The Best AI Design Digitizes Trust Instead of Replacing It

On September 28, Disrupt Africa reported on Sandi AI, a Ugandan agri-tech startup founded in 2021 by Douglas Albert Kikonyogo, that lets smallholder farmer savings groups buy solar irrigation equipment with no loan, no collateral and no credit score. A group leader registers the group through WhatsApp, USSD or Sandi's mobile platform. Each member picks the equipment they want. Then the platform does something almost boring: it digitises the rotating savings round the group already runs. Instead of one member taking the pooled cash home each round, the pool buys equipment for one member at a time, in sequence, until everyone has theirs.

There is no underwriting in that description. No income statement, no bureau check, no interest rate. The reason Sandi AI can skip all of it is not that its software is unusually clever. It is that the group already solved the trust problem before Sandi arrived. A farmer savings group is a small, closed, socially enforced credit system: everyone knows who has paid in, everyone knows whose turn is next, and default carries a social cost that a bank's default rate never captures on paper. Sandi is not building a new institution the farmers have to learn to trust. It is putting a digital layer over an institution they already trust, one that predates the app by years.

The design choice that actually matters

Most coverage of AI in African finance runs the other direction. TheCable's recent look at Kenya's approach to banking AI treats the frontier as banks getting better at decisioning: faster credit scoring, better fraud detection, more automated approval. That is a real and useful frontier, and it is the one most executives picture when someone says "AI for financial inclusion." It is also the frontier that requires the most infrastructure a smallholder farmer in Uganda does not have: a formal income history, a national ID linked to a credit bureau, a bank account with transaction depth. Sandi AI's actual contribution is refusing that frontier. It treats the absence of a credit file not as a data problem to solve but as a design constraint to route around.

That is the argument worth making to any executive weighing an AI build for this market: the hard part is rarely intelligence, it is the collateral that intelligence usually needs to operate on. A scoring model is only as good as the financial history it scores. Where that history does not exist, or exists in a form no institution recognises, the model has nothing to work with regardless of how well it is trained. Sandi's answer is to stop asking the model to predict risk and instead ask the software to schedule and verify a savings round that already carries its own risk management, built by the people in it.

The objection that survives contact

The strongest objection to this design is capacity, not trust. A savings group can only buy what it can save. There is no external capital entering the system, no expansion of what the group can afford beyond the sum of its own contributions. A model that expands credit, even imperfectly, moves more capital into the hands of more farmers faster than a model that only reorders capital the group already has. If the goal is closing an investment gap in irrigation, digitising a savings round does not close it, it just makes existing savings more usable.

That objection is correct and does not change the design decision, because the two models solve different problems and carry different costs to operate. A credit model needs loan officers, and loan officer turnover is one of the quiet reasons microfinance portfolios in this region degrade: the person who understood a client's household cash flow leaves, and the next one starts from a spreadsheet. A savings-group model needs no loan officers at all. The verification work is distributed among members who already do it for free, because they were doing it before the app existed. For a business built on the constraint set that matters here, unreliable staffing and thin cash flow rather than headcount and compliance cycles, that is not a smaller version of the credit model. It is a cheaper one to run for as long as the business exists, which is the metric that decides whether a system survives past its pilot.

The USSD and WhatsApp registration path does the same kind of quiet work on the infrastructure side. It asks nothing of bandwidth that a feature phone cannot supply, which means the product does not degrade the moment a farmer moves out of range of reliable data. An executive evaluating a vendor pitch for an AI system in this market should be asking a version of the question Sandi AI answered without being asked: does this system require me to build trust from nothing, or does it let me borrow trust that already exists. Every answer that requires building trust from nothing is going to cost more, take longer, and fail more often than the coverage of AI adoption tends to admit.

The lesson for Monday is not to look for the next scoring model. It is to look for the informal structure your customers already use to solve the trust problem your product is trying to solve, and to ask why you would replace it instead of encoding it.

Sources

Uganda’s Sandi AI helps farmer savings groups access solar irrigation without loans (Disrupt Africa, 2026-09-28)