What AI Can Actually Do for an African Business Today

Walk into any business breakfast in Nairobi, Lagos or Johannesburg right now and you will hear two contradictory sermons about AI. One says it will replace your staff by Friday. The other says it is all hype and you should ignore it. Both are wrong, and both are expensive to believe. The honest truth sits in a narrow, unglamorous middle: AI can already do specific, valuable things for an African business today, and it cannot do many of the things you have been sold.

My thesis is simple. AI is a capability multiplier, not a decision-maker. It is brilliant at chewing through repetitive language and pattern tasks at a scale no human team can match, and it is unreliable at anything that requires accountability, local judgement, or clean data you do not actually have. The businesses winning with it are the ones who understood that distinction before they signed a contract.

What AI can genuinely do for you today

Let us be concrete, because generic promises are how money gets wasted. Here are things that work right now, at prices an SME can survive.

1. Handle the first layer of customer conversation

Across Ghana, Nigeria, Kenya and South Africa, WhatsApp is the default way customers talk to a business. More than 175 million people message a WhatsApp Business account every single day, and WhatsApp messages open at rates above 90 percent compared with 20 to 25 percent for email. An AI assistant sitting on your WhatsApp line can answer "are you open," "how much," "where are you," and "has my order shipped" at 2am without a salary. Nigerian logistics firms like GIG Logistics and Kwik already push automated delivery updates this way, at roughly a cent and a half per utility message. That is real, and it is affordable.

2. Draft, summarise and translate faster than any person

Product descriptions, invoice reminders, social captions, a first draft of a proposal, a summary of a 40-page tender document - these are language chores, and language is exactly what large language models do well. A one-person marketing team can now produce a week of content in a morning. That is not a gimmick; it is genuine leverage for the thinly staffed businesses that make up most of the African economy.

3. Spot patterns in data you already collect

If you run an M-Pesa till or a POS, you are sitting on transaction history. AI is good at ranking your best customers, flagging unusual activity, and predicting which stock will move. Safaricom itself moved M-Pesa onto a cloud-native, AI-powered platform in 2025 and, with AWS, deployed graph neural networks that hit 89 percent accuracy at catching social-engineering fraud. You will not build that, but the same category of pattern-detection is now available to smaller businesses through off-the-shelf tools.

4. Automate the boring back office

The most reliable returns from AI are unglamorous. MIT's 2025 research on business AI found that back-office automation - reconciling records, sorting emails, extracting data from receipts - produces the highest returns precisely because the work is repetitive and low-stakes. Start where a mistake is cheap.

What AI cannot do (no matter what the demo showed)

Now the other half, which vendors will not volunteer.

It cannot be accountable

AI does not carry risk. If a chatbot quotes a wrong price, promises a refund you never authorised, or gives a customer bad medical or legal information, that liability is yours. It cannot sign off a loan, fire a supplier, or make an ethical call. Anything with legal or financial consequence needs a human name attached to it.

It cannot fix bad or missing data

This is the one that quietly kills African AI projects. If your inventory lives in a WhatsApp chat and three exercise books, AI has nothing clean to learn from. Gartner projects that through 2026 organisations will abandon 60 percent of AI projects that are not backed by AI-ready data. Your data is the raw material. No model overcomes its absence. This is exactly why I keep telling founders to not get pressured into buying AI before their business is ready.

It does not truly understand your local context

Most mainstream models were trained on English-heavy, Western-heavy data. They can stumble on Sheng, Pidgin, code-switching, local place names, and cultural nuance. African-language efforts like Lelapa AI's InkubaLM, covering Swahili, Hausa, Yoruba, isiZulu and isiXhosa, and the Masakhane research community are closing that gap - but slowly. Do not assume the model "gets" your customer the way you do.

It hallucinates with total confidence

An AI will invent a policy, a statistic or a product feature and state it in a calm, authoritative tone. In customer-facing settings this is dangerous. Treat every factual claim it makes about your business as a draft to verify, not a fact to publish.

The question is never "is AI powerful." It is "is this specific task one where a confident, fast, occasionally-wrong language machine adds more than it risks."

A simple test before you adopt anything

Run every proposed AI use case through three questions:

1. Is the task repetitive and language- or pattern-based? If yes, AI is a candidate. If it needs judgement, human first.
2. Is a mistake cheap and reversible? Sorting emails, cheap. Approving credit, not cheap. Start where errors do not hurt.
3. Do I already have clean data or a clear process for this? If no, fix that before you automate it. Automating a mess just produces a faster mess.

The realistic picture for African SMEs

Context matters. Only South Africa had recorded AI adoption above 20 percent among businesses by the end of 2025, even as consumer usage exploded - Kenya reached a 42 percent ChatGPT usage rate among internet users. That gap between individual use and business use is the opportunity. McKinsey estimates generative AI could add 61 to 103 billion dollars of annual value to Africa. Very little of that goes to the businesses that bought the flashiest tool. It goes to the ones who used a boring tool on a well-chosen problem.

So here is my advice. Pick one task where you are drowning - customer replies, content, reconciliation - and pilot AI there for 90 days with a number you can measure. Keep a human on anything with consequence. Fix your data as you go. Ignore the person telling you AI will run your whole company, and equally ignore the person telling you it is a fad. The reality is more useful than either: a cheap, fast, tireless assistant that is right most of the time and needs supervision the rest. Used with that clear head, it is one of the best deals a resource-constrained business has ever been offered.