by our AI model, and we would come across a borrower with almost no formal income record, no employment stability, and minimal digital footprint, yet the model marked them as low risk and approved the loan. Everyone in the room would go quiet. We would go back and forth, digging into the features, the weights, the training data. Eventually, we would settle on a vague explanation that the model saw something in a combination of variables, perhaps a stable device usage pattern, regular mobile money inflows, or some signal buried in their geolocation metadata. But the truth was, we didn’t really know.
That lack of clarity is not just a technical concern. In lending, it becomes an ethical one. These decisions affect people’s lives. If someone is denied a loan, they deserve to know why. If they are approved for one they cannot afford, the consequences fall heavily on them and on the lender. Yet we are now in a situation where we cannot always explain our own systems. This is a real problem. In places like Brazil, regulators have already begun scrutinizing AI-based credit scoring models specifically because of their opacity.
We have even seen cases where a single, seemingly harmless adjustment to the data, removing a column of zeroes or changing the format of a timestamp, caused the model to behave in completely unpredictable ways. These are not hypothetical risks. They are the kinds of things that creep up silently and only become visible when your portfolio starts leaking. If your default rates begin to rise and you cannot trace the root cause, then you are flying blind.
The core issue here is that AI, while powerful, is extremely sensitive. It is not robust in the way many people assume. It relies on data that is noisy, incomplete, or sometimes manipulated. It draws conclusions from correlations that may not hold in the real world. And once it begins to drift, it does not send up a red flag, you only know something is wrong when the damage has already begun.
So what do we do? Use AI, but use it with care
I want to be clear: I’m not against AI. Far from it. In fact, I am staking a lot on it because I actually believe it’s one of the most important technological shifts we’ve seen in the push for financial inclusion. In markets where traditional underwriting has always excluded informal workers, students, rural families, and the self-employed, AI gives us a real chance at changing that. It allows us to look beyond bank statements and employment records and build lending models around real-life behaviors. That’s powerful.
But it also demands humility. Because we are not working with perfect systems. We’re still in experimental territory, and pretending otherwise is reckless. When we deploy AI in underwriting, we’re making decisions that directly impact people’s ability to survive, run their businesses, or support their families. That’s not the kind of power you hold lightly. It requires discipline, caution and most importantly, it requires honesty; about what we know, what we don’t know, and where we might be wrong.
That means putting systems in place not just technically, but institutionally. Teams need to monitor models continuously, not once a quarter. Decisions that look off need to be flagged and traced back. When we see drift, we must be able to course-correct quickly, even if it means pausing lending altogether. It also means communicating clearly with your team, your board, your regulators, and the customers whose lives your systems affect. AI doesn’t get a free pass just because it’s complicated. If your model denies someone credit, you should be able to give a meaningful reason, or at least admit when you don’t fully understand it. That transparency is what builds long-term credibility.
I’ve seen too many teams, startups and large institutions alike, fall into the trap of overconfidence. A model works in one market or segment, and suddenly everyone assumes it can be scaled everywhere. But most of us working in emerging markets know that every geography is different. Assuming your model is universally valid without validating it at every step is how things fall apart.
India’s Reserve Bank has already raised concerns about algorithmic bias and lack of explainability in credit decisions. In Brazil and Colombia, fintech lenders are now being asked to justify their use of alternative data and prove their models do not discriminate. These conversations are going to keep happening, and rightfully so. The new EU AI Act is one more reminder that regulators are starting to take these risks seriously. Under this law, credit scoring models that use AI have been explicitly classified as high-risk systems. That means providers will be required to meet strict obligations around transparency, accountability, and human oversight. Whether or not you operate in Europe, this signals where global expectations are headed. And sooner or later, every serious player will need to align with that level of scrutiny.
As lenders and providers of lending tech, we owe it to ourselves, and to the customers we serve, to approach AI with a mix of ambition and responsibility.
So yes, let’s use AI. But not like it’s infallible, let’s use it with our eyes open, and with the kind of accountability this space actually demands. Because inclusive finance is not just about scale. It’s about doing right by the people we are trying to serve, especially when the systems we’ve built still don’t fully understand them.