by Lanre Shittu
Senior Machine Learning/AI Engineer and Researcher | IEEE Senior Member, MBCS
Years before I wrote any machine learning code, I learned a rule I’ll never forget. That is, the roof will not collapse first. It fails from the ground up long before anyone realizes anything is wrong. You can put the most beautiful floor in the world on top of a foundation that isn’t built to hold weight and it’ll look perfectly fine for a while. Structural defects rarely manifest themselves. It silently accumulates stress until the day the stress is gone.
I’m thinking a lot about that rule now, seeing how quickly African tech companies are incorporating AI into systems that were never built to carry it.
An example of this is currently happening in South Africa. The South African Reserve Bank is rolling out a payments ecosystem modernization program. This will allow fintechs and other non-bank companies to connect directly to the country’s payment and settlement system, which previously had to go through a sponsoring bank. In reality, its sponsor banks had quietly absorbed a lot of safety work, including fraud checks, reconciliations, and layers of “what happens if something breaks”. Once the sponsor is no longer involved, all of that responsibility will be placed directly on the fintech’s own systems. No one puts up with it anymore.
It’s not just a metaphor. This is a real engineering transformation, and it’s already underway.
Gap between prompt shipping and stable service
This is where a lot of AI conversations go wrong. We’ve been very good at delivering intelligence quickly, but not so good at making sure it’s maintained once it’s out there.
Even though a model may seem completely production-ready in a demo and has been trained on clean, working data, it may fail in ways that have nothing to do with how smart the model is. In real systems, three patterns appear repeatedly:
- Decision making without paper records: If a system rejects a transaction, flags an account, or approves a loan but can’t explain why, that’s no small technology gap. In a regulated payments environment, this is a compliance issue waiting for auditors to discover.
- Quiet drift: A system trained on last year’s fraud patterns won’t raise its hand when this year’s patterns change. While you continue to answer confidently, you gradually make more mistakes and no one notices until the losses show up on your report.
- Old beliefs and new realities: Signals that reliably caught fraud when the sponsoring bank’s own controls were doing some of the work can have a completely different meaning once that sponsor is gone and the fintech itself is verifying everything. No one was looking back to see if the old assumptions still held after the surrounding settings changed.
None of these are real “AI problems” as people often think of them. These are the same types of oversights that structural engineers are trained to spot before a building is occupied, rather than after it is occupied.
Train your body according to the weight you are actually carrying
The fix isn’t about being more careful or releasing later. We’re asking a better question than the earlier question, “Does this model work?” But, “What is this system now responsible for maintaining? And did we actually design it for that?”
For fintechs that take on direct access to payment rails, that means going back and revisiting the fraud and validation logic that sponsoring banks used to help execute, in the same way that engineers recalculate how weight moves within a building after supporting walls are removed. It actually looks like this:
- Clarify who now owns each safety check that was previously handled by a third party
- Double check what the data signals actually mean every time the system supplying the data signals changes
- Test not just clean situations, but real-world situations such as slow connections, data lag, and people actively trying to trick the system.
This also means changing how the system determines when it’s “ready.” Preparation must include being able to explain decisions after the fact, being able to demonstrate that the system can withstand changes in data, and having processes in place to catch quiet fluctuations before they result in real losses.
Trust is not a feature. It’s what you build into the structure.
No one asks whether a bridge is “innovative” before driving across it. They ask if it holds true. As fintechs take more direct responsibility for the funds moving through their systems, it’s the same question worth asking of every AI system in the chain. It’s not about how fast it was built, it’s about what it was actually designed to carry.
I never expected that my structural engineering degree would be the most useful thing I brought to my AI career. But the instinct it left me with, to see what underpins a structure before trusting what is built on top of it, turned out to be exactly the instinct needed at this moment in African fintech.
About the author
Lanre Shittu I am a senior machine learning/AI engineer with a background in civil and environmental engineering. He focuses on production-grade, reliable systems engineering, building and deploying intelligent systems that work across LLMOps, real-time anomaly detection, and fraud detection pipelines at scale.
