top of page
Blurry Blue_edited_edited.jpg
Search

AI Done Right in Africa: What Happens When AI Stops Pretending Conditions Are Ideal

AI usually gets talked about in terms of huge valuations, futuristic promises, and worries about machines taking jobs. But some of the most interesting work is happening nowhere near Silicon Valley. Across Africa, developers and researchers are building systems under genuinely hard conditions. Conditions such as unreliable power, dozens of local languages in a single country, patchy internet, and almost no budget. In settings like these, AI tends to work alongside people rather than replace them.

In this post, Johan Botha looks at a few projects making a real dent, such as communication in healthcare setups, crop disease detection, flood warnings, applications in education, or multilingual communication. He shows us that there can be a different idea of what progress looks like, one that has to actually earn its keep.

AI in Healthcare

A patient walks into a clinic in South Africa, trying to explain symptoms in a language the nurse doesn’t speak fluently. In a country with 12 official languages, countless dialects, strong regional accents, and the presence of multilingual speakers, that situation is not unusual. Anyone who has spent enough time in multilingual African environments (trilingualism is pretty common) knows how quickly conversations switch between languages mid-sentence, especially when people are stressed, tired, scared, or trying to explain something medical.

Healthcare communication becomes complicated very quickly under those conditions. Symptoms are nuanced, patients describe pain differently depending on language and culture, and medical terminology often does not translate neatly into direct equivalents. Even where people technically share a language, comprehension can still collapse surprisingly fast.

That is part of what makes projects like AwezaMed interesting, although not for the reasons many AI headlines would probably prefer. For those who don’t know, AwezaMed is a mobile app and speech-to-speech translation tool designed to help healthcare professionals and patients communicate in South Africa. It features automatic speech recognition, text-to-speech, and machine translation, bridging language divides across the country's many official languages. While human interpreters remain the gold standard, they are an expensive solution, and trained interpreters are not available in many regions. The app does not magically solve multilingual healthcare communication, and even the developers acknowledge limitations around accents, phrasing, and conversational depth. What it does show is that AI can help reduce friction in communication when systems are designed around real-world conditions instead of ideal ones. The technology still depends heavily on humans. Nurses verify meaning, healthcare workers provide context and judgement, and the app functions more like support infrastructure than replacement. In many ways, that restraint is exactly why it feels credible.

If you spend enough time reading global AI headlines, you start noticing that the loudest stories usually revolve around valuation, scale, compute power, or increasingly dramatic predictions about the future of humanity. Meanwhile, some of the most practical AI deployments are quietly solving immediate problems in places where infrastructure is inconsistent, budgets are tight, and patience for technology theatrics is fairly limited. African AI works best when it stops pretending conditions are ideal.

That pattern appears repeatedly across some of the continent’s most inspiring ventures. The systems gaining traction are certainly not the most futuristic. They are usually more focused and built around constraints from the beginning. They assume unstable connectivity. They assume, or even expect, older devices. They assume multiple languages, uneven infrastructure, and environments where human oversight must be the de facto. Because reality in Africa is messier than what all those product demos assume.

AI in Agriculture

Apollo Agriculture in Kenya is one example of this. The company combines machine learning, satellite data, mobile technology, and field agents to support smallholder farmers who often lack access to formal credit systems, agricultural advice, crop insurance, or reliable financial records. More than 350,000 farmers across Kenya and Zambia now use the platform. Yes, AI is there, but it was designed around the above assumption that connectivity may fail, documentation may be incomplete, and, in this case, farming conditions may change faster than the data can keep up.

The company still relies heavily on field agents who verify information, assist users, and bridge the inevitable gap between recommendations from the algorithm and “messy” real-world farming. Human involvement was never treated as evidence that the automation had somehow failed. It was treated as part of the operating reality, which is a far more mature way of thinking about AI than most of the louder discussions we’re all bombarded with daily.

