Is it too late to build an AI startup in Africa?

By Bonaventure Ogeto, Founder and CEO, Mctaba Labs
Published

The short answer

No. It is too late to compete on building AI models, and nearly too late for generic chat tools that the big providers already give away. It is early for products that apply existing models to specific local work, such as paperwork, customer messages and records, in businesses those providers will never visit. Start from the work, not the model.

Every few weeks a company in California or Beijing announces a model that does something the last one could not. From a desk in Nairobi or Kampala it is easy to conclude that the train has left. They have the money, the chips and the researchers. What is there for you to build?

Quite a lot, as it happens. But the question "is it too late?" has different answers depending on which kind of AI business you mean. For one kind it was always too late. For another, the window is closing. For a third, almost nobody has begun.

Too late for what, exactly?

There are three layers to this industry, and they behave very differently.

Layer What it is Who wins Verdict for a new founder here
Models Building the AI itself, such as the systems behind ChatGPT, Claude and Gemini A handful of very well funded companies Closed. Do not try
General tools Assistants that chat, write, summarise or translate for anyone The model providers, who include these for free or cheaply Closing. Avoid
Applied products A tool that does one specific job for one specific kind of customer, using someone else's model Whoever understands the customer and the work best Open, and early

The third layer is where ordinary founders have always built. Nobody starting a booking product had to invent the database or the mobile network first. They used what existed and applied it to a problem they understood. AI models are now one more thing that exists. You rent one by the task, the way you buy data bundles, and build the product around it.

So it is not too late, provided you pick a specific job for a specific kind of customer.

Where is the opening for a founder in Africa?

In the distance between what a general assistant can do and what a business here needs done.

A model can read a document and answer questions about it. It does not know that a clearing agent in Mombasa receives the documents as photographs on WhatsApp, that half are handwritten, or which three fields matter. Someone has to build that. The same is true across hundreds of small, specific jobs.

Your advantages are practical ones:

  • You know the workflow. You have seen how the school bursar, the sacco clerk or the clinic receptionist does the job, including the parts that never get written down.
  • You can get the meeting. Trust here travels through people. A recommendation from someone the owner knows beats any advert.
  • Your customers live on WhatsApp. A product that works inside a chat the customer already uses needs no app download and no training.
  • Payments are local. Customers pay by M-Pesa. A foreign tool that bills a card in dollars loses many of them at the first step.
  • Language is mixed. Real messages move between English, Swahili and Sheng in one sentence. Models tend to be strongest in English, so test how well they handle the way your customers write, using real examples. Where they struggle, the work you do to fix it is an advantage others lack.
  • Records are on paper. A great deal of business information here sits in exercise books, receipts and photographs. Turning that into something usable is real work that general tools do not do.

None of this is about being cheaper than global companies. It is about being close to a customer they will never meet.

Which AI ideas should I avoid?

The ones where your whole product is a thin layer on a model, with nothing else behind it.

  • "ChatGPT, but for Kenya." A general assistant with a local name. The original already works here.
  • A generic chatbot for any business. If it could be sold unchanged to a law firm and a butchery, it is not specific enough to defend.
  • Content generators. Captions, essays, logos, CVs. The free assistants do these well, and customers know it.
  • Anything that is one instruction away. If a customer could get the same result by typing a sentence into a free assistant, they eventually will.
  • A feature the provider is likely to add. Summaries, meeting notes, translation and document chat keep being absorbed into the main products.

A fair test: if the model providers released a much better model tomorrow, would your product become more valuable or pointless? Applied products get better when models improve. Thin wrappers get replaced.

How do I test whether my AI idea is worth building?

Put it through six questions. Be strict with the answers.

  • Would the customer pay for the result if a person did it? If nobody would pay a human to do this task, AI does not make it worth paying for.
  • Does it take more than a free assistant to do it? Your product must add something: the customer's own records, a connection to their systems, checks, or delivery in the place they work.
  • Do you have access others do not? A route to the customers, an understanding of the trade, or data that is not on the public internet.
  • What happens when the AI is wrong? It will be, sometimes. A wrong caption is an annoyance. A wrong dosage, loan decision or legal document is a disaster. Either choose tasks where errors are cheap, or put a person in the loop to approve.
  • Does the cost per task leave a margin? See the next section.
  • Does it fit into what they already use? WhatsApp, SMS, the till, the spreadsheet. A new habit is a harder sale than a new tool.

