Why English-First AI Tools Are Failing Township Entrepreneurs

South Africa has 11 official written languages. Yet research from the University of Cape Town's MzansiLM project found that only 8.7% of South Africans speak English at home. This is not a minor statistical footnote — it means the vast majority of the population, across every sector including small business, operates day to day in a language other than the one most major AI tools were originally built and optimised for.

For a township entrepreneur serving customers, negotiating with suppliers, and thinking through business decisions primarily in isiZulu, isiXhosa, Sesotho, Afrikaans, or a natural mix of local language and English, an AI tool that only performs well in English isn't just less convenient — it's structurally less useful for the actual work of running the business.

This isn't just a theoretical concern. A 2025 study presented at the AAAI conference specifically benchmarking AI performance across low-resource African languages found a substantial, measurable gap: even the best-performing AI model tested showed an absolute performance drop of 12.0 to 19.9 percentage points when comparing English performance to the average across 11 African languages.

The same research found meaningful differences between African languages themselves — Afrikaans, which has more digital text available for AI training, performed substantially better than lower-resource languages with less available training data. This matters practically: an entrepreneur working primarily in isiZulu or Sesotho is likely facing a larger accuracy and quality gap than one working in Afrikaans, simply due to how much language data currently exists to train AI systems on each language.

In practical terms, this gap shows up as AI-generated content in local languages that reads slightly unnatural, occasionally mistranslated, or grammatically imperfect in ways that a native speaker would immediately notice — even when the same tool produces fluent, natural English. For marketing content specifically, where tone and authenticity matter enormously for connecting with a local customer base, this gap can undermine exactly the kind of trust-based, community-rooted marketing that works best in township contexts.

The good news is that this problem is actively being worked on, even if the solutions aren't yet fully packaged for everyday business use. The University of Cape Town's MzansiLM project is the first public AI model specifically aimed at all 11 of South Africa's official written languages — a foundational step toward more accurate, natural South African language AI.

Lelapa AI, a South African company, has built VulaVula, a translation tool for African languages, and has released InkubaLM, a multilingual African-language AI model that specifically includes isiXhosa and isiZulu among its supported languages. Additionally, organisations like PanSALB (the Pan South African Language Board) have worked on AI-term translation specifically in isiXhosa, Afrikaans and isiZulu, while SADiLaR continues building broader digital language-resource infrastructure for South African languages.

It's important to be precise about where things currently stand. The research and foundational technology for better South African language AI genuinely exists and is actively improving. What doesn't yet exist, in any widespread way, is this technology packaged into low-friction, business-ready products specifically for township entrepreneurs — tools that let a spaza shop owner or salon operator draft marketing content, customer messages, or compliance documents naturally in their working language without needing to understand the underlying AI infrastructure at all.

This is a genuine, current gap — not a solved problem waiting to be discovered, but an active area where the foundational work is ahead of the practical, accessible product layer most entrepreneurs would actually use.

In the meantime, there are practical steps that help. Using AI tools to draft content in English first, then having that translated and adapted into the specific local language a business's customers actually speak — rather than expecting native-quality output directly in the local language — currently produces more reliable results. Reviewing AI-generated local-language content with a native speaker before using it publicly remains a sensible practice, given the documented performance gap.

kasiAIhub is built with an explicit awareness of this gap. Sessions incorporate translation and local-language adaptation as a deliberate step in content creation — not an afterthought — recognising that marketing content, customer messaging, and even compliance explanations often land better, and build more trust, when delivered in the language a business's actual customers use every day.

See how content gets built across Think & Plan, Build & Create and Brand & Grow. This piece sits inside the broader State of AI Adoption Among Township Entrepreneurs (2026) hub.

See the full 3-session programme →

University of Cape Town MzansiLM research (2026); AAAI 2025 research on low-resource African language AI performance; Lelapa AI (VulaVula, InkubaLM); PanSALB; SADiLaR.