Across African institutions, there is a pattern where vendor demos arrive before anyone asks the key questions about training data and model origins. Unfortunately, this results in the adoption of AI models that do not reflect the reality of African organisations. The solution lies in standardisation and governance before adoption of models
I recently signed up for ISO/IEC 42001 training and a colleague asked why someone who already spends her days thinking about cyber risk would sign up for yet another certification. My answer was that AI is already inside our organisations, and most of it arrived without a governance plan.
Working across six African jurisdictions, I keep seeing vendor demos and impressive dashboards. However, questions of where the model was trained, on whose data it was trained and who is accountable when the model makes a wrong decision arrive late. Too often, the honest answer to these questions is "we are not sure."
Much of the AI running in African institutions is designed, trained and hosted far from the people it serves. The data rarely reflects our languages or how our customers actually live. A credit model trained on salaried borrowers elsewhere may misread a market trader whose income arrives daily through mobile money. When models live offshore, our regulators and data protection laws are always one step removed.
The need for standards
While it is tempting to jump straight to building our own, ownership without standards just moves the problem closer to home.
Standards give us a common language. When every system, local or imported, is documented and monitored the same way, we can question it and, when needed, say no.
Standards also give buyers leverage. Today, many African institutions negotiate AI contracts on price, tokens and features. A recognised framework lets a procurement team ask for impact assessments, data provenance and accountability before signing.
In addition, standards travel. Our continent is a patchwork of data protection laws and emerging AI policies. A shared international benchmark gives regulators, boards and partners across different countries Nairobi, Lagos and Kigali one reference point.
S𝘁𝗮𝗻𝗱𝗮𝗿𝗱𝗶𝘀𝗮𝘁𝗶𝗼𝗻 𝗶𝗻 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲
ISO/IEC 42001 is one of the international management system standards for AI. It treats AI as something that must be planned, risk assessed, controlled, monitored and improved just like we treat information security. It does not ask whether AI is good or bad. Rather, it asks whether you are in charge of it.
We have done this before. Many African banks adopted ISO 27001 for information security long before local rules required it. It raised the floor, built a generation of skilled practitioners and made our institutions credible to global partners.
Successful standardisation has the following features:
- An inventory of every AI system in use, including the ones hiding inside vendor products.
- An impact assessment before deployment, not after the complaint.
- Model documentation and accountability clauses written into contracts.
- Data governance that makes local data trustworthy, and therefore usable.
- Certified implementers and auditors based on the continent, not flown in.
The order in standardisation matters because teams that govern AI well learn how it works. This is important because investors and partners will back infrastructure that can show how its AI is controlled. Standards turn "trust us" into evidence, and evidence is what attracts the capital for local data centres and models.
Standards are demanding, and turning them into everyday practice in busy institutions is real work. Some days, the gap between policy and practice feels wide. But every risk assessment and every honest conversation with a product team narrows it a little. Tech leaders across Africa must adopt a recognised AI governance standard before regulators ask them to. The same goes for anyone who wants to buy AI; they must make governance evidence part of every contract. Further, policymakers must align with international standards so that our institutions speak one language.