Then run the test that costs almost nothing. Before you build, do the job by hand for five customers using an ordinary AI assistant. Say your idea is to answer customer enquiries for furniture workshops on Ngong Road. Ask five owners to forward you a week of enquiries. Draft the replies yourself with the assistant, using each workshop's price list, and send them back for the owner to approve. Charge a small fee.

In two weeks you will know whether the replies are good enough, what the model gets wrong, how much correcting is needed, and whether owners value it enough to pay. You will also have real examples to test any future product against. That is worth more than a prototype.

What does it cost to run, and can customers afford it?

Unlike ordinary software, an AI product costs you money every time it is used. Providers generally charge by the amount of text the model reads and writes. Prices differ by provider and model and change often, so read the provider's current pricing page instead of trusting any figure in an article.

The method matters more than the number. Here it is with made-up figures.

Say each customer enquiry your product answers costs you KES 2 in model fees. A workshop receives 400 enquiries a month. Your model cost for that customer is KES 800 a month. Add messaging fees, hosting and your support time, say another KES 700. Your cost to serve is about KES 1,500.

If you charge KES 2,000 a month, you keep KES 500 per customer, which is too thin. At KES 5,000 the business works, provided the owner believes the product recovers more than that in sales. So the real question is on the customer's side: do unanswered enquiries cost this workshop more than KES 5,000 a month? If one lost sofa order is worth KES 40,000, the answer is plainly yes.

Three ways to keep the cost down: use a smaller, cheaper model for simple tasks and a stronger one only when needed, send the model only the information it needs, and do not call the model at all for things ordinary code can do, such as looking up a price.

What should I build so that the next model release does not replace me?

Build the parts that are not the model.

  • The workflow. The steps before and after the AI: receiving the photograph, checking the result, sending it to the right person, recording that it was done.
  • The customer's records. Their price lists, past orders, client history. The longer they use you, the more the product knows that no general assistant does.
  • The connections. M-Pesa payments, WhatsApp or SMS delivery, the accounting or stock system they already have.
  • The checks. Tests built from real examples that tell you whether answers are right, and a way for a person to approve anything risky.
  • The trust. Clear handling of customer data. If you will store people's personal information, read the guidance published by Kenya's Office of the Data Protection Commissioner, or the regulator where you operate, before you collect anything.

Keep the model itself replaceable. Have your developers build so that switching provider is a small change. Then each new release becomes an upgrade for you instead of a threat.

The founders who do well with AI here will look less like AI companies and more like people who understand clearing and forwarding, or school administration, or dairy cooperatives, and who happen to use AI to do the job better. If you know a trade and its customers, you are not late. Pick one task in it and do it by hand for five customers this month.

Key takeaways

  • The race to build models is closed to almost everyone. The race to apply them to specific local work has barely begun.
  • Your advantage is access to a customer and a workflow that a global company will never study.
  • An AI product is still a product. Someone must have a costly problem and be willing to pay.
  • Do the task by hand with an AI assistant for five customers before you build anything.
  • Work out the model cost per task and check that your price leaves a margin.
  • Build the parts a model release cannot replace, which are the workflow, the records, the payments and the trust.

Frequently asked questions

Do I need to train my own AI model?
Almost certainly not. Training a general model takes money and specialist skill far beyond a new startup. Nearly every useful AI product is built by sending instructions and data to an existing model through its provider's service, and then doing careful work around it. Put your effort into that surrounding work.
Can a non-technical founder build an AI product?
Yes, if you bring the knowledge of the customer and the task. You can test the idea yourself with an ordinary AI assistant and real examples. You will then need a technical partner or team to make it reliable, secure and connected to payments and messaging, because the gap between a demo and a product people depend on is large.
Will the big AI companies simply build my idea into their own products?
They will if your idea is general, such as summarising, drafting or answering questions. They are unlikely to if it depends on a specific trade in a specific country, local payment and messaging systems, and relationships with customers. The narrower and more local the workflow, the safer you are.
Is AI too expensive for customers in Africa to pay for?
Not when the product replaces something that already costs them money. A business that pays a clerk to do a task by hand, or loses sales because nobody answers messages at night, has a budget for the result. Price against that cost, and check that your own cost per task leaves room.

Sources

  1. The Lean Startup: How Today's Entrepreneurs Use Continuous Innovation to Create Radically Successful Businesses, Eric Ries (book, 2011). Checked on 19 November 2026.