By Lanre Shittu
Senior Machine Learning/AI Engineer & Researcher | IEEE Senior Member, MBCS
Years before I wrote a line of machine learning code, I learned a rule that never really leaves you: a building doesn’t fail at the roof first. It fails at the foundation, long before anyone notices anything is wrong. You can put up the most beautiful floors in the world on top of a foundation that was never built to hold their weight, and for a while, it will look completely fine. Structural failure rarely announces itself. It builds up quietly, stress by stress, until the day it doesn’t.
I think about that rule often now, watching how fast African tech is racing to plug AI into systems that were never built to carry it.
There’s a real example of this playing out in South Africa right now. The South African Reserve Bank is rolling out its Payments Ecosystem Modernisation Programme, a shift that will let fintechs and other non-bank companies plug directly into the country’s payment clearing and settlement systems, something that used to require going through a sponsoring bank. In practice, that sponsor bank used to quietly absorb a lot of the safety work: fraud checks, reconciliation, the “what happens if something breaks” layer. Once that sponsor is out of the picture, all of that responsibility lands squarely on the fintech’s own systems. Nobody else is holding it up anymore.
That’s not just a metaphor. It’s a real engineering shift, already underway.
The Gap Between Shipping Fast and Standing Firm
This is where a lot of AI conversations go wrong: we’ve gotten very good at shipping intelligence quickly, and not nearly as good at making sure it holds up once it’s out there.
A model can look perfectly production-ready in a demo, trained on clean, well-behaved data, and still fail in ways that have nothing to do with how clever the model is. Three patterns show up again and again in real systems:
- Decisions with No Paper Trail: If a system declines a transaction, flags an account, or approves a loan and can’t explain why, that’s not a small technical gap. In a regulated payments environment, that’s a compliance problem waiting for an auditor to find it.
- Quiet Drift: A system trained on last year’s fraud patterns doesn’t raise its hand when this year’s patterns change. It just keeps answering confidently while slowly getting things wrong more often, and nobody notices until the losses show up in a report.
- Old Assumptions, New Reality: A signal that reliably caught fraud when a sponsor bank’s own controls were doing part of the work might mean something completely different once that sponsor is gone and the fintech is validating everything itself. Nobody went back and checked whether the old assumptions still held once the setup around them changed.
None of these are really “AI problems” the way people usually frame them. They’re the same kind of oversight failures a structural engineer is trained to catch before a building is ever occupied, not after.
Building for the Weight You’re Actually Carrying
The fix isn’t more caution or slower releases. It’s asking a better question, earlier: not “does this model work?” but “what is this system now responsible for holding up, and did we actually design it for that?”
For a fintech taking on direct access to payment rails, that means going back and re-checking the fraud and validation logic that a sponsor bank used to help carry, the same way an engineer recalculates how weight moves through a building once a supporting wall is taken out. In practice, that looks like:
- Being clear about who now owns each safety check that a third party used to handle
- Re-checking what your data signals actually mean whenever the systems feeding them change
- Testing for real-world conditions, not just clean ones, things like slow connections, delayed data, and people actively trying to trick the system
It also means changing how you decide a system is “ready.” Readiness should include being able to explain a decision after the fact, proving the system holds up when the data shifts, and having a process in place to catch quiet drift before it turns into real losses.
Trust Isn’t a Feature. It’s Something You Build into the Structure.
Nobody asks whether a bridge is “innovative” before they drive across it. They ask whether it will hold. As fintechs take on more direct responsibility for the money moving through their systems, that’s the same question worth asking of every AI system in the chain: not how fast it was built, but what it was actually designed to carry.
I didn’t expect a structural engineering degree to be the most useful thing I brought into an AI career. But the instinct it left me with, check what’s holding the structure up before you trust what’s built on top of it, turns out to be exactly the instinct this moment in African fintech calls for.
About the Author
Lanre Shittu is a Senior Machine Learning/AI Engineer with a background in Civil and Environmental Engineering. He builds and deploys intelligent systems at scale, working across LLMOps, real-time anomaly detection, and fraud detection pipelines, with a focus on production-grade, reliable systems engineering.
