Follow Us on Google
By Rita Okoye
Amid the rapid adoption of artificial intelligence in financial services across Africa, fintech expert Kelvin Efosa Obasuyi has lifted the lid on why AI-powered credit scoring continues to fail millions of borrowers in low-data environments, warning that applying models built for advanced economies to African markets has come at a significant cost.
In his recommendation, which has earned recognition in fintech and thought leadership circles for its practical approach to AI-driven lending, Obasuyi argued that financial institutions must abandon assumptions designed for data-rich economies and instead build credit models around the realities of African markets.
According to him, one of the biggest mistakes made by fintech companies is assuming that machine learning models developed in countries with extensive credit bureau records and formal employment systems can deliver the same results in markets where millions of people have little or no documented financial history.
“The models are sophisticated. The markets are not what the models assume. That gap has been costing African borrowers for years,” he said.
Drawing from years of experience building and scaling fintech products across Africa, Obasuyi said the failures of many AI credit models have less to do with poor engineering than with the misconception that methods created for credit-bureau-rich markets can simply be transferred to low-data environments.
“I have spent the past several years building and scaling fintech products across African markets, watching credit models fail in ways that consistently have less to do with engineering quality and more to do with a single misunderstanding: that methods developed for credit-bureau-rich, formal-employment-heavy, data-dense markets can be translated into low-data environments without fundamental modification.
They cannot. The teams that discover this after deployment discover it expensively,” he stated.
He explained that while customers in developed economies typically generate extensive financial records through payroll deposits, credit histories and banking transactions, the average customer in countries such as Nigeria, Kenya and Ghana often has only a “thin file, or no file at all.”
According to Obasuyi, the consequence is that AI systems trained on data-rich assumptions do not merely become less accurate when deployed in Africa but become “confidently wrong.”
“What a model trained on bureau-dense assumptions does in that environment is not become less accurate. It becomes confidently wrong. Confidently wrong credit models cause real damage—creditworthy customers locked out of formal credit, and lenders booking losses the risk models never flagged because the training data was never representative of the customer base being served,” he said.
The fintech expert noted that the industry’s response has largely focused on sourcing more alternative data, including phone usage, app activity and social media behaviour. While some of these indicators have predictive value, he said many fail to remain reliable across different cities, income groups and changing economic conditions.
Instead, Obasuyi argued that the strongest indicator of creditworthiness in low-data markets is cash flow.
Other News
“What actually works in low-data markets is unglamorous, and it has been working for longer than the current AI wave has been a conversation. Cash flow is the signal. Specifically, the rhythm of money moving in and out of a mobile wallet, a merchant till, an agent network, over time,” he explained.
Although such transaction records may be informal, incomplete and seasonal, he stressed that they reflect real economic activity and already exist at scale across Africa’s digital payment networks.
He further recommended that lenders build underwriting systems capable of analysing noisy and irregular transaction data rather than forcing models designed for clean, structured datasets to operate in environments they were never built for.
“It is less dramatic than a deep neural network on a pitch deck slide. It is considerably more reliable,” he noted.
Obasuyi also urged fintech firms to treat data availability as a business partnership challenge rather than solely a data science problem. He said institutions that have achieved better lending outcomes are those that collaborate with telecommunications companies, merchant networks and payment agents to access authentic transaction histories.
“No architectural decision compensates for a bad input problem. A logistic regression on five years of real transaction data will outperform a transformer model trained on six months of synthetic proxies, every single time,” he said.
Beyond improving data quality, Obasuyi called for greater transparency in AI lending systems, warning that credit models incapable of explaining why an applicant was rejected could expose lenders to regulatory risks, especially as African governments continue developing frameworks for algorithmic lending.
“A credit model that cannot explain a decline is a liability in any market. It is a particular liability in markets where regulators are still building the frameworks to govern algorithmic lending and where a wrongly declined customer has fewer channels for redress,” he said.
While maintaining that machine learning remains vital to expanding financial inclusion across Africa, Obasuyi insisted that success depends on designing AI systems around local realities rather than importing foreign assumptions.
“It argues against the order of operations most teams have been following: import the model, then discover the environment doesn’t fit. The right sequence starts from the ground up. What does repayment behaviour actually look like for someone who has never held a formal bank account?
What data genuinely exists at the scale needed to train on? What does an interpretable, auditable decision look like for this jurisdiction, for this regulator?
“Those questions are harder than tuning a model. They are also the questions that determine whether the model works when real customers’ livelihoods depend on the answer,” he added.