A similar idea can be found in PlantVillage’s Nuru platform, which helps farmers identify crop diseases through image recognition on mobile phones. The machine learning system embedded in the platform is getting better with the help of AI. One practical engineering decision was ensuring the system could operate offline in rural areas with poor connectivity. Farmers photograph crops, the AI assists with diagnosis, and the user still decides what action to take next. According to CGIAR-backed reporting and field studies, the platform has shown strong results in helping farmers identify cassava diseases and improve yields under difficult agricultural conditions. In most situations, though, the system still depends on decent image quality, local training, data, and, importantly, human insight.

What makes these solutions noteworthy is that human involvement is a given. They seem to understand where automation stops being useful. A surprising number (surprising to people outside the continent, that is) of successful AI initiatives in Africa follow this hybrid model. AI handles pattern recognition, forecasting, scoring, or analysis while humans continue handling context, relationships, improvisation, trust, and all the variables that are absent from most datasets. While all this might not sound as flashy as other AI solutions out there, these systems have a more profound impact on the reality of people using them.

AI for Disaster Prevention

With the recent flooding in my hometown still front and centre for me, flood prediction systems present another use case on how AI can make a difference. Google’s Flood Hub now provides forecasting systems and alerts across large parts of Africa where traditional flood monitoring infrastructure has historically been non-existent or, at least, inconsistent. The platform combines machine learning with remote sensing and forecasting models to provide earlier warnings through mobile systems and APIs. The engineering challenge changes significantly when the infrastructure itself is incomplete. Forecasting floods in regions with fragmented datasets, limited sensors, unreliable connectivity, and rapidly changing environmental conditions is very different from modelling environments filled to the brim with expensive monitoring systems and lots of clean data.

AI in Education

Another surprising and inspiring use case of how AI is being leveraged in Africa came from research involving teachers in Sierra Leone. Researchers studied how teachers used an AI chatbot through WhatsApp, instead of relying on a usual web search, to support them with their teaching activities. The findings were striking. Researchers found that individual web pages consumed 3,107 times more bandwidth than AI responses on average, while querying AI through WhatsApp was approximately 98% less expensive than loading traditional web pages, even after including the AI compute costs. Teachers also rated AI responses as more relevant and more helpful than standard web search results.

Priorities shift dramatically, and systems become lighter because they don’t have another option. Smaller, specialised models are more useful and easier to maintain. Not to mention cheaper. Things like having an offline capability matter to us. Language support, reliability, a mobile-first(!) design. In many cases, usefulness starts outperforming novelty surprisingly quickly.

AI in Translation

This also helps explain why multilingual AI work on the continent deserves far more attention than it usually receives. Projects like Lelapa AI and Masakhane are trying to address one of the largest blind spots in global AI development, namely language exclusion. Most major AI systems were trained primarily on dominant global languages, leaving large parts of Africa digitally underrepresented or ignored altogether.

The challenge here is not simply translation. African language environments are complex, and building scalable solutions that function under those conditions is technically difficult, risky, and not particularly glamorous. All of which makes funding a real challenge. But it may ultimately prove far more useful than yet another chatbot optimised for polished English marketing copy.

Lessons Learned of AI in Africa

None of these successes means Africa has somehow solved AI or escaped the problems surrounding it. There are already examples of poorly designed and rolled out systems, with imported assumptions and harmful outcomes. Facial recognition technologies have faced criticism for racial bias and inaccurate identification. Data workers across parts of Africa have reported exploitative conditions while performing emotionally taxing moderation and annotation tasks for global AI companies. Some imported models fail simply because they do not understand the African reality.

Another “uncomfortable” truth is that AI alone does not miraculously fix broken systems. You cannot automate and scale what you have not yet organised properly in the first place. That principle applies to people as well. A growing number of AI researchers and practitioners have started pointing out that meaningful AI adoption rarely comes from generic corporate training alone. It usually comes from people experimenting and adapting systems to local realities. You try to reduce friction around a specific problem. In the above discussion, those are crop disease, flood warnings, credit access, or information gaps.

In many parts of Africa, AI needs to justify itself quickly, operate under pressure, and, most importantly, prove useful outside conference stages and investor presentations. It gets rewarded when it still works after the power flickers, the bandwidth drops, the phone is five years old, and the user speaks a language the rest of the digital world forgot to support.

Sources:

 
 
bottom of page