The Fintech Playbook That AI Is Following

The M-Pesa story begins with a development loan. The UK Department for International Development gave Vodafone £1 million in 2003 to explore whether mobile phones could be used to facilitate microfinance repayments in Kenya. The original application — using mobile minutes as a store of value for loan repayments — was awkward and never properly launched. What Nick Hughes, the Vodafone engineer who led the project, discovered was that Safaricom’s airtime resellers were already informally using airtime as a currency: sending it as a proxy for money transfers, trading it for goods and services. The application he was supposed to build existed; it just needed formalization.

M-Pesa launched in 2007 as a formalization of what the market was already doing informally. Safaricom added the infrastructure — the agent network, the liquidity management system, the mobile interface — that made the informal practice reliable, scalable, and accessible to people beyond the small network of airtime traders. By 2012, M-Pesa had 17 million registered users in Kenya (a country of 40 million people), and the academic literature on its poverty-reduction effects — summarized in a landmark 2016 paper by Jack and Suri — estimated that it had moved 2 percent of Kenyan households out of poverty.

The AI parallel is being constructed with deliberate reference to the M-Pesa model. Several of the organizations building AI tools for low-income markets explicitly invoke mobile money as their template: find the problem the market is already trying to solve informally, add infrastructure that makes it reliable and scalable, price it at a point that the market can afford, and distribute it through existing social networks rather than building new ones from scratch.

The cases where this explicitly works: AI customer service tools that formalize the advice that informal networks already provide, reaching people who currently rely on asking a relative who works in the city or a neighbor who has been through the same situation. Agricultural AI advice that formalizes the knowledge that extension workers have been trying to distribute informally through limited networks of human contact. Medical AI triage that formalizes the “ask the nurse at the clinic” function for conditions that community health workers encounter repeatedly.

The structural similarity to M-Pesa is real. The question is whether the infrastructure requirements for AI parallel the infrastructure requirements for mobile money in ways that make the comparison predictive, or differ in ways that make it misleading.

Mobile money worked with a specific infrastructure equation. The mobile network was already built (for voice and SMS). Handsets were already widely owned. The agent network for cash-in/cash-out was built by Safaricom over two to three years, leveraging the existing airtime retailer network. The marginal cost of extending M-Pesa to an additional user was very low once the network was established. The service became more valuable as more people used it (strong network effects), which drove adoption.

AI tools for developing markets have a partially different infrastructure equation. The mobile network and handsets exist — this is the part that is genuinely analogous. What differs is the compute requirement. M-Pesa’s transactions are computation-light: moving a number from one account to another, checking a balance. AI inference — running a language model to answer a question or analyze an image — requires meaningfully more computation than a money transfer. On-device AI inference is possible for small models, but the quality of small models is significantly lower than the quality of large models running on cloud infrastructure.

This means that AI services for low-income users face a tradeoff that M-Pesa did not: delivering high-quality AI (which requires cloud compute and therefore internet connectivity) versus delivering reliable AI (which might run on-device with lower quality but without connectivity requirements). Mobile money never had to make this tradeoff — all it needed was an SMS connection, which worked even on the most limited handsets with the most limited connectivity.

The credit scoring application of AI is the one that most directly follows the M-Pesa playbook, and it is also the one with the most complicated legacy. Mobile money created a transaction history for people who had no credit history in the formal banking sense. That transaction history is data. AI models trained on that data produce credit scores that can extend credit to people who would have been declined by traditional credit scoring — a real advance in financial inclusion.

The practice has scaled across sub-Saharan Africa and South Asia in the form of micro-lending apps — Branch, Tala, Jumo, and dozens of smaller local players — that use smartphone data (call patterns, app usage, location history, mobile money transactions) to build credit scores. In Kenya, Tanzania, Nigeria, and India, these services have extended credit to millions of people who had no prior credit access. Interest rates are high by developed-country standards — typically 20 to 50 percent annualized — because the default risk is genuinely higher and the operational costs are proportionally larger for small loan amounts. But for someone who would otherwise borrow from an informal moneylender at higher rates, or not borrow at all, the services represent genuine financial access.

The complication is the predatory lending dynamic that has emerged in several markets. Aggressive push marketing, unclear terms, debt collection practices that shame borrowers by contacting their social networks, and credit scoring that does not accurately reflect the user’s actual repayment capacity have produced significant consumer harm in Kenya and India specifically, generating regulatory responses (Kenya’s CRB regulations, India’s RBI digital lending guidelines) that have shut down the worst actors while constraining the legitimate ones.

The M-Pesa story had a relatively clean impact narrative. The micro-lending AI story has a more complicated one: real financial access for many, predatory harm for some, regulatory responses that are still being calibrated. The AI fintech model has followed the fintech playbook precisely enough to replicate both its benefits and its failure modes.

The education AI application is the one where the M-Pesa analogy is most strained, and understanding why is instructive. M-Pesa worked because the underlying service (money transfer) is a transaction with no quality dimension — either the money arrives or it doesn’t. Education is fundamentally different: the quality of what is learned matters enormously, and quality is hard to measure in real time.

AI tutoring tools for developing markets — Khan Academy’s AI tutor (Khanmigo), the various local-language tutoring apps in India and Nigeria, Kenya’s Tangerine learning assessment AI — face the challenge that mobile money never did: how do you know if it’s working? A money transfer either succeeds or fails. A tutoring session either teaches the concept or it doesn’t, and knowing which requires assessment that is harder than a transaction log.

The AI tutoring deployments that have produced the clearest evidence of learning impact are the ones that have been most aggressively paired with human teacher feedback — not AI that replaces teacher interaction, but AI that structures and augments it. The evidence on purely AI-driven tutoring without teacher integration, in low-income classroom contexts, is more mixed. The technology can produce engagement. Whether it produces durable learning at the same rate as teacher-led instruction is still genuinely uncertain.

The mobile money comparison ultimately illuminates more than it conceals, if used carefully. The correct lesson is not “AI will work in developing markets the same way mobile money did” but “the conditions that made mobile money work — existing distribution infrastructure, real unmet need, affordable pricing, network effects — are also required for AI deployment, and they are present to different degrees for different AI applications.”

Where those conditions are most clearly present — agricultural advice, basic financial services, medical triage, government information services — the M-Pesa analogy is most predictive. Where they are least present — high-quality education, complex legal services, anything requiring high-stakes accuracy — the analogy breaks down in ways that require a different analysis.

The fintech playbook works until it doesn’t. Knowing which case you’re in requires more analytical honesty than the development technology community typically applies.