African enterprises don't need ChatGPT; they need domain models.
There is a category error at the centre of most African enterprise AI strategies, and it is costing them real money and competitive position. The error is this: they are trying to use general-purpose AI tools to solve specific operational problems; and then concluding, when it does not work well, that AI is not ready for Africa. AI is ready. General-purpose AI is just the wrong instrument.
What general-purpose AI is actually good at
ChatGPT, Claude, Gemini, and their peers are extraordinary tools. They can write code, summarise documents, draft communications, answer general knowledge questions, and assist with a broad class of cognitive tasks across virtually any domain. For individual productivity; the knowledge worker who needs a thinking partner, a writing assistant, or a research aid; they are genuinely transformative.
They are trained on the internet. The internet, as it turns out, is not a representative sample of Nigerian enterprise operations, Ghanaian trade finance, Senegalese agricultural supply chains, or Kenyan health insurance claims. The training data for every major general-purpose model is overwhelmingly Western, overwhelmingly English, and overwhelmingly consumer-facing. The models know a great deal about what a lot of people have written on the public internet. They know almost nothing about how your business actually works.
The specificity gap
Consider what a Nigerian Tier-2 bank actually needs from AI. It needs models that understand the credit risk profile of a small business owner in Kano with three years of mobile money history and no formal credit record. It needs models that can process loan applications written partly in Hausa, partly in English, and partly in the shorthand of a 12-year bank officer. It needs models calibrated to the fraud patterns specific to Nigerian mobile banking, not the fraud patterns from US credit card data.
None of this knowledge lives in ChatGPT. It lives in the bank's own transaction history, its own loan books, its own fraud incident logs. The model that will outperform a general-purpose AI on these tasks is a model trained on that data; and that is a domain model, not a general-purpose one.
The model that will outperform ChatGPT in your business is already inside your data warehouse. You just haven't trained it yet.
This pattern holds across every sector I can name. A Lagos logistics company's demand forecasting model needs to understand the relationship between school holidays, NEPA outages, and delivery success rates in specific LGAs. A Nigerian HMO's claims processing model needs to understand the difference between a legitimate claim for malaria treatment and the specific fraud patterns that are unique to the Nigerian health insurance ecosystem. A West African agricultural trading platform's price prediction model needs to know about seasonal road access in the North-West states.
ChatGPT does not know any of this. Your data does.
The language problem is worse than you think
Nigeria has over 500 languages. Yoruba, Hausa, Igbo, and Pidgin are spoken by tens of millions of people and are central to how Nigerian enterprises actually communicate; with customers, with suppliers, and internally. General-purpose AI models perform well in English. Their performance in Nigerian Pidgin is inconsistent. Their performance in Hausa and Yoruba is significantly below their English performance. In Igbo and minority languages, the performance is often unreliable for production use.
This is not a temporary gap that will close in the next ChatGPT release. It is a structural consequence of training data distribution. The organisations; and there are now several across West Africa; that are building domain models with local language capability are building a competitive moat that general-purpose AI companies will not close for years, if ever.
The data sovereignty argument compounds this
There is a second argument for domain models that is separate from performance: control. When you use a general-purpose AI API, you are sending your operational data; your customer queries, your document contents, your decision inputs; to a server owned by a US or European company, processed under their privacy policy, potentially used to improve their models, and stored in their infrastructure.
A domain model that you own, fine-tuned on your data, running on infrastructure you control, creates none of these exposures. Your operational intelligence stays inside your organisation. Your data does not subsidise your competitors. And you are not at the mercy of a vendor's pricing decisions, API rate limits, or terms of service changes.
Under NDPA 2023, the data sovereignty argument also has a compliance dimension. General-purpose AI APIs that transmit personal data outside Nigeria require cross-border transfer mechanisms that many organisations have not established. A domain model running in-country eliminates this compliance exposure entirely.
But domain models are expensive to build; right?
This objection was correct in 2022. It is less correct in 2026, and becoming less correct every quarter.
The economics of fine-tuning a capable open-source foundation model on domain-specific data have improved dramatically. A credit-scoring model fine-tuned on a Nigerian bank's loan book does not require the compute infrastructure of OpenAI. It requires good data, a competent ML engineering team, and a thoughtful evaluation framework; investments that are within reach of any Tier-1 or Tier-2 Nigerian financial institution.
The key insight is the distinction between foundation models and domain models. You do not build a foundation model. Building a foundation model requires billions of dollars and access to internet-scale training data. You fine-tune a foundation model; taking a capable open-source base (Mistral, LLaMA, Qwen, and their successors) and adapting it to your domain, your language, your data, and your task. This is a fundamentally different economic proposition.
You don't build a foundation model. You build on one. The distinction is the difference between Boeing and a private pilot.
The right architecture for 2026
None of this means that general-purpose AI has no place in African enterprise operations. It has a clear place: internal productivity, content generation, research assistance, and the class of tasks where African-specific context is not required and general language capability is sufficient.
But for the decisions that drive enterprise value; credit risk, fraud detection, supply chain optimisation, customer retention, clinical decision support; the architecture should be a domain model tuned on your data, evaluated against your specific failure modes, and owned by your organisation.
The organisations that are building this now are accumulating a data and model advantage that compounds. Every transaction processed, every claim adjudicated, every delivery completed makes the model better. Competitors using general-purpose AI from the same API endpoint are not accumulating this advantage. They are renting a commodity.
The question is not whether to use AI. The question is whether the AI you use belongs to you or to someone else. In 2026, that distinction will determine your competitive position in 2028.