An AI trained on the world’s software still arrives at an African bank knowing nothing that matters there. For banking boards, the real task is teaching it what does.
A single change to a lending rule can take months to ship inside a bank, and the code is the easy part. What takes the time is knowing what else that rule touches: an AML report three systems downstream, a pricing model built by a different team, a reconciliation job that runs overnight and depends on a field nobody has looked at since it was written into the system twelve years ago.
Generative AI can write that code in minutes, but it cannot answer any of those questions. AI knows the world. It doesn’t know your bank. In Africa, it knows even less.
Intellect Design Arena has placed its latest bet on that difference. This week the company launched MSOCK, a knowledge framework backed by 39 filed patents.
Why the knowledge that matters here is African
South Africa runs one of the most sophisticated banking sectors on the continent, competing against digital banks and fintechs while carrying decades of regulatory and technical complexity most challengers never had to build. Move into the rest of Africa and the picture gets harder: banking groups spanning multiple countries, currencies and regulatory regimes, each balancing group-wide standardisation against local rules while running older and newer technology side by side.
None of that complexity is generic. A model trained on the world’s banking software still knows nothing about how a cross-border group reconciles across three central banks, how a mobile-money rail settles overnight, or how a bank prices credit for a customer who has never held a formal account. That knowledge is African, it lives inside each institution, and it does not transfer from a bank in New York or London. Teaching an AI to change an African bank means teaching it what that bank already knows, and almost all of what it knows is African.
Institutional knowledge as the edge
For decades, banks competed on capital, distribution, product and, more recently, data. Two banks can run the same foundation model on the same cloud infrastructure, staffed by engineers of similar calibre, and still end up in very different places. The bank that gives AI a governed, detailed understanding of its own architecture, rules and dependencies moves faster and with more confidence than the one that doesn’t.
Whatever a bank adopts has to coexist with what it already runs, not replace it. I think banks will soon worry less about which model they have bought than about whether their own institution can be read by any of them.
Making the bank computable
MSOCK – short for Multi-Dimension, Multi-Layer System of Connected Knowledge – sits between AI tools and a bank’s own institutional knowledge, rather than replacing the foundation models, copilots or agents a bank has already bought. It asks whether the AI understands this specific institution, or only software in general.
Relational databases solved a version of this problem for computing decades ago, turning fragmented information into something structured and queryable. A bank holds the same kind of problem at a much larger scale: business rules, application code, APIs, customer journeys, infrastructure, regulatory obligations and years of architectural decisions, most of it scattered across documents, code repositories and people’s heads. Internally we describe MSOCK’s architecture as seven connected knowledge layers, running from user journeys down to infrastructure and operations, mapped using what we call 21-dimensional visual spatial intelligence.
The result is an AI that can answer “if I change this, what breaks?” before a developer writes a line. Say a bank wants to change how it assesses a lending application. Today that means weeks of manually tracing dependencies across applications, interfaces and databases. Mapping the “blast radius” of that change surfaces every upstream and downstream dependency first, so nobody starts building blind.
Our own early projects have compressed an eleven-month build into six, with one project’s effort dropping from 170 person-months to 50. Even a fraction of that holding at scale would change what banking technology costs to run.
Africa can leapfrog again
Mobile leapfrogged fixed-line telecoms here. So did mobile money, skipping branch banking in markets that never built one. Each of those leaps worked because it was built for how money actually moves here, not lifted in from somewhere else. AI-native engineering could be next, letting African banks skip a decade of inherited software complexity instead of carrying it forward.
But that will take governance regulators can trust, with audit trails that hold and a human who stays accountable when something breaks. Probabilistic experimentation alone will never be enough for a system that moves people’s money and livelihoods. Deterministic engineering and generative speed can run together, each keeping the other honest.
Every rand, dollar, kwacha or pula lost to avoidable technology complexity is capital an African bank cannot spend on reaching an SME faster, or on building something a customer would actually notice. That is the prize on the table, and it has very little to do with which large language model ends up winning.
