AI for African founders, without the hype

The first thing I tell African founders about AI is that almost none of the advice they are reading was written for them. The case studies assume cheap compute, abundant capital, and a workforce that thinks in English-only contexts. None of that maps cleanly onto a business in Tembisa, or a startup in Lagos, or a SME in Nairobi. So the playbook needs translating. Here is the version I share with the founders I actually work with.

## Start with the bottleneck, not the model

Every founder I meet wants to talk about which model to use; GPT-5, Claude, Gemini, Mistral. Almost none of them have first answered the question that actually matters: where is the bottleneck in the business right now. AI is leverage. Leverage applied to the wrong thing is just noise.

I run a simple exercise with founders. Walk me through your week. Tell me where you lose the most hours, or the most revenue, to a task that is mostly pattern-matching. That is your candidate. It is almost never "let''s build a chatbot." It is almost always something less glamorous; quote turnaround, ticket triage, lease analysis, supplier reconciliation. Boring, valuable work.

## Buy before you build

In 2026, the right default for African founders is to buy. The model providers have made the underlying technology a commodity. The differentiator is your data and your distribution, not the model. If a $30/month tool solves 70 percent of your problem today, take the 70 percent now and put the saved engineering hours into the thing only you can do.

I have seen too many small teams spend six months building a custom RAG system that ChatGPT''s built-in retrieval would have handled for the first year. By the time they shipped, the market had moved twice. Buy the boring infrastructure. Build the proprietary slice.

## Treat data like inventory

If you are running a business in Africa, your data is more uneven than the US case studies assume. You have phone-based receipts, WhatsApp orders, Excel files written by a bookkeeper who left in 2019, and an ERP your operations lead never trusted. That is normal. Do not let a consultant tell you that you need to "get your data house in order" before you can use AI. You can use AI on messy data; the trick is to know what good enough looks like for your specific use case.

What I tell founders is this: treat data like inventory. You don''t need to inventory the warehouse to start selling. You need to know where the next ten orders are coming from, and you need to be able to find the SKU on the shelf when you need it. Same with data. Pick the use case, identify the few datasets that matter for it, clean those, ship.

## Build for the language you actually speak in

The single most underused move I see is fine-tuning a model on the actual language your customers speak. Not a research-grade fine-tune; a small, practical adaptation. South African retail customers speak a mix of English, isiZulu, and Sesotho in the same WhatsApp message. Models trained mostly on US English handle that badly. A small layer of prompts and examples in your customers'' actual register can move conversion rates by ten to twenty percent.

This is one of the most concrete arguments I make for African builders to stay in the AI conversation. Tools tuned in California will keep being tuned for California. The tools your business needs will be tuned by you, or by builders who understand your market.

## Hire one person, not a team

Almost every African SME I have advised has been told by a consultancy that they need an "AI team." For most of them, the right answer is one technically literate operator who can buy, configure, and integrate three or four tools. Title doesn''t matter. Disposition does. They should be curious, sceptical of hype, and comfortable shipping rough first drafts.

That person, paid well and given real autonomy, will outperform a six-person team running a slide deck for the next eighteen months.

## What I would do this quarter

If you are running a business and you have not yet shipped one AI-driven workflow, here is the ninety-day version of what I would do.

Pick one workflow that costs you the most hours per week. Buy a tool that does 70 percent of it. Spend two weeks integrating, not building. Spend the next four weeks measuring. If the numbers move, keep it. If they don''t, kill it and pick the next one.

That is the entire playbook. The rest is patience and attention.