“AI is revolutionizing everything”. This sentiment dominates a standard headline and forms the crux of every pitch deck. While it is indeed a digital revolution unfolding globally, the narrative that it is revolutionizing everything equally is a far-fetched one, especially within the African region.

Although AI is being deployed in vital sectors such as agriculture and public health, progress continues to stall. In these areas, which are vital to every state and the continent, there exists a common obstacle to the success of AI. With records scattered across disparate systems and huge infrastructure costs, there is no digital foundation for an AI to run on. Conversely, in the mobile sector, AI continues to transform with every passing second, contributing about $220 billion to the continent’s economy, an indication of real digital momentum.

With the drive to produce technology, it is important to consider whether funding, such as venture capital allocated to superficial AI wrappers, is making the continent a passive consumer. And thus, there is a need to know where AI actually works and doesn’t.

First, it is essential to establish the landscape in question: Africa. This is the environment and its components that determine what survives. A region currently limited to a few local data centres, yet one that lacks the historical datasets to train these Machine learning models. This is particularly prevalent in sectors that largely depend on paper ledgers. In data-poor environments, the machine has to work twice as hard to extract meaningful patterns. More GPU hours are then required to compensate for being trained with uncleaned data, the cost paid for the absence of digital records. As fast-paced as this revolution is, the existing facilities are no longer enough for the estimated 7 million GPU hours needed to continuously train these models. Productivity in such an environment slows down, and here a problem stems: the ‘Compute Paradox.’ This paradox ensures that returns on the large investments into computing technology are likely below par due to this fundamental slowdown.

It is, therefore, unsurprising that the advanced compute resources required to build and scale models are available to only about 5% of local AI innovators. What then happens to the others? The majority of local startups must rely on foreign cloud APIs like AWS, and since a free lunch exists nowhere, they suffer an ‘infrastructure tax’. This is a tax because these are rates a typical Lagos-based start-up with almost zero revenue is unlikely to afford. Adding salt to the wound, local currency devaluations against the USD further cripple their capacity. This clearly paints the picture of the space that determines where AI truly functions.

Predictive AIs feed on historical datasets to make meaningful patterns. This means that they perform exceptionally well where data is organised digitally rather than scattered. Consequently, the application of AI thrives in data-rich hubs, most notably, the African telecommunication sector. With about 47% of the continent’s population subscribed to mobile services, representing roughly 710million unique subscribers is access to massive volumes of structured mobile data. Using payment histories, mobile and internet subscription details and behavioural data like app usage, AI can predict the ‘churn rate’, which identifies subscribers likely to cancel or switch providers. Operators also leverage them for their backend predictive maintenance and intelligent traffic management by mapping real-time human and vehicular mobility.

In the African fintech market, AI continues to transform financial services and promote accessibility. To drive inclusion, AI is being leveraged in credit-worthiness assessment. Alternative data-based credit scoring is employed instead of the adoption of western standards. Non-traditional data such as mobile usage patterns, utility payment history, e-commerce purchases, and even device usage now help assess the credit-worthiness of the unbanked, as these reflect a person’s financial behaviour. While mobile money platforms address the gap of financial exclusion, the threat of fraud remains a hurdle. With AI, these fintech platforms are able to detect fraud and prevent it. These Algorithms identify anomalies in transactions and signal for fraud based on past transactions, kept clean in a standard system. Again, AI systems succeed here because they have access to their favourite meal: Clean data.

On the Contrary, it is difficult for AI to progress when deployed in areas where data is scarce and foundational Digital Public Infrastructure (DPI) is non-existent. This is why areas like agriculture continue to struggle. While startups bring forth intelligent ideas, implementation remains static because the ‘digital soil’ has not been prepared.. Smallholder farmers lacking ground-level sensor networks for soil moisture monitoring and irrigation control often have to rely on remote satellite data. A machine cannot identify a farmer or the coordinates of his plot without his digital IDs linked to land registries and without stable electricity and internet connection, these insights are delivered through sms text messages.

For a critical sector like Healthcare, deploying AI could make delivery more manageable, but the absence of interoperable, digitsed health records renders this impossible. The smartest machine would have nothing to ingest when patient records are fragmented in physical folders or a siloed software. Consequently, healthcare AI often ignores these poor datasets and rather relies on narrow use cases such as computer vision for localised medical image screening. Even though they can detect a fractured bone, they cannot predict a disease outbreak or manage care across a population. Having effective AI in healthcare means enabling workers to attend to more patients through faster diagnosis and standardised care requiring a digital backbone for the seamless flow of data from a rural clinic to a national database.

Amidst these deployment challenges lies the phenomenon of ‘AI washing’. Some startups pose as AI- driven firms but actually rely on simple rule-based software that uses predefined “if - else” statements to make predictions. By using systems that operate on current data and are hard-coded by humans, they can make consistent and accurate predictions. For others, the training process relies entirely on underpaid gig workers, the ‘mechanical turks’ posing as algorithms to correct errors and shape responses. This AI hype creates a mask for some startups to market automated efficiency while depending on low-wage labour. The use of these human-powered ‘AI’ undermines digital sovereignty since it hinders the development of local intellectual property creating a political economy where the cost of computing is offloaded onto the human algorithm.

AI washing is often a response to the prohibitive cost of computing and APIs. Developers pay for cloud services, for storage, networking, and even training the AI. The majority of these API’s, like AWS, Azure or OpenAI, can exceed the price of the local developer’s laptop. This encourages falling back on simple rule-based systems to avoid being blocked by these financial barriers, hiding behind the AI hype.

In Ghana, mobile money provides telcos and banks with large datasets required to build functional AI-driven chatbots for service provision. Telcos now deploy these bots to manage customer complaints and inquiries as a human customer representative would. Banks, meanwhile, use them to foster self-service, allowing customers to perform transactions, transfers and bill payments. This structured data extends even to the Ghana Statistical Service, which leverages de-identified de-identified telecommunications data to generate predictive models for official economic datasets. It indicates that institutions can shift from traditional, slowing moving survey methods when they gain access to clean, structured digital ledgers. Real time algorithmic analysis can generate economic statistics faster and accurately.

In the capital, Accra, AI specifically navigates the mapping of human and vehicle mobility. Algorithms on ride-hailing apps like Bolt or Yango predict traffic patterns and peak hours, adjusting prices based on congestion at Spanner junction between 4 pm and 7 pm. These models are so well-fed that they recognise which local drivers to favour during midday or late nights: those with longer working hours.

Ironically, the thousands of market traders at the centre of Makola who process tons of business transactions ranging from payments and supplies to the movement of goods benefit little to nothing from AI deployment. They generate immense datasets daily, yet this information remains ‘littered’ in memories, on shredded cartons and whatsApp messages. These records are too scattered for an AI to digest. However, a high-tech banking app that processes a fraction of Makola’s volume is able to leverage AI to improve its services because it maintains structured ledgers and clean historical datasets.

The bottom line remains that AI only works where digital and traceable historical datasets exist. Where data is poor, it does not compromise to fit, hence the stall in its deployment, even in the most crucial sectors. This is almost a non-negotiable reality.

Key Takeaways

• AI in Africa is not ubiquitous. It only works where robust capital investment and massive, structured digital ledgers meet.
• Local innovators will continue to be forced to pay “infrastructure tax” without clean, digital records.
• To foster production of its own Technology, stakeholders must prioritise building foundational digital public infrastructure and affordable local compute facilities to prevent the flow of venture capital toward Superficial AI wrappers.