S Labs Alternative Credit Infrastructure for the Informal Economy

Abstract

The informal economy of Sub-Saharan Africa accounts for roughly 85% of total employment in Kenya, Uganda, and Ghana, and most of the people working in it cannot get formal credit. This review is bounded to those three markets; Nigeria, the region's largest credit market, is excluded and identified below as the most consequential gap in market coverage. The decade of mobile-based digital lending that followed the launch of M-Shwari in 2012 was a partial correction. It widened access, but it also produced mass blacklisting, over-indebtedness, and outcomes that landed hardest on women and other underserved borrowers. It ran, moreover, on data-extraction practices (contact-list harvesting, device fingerprinting, behavioural telemetry) that are now prohibited across all three target markets. This review sets out the technical and theoretical groundwork for a successor architecture: a privacy-preserving, edge-native credit system built on consented data. The organising claim is that credit exclusion in low-information markets is, at bottom, an information asymmetry problem, and that the structure of peer transaction networks is a form of quantifiable social collateral that can partially close the gap without the extractive practices regulators have moved to stop. Against that frame, we work through four literatures: the empirical record of alternative credit scoring and its failure modes; the regulatory shift that has made extractive architectures legally untenable; the privacy-preserving machine learning stack (federated learning with differential privacy, fully homomorphic encryption, and zero-knowledge proofs) and the real cost each guarantee carries; and graph neural network architectures for financial risk, including their vulnerability to adversarial manipulation. We also ask whether sub-2B-parameter models, quantized to INT4, can run on the low-end Android hardware that target borrowers actually own. Two further problems sit underneath these four and are treated as first-order rather than incidental. The first is that consent in a relational setting is not the same object as consent in an individual one: scoring a borrower from the structure of their transaction graph implicates the counterparties in that graph, and the literature on consent has barely begun to model this. The second is that an architecture built to expand across domains, from finance into agriculture and eventually health, expands its governance surface at the same rate, and the contextual-integrity principle that justifies the credit model also constrains where that data may travel. The most consequential research frontier is not inside any one of these areas. It is at their meeting point: private training of graph-structured models is unsolved, benchmark results have never been tested against African mobile money networks, and the compounding accuracy costs of privacy, quantization, and fairness have not been characterised jointly. We identify the gaps S Labs Finance AI will address, while refusing throughout the comfortable assumption that a privacy-preserving credit model is automatically a welfare-improving one. This revision adds a consolidated execution-risk register and a prioritised contribution roadmap (Sections 5 and 6) synthesised from an internal review of the Phase 1 draft. Coverage is primarily peer-reviewed work from 2019 to 2025, with grey literature (regulatory instruments, industry benchmarks, arXiv pre-prints) included where peer-reviewed equivalents do not yet exist. Jurisdictional coverage is similarly bounded: this review treats Kenya, Uganda, and Ghana as the target markets and does not examine Nigeria's regulatory regime (CBN consumer-protection guidelines, the NDPA) or its competitive landscape (FairMoney, Carbon, Renmoney, Branch Nigeria, Kuda). Given Nigeria's scale, this is flagged as a priority extension rather than a settled exclusion