Africa is Redefining AI Usability: A Call for Multilingual, Multimodal AI.

While product teams in San Francisco picture AI as a sleek chat window, crisp English prompts, instant responses, a farmer in rural Malawi is quietly reinventing what an interface can be. Here, the "user" doesn't type. She records a WhatsApp voice note.[1] She doesn't describe a problem; she snaps a photo of a wilting maize leaf. For millions across Africa, AI is not a destination website or a novel toy. It is an urgent, voice-native query in Chichewa or Yoruba, sent while walking through the fields.

How People Actually Use AI

Across the continent, AI enters daily life through the channels people already trust. WhatsApp dominates because it works on feature phones, survives spotty data, and lets users speak rather than type.[2] But the true shift is in the intelligence running behind the chat. When a trader in Lagos sends a voice note, a Large Language Model (LLM) isn't just transcribing words; it's performing real-time intent recognition across three languages. The AI must follow this fluid "code-switching" or fail.

Real Systems Meet Real Constraints

UlangiziAI: Voice and Image AI for Malawi Farmers

Consider UlangiziAI, deployed in Malawi by Opportunity International. Farmers send questions via WhatsApp: voice notes about pests, photos of diseased cassava, text in Chichewa or English. The system transcribes the audio, translates it (often via fine-tuned Whisper models), interprets the image, and pulls answers from the national agriculture manual. Responses return as text or audio in Chichewa.

In a three-month pilot, it handled over 4,000 queries, cutting advice delivery from days to seconds. Chichewa's low-resource status[3] means standard models stumble until locally optimized. And because many farmers are illiterate or semi-literate[4] and use phones mainly for calls and messages,[5] voice and photos are a far lower barrier than text. The system still breaks on background noise or thin training data, gaps local agents quietly bridge.

Jokalante NAFOORE: Voice-First Climate Advice in Senegal

In Senegal, Jokalante's NAFOORE platform takes a different but parallel route. Farmers access climate and agronomic advice through voice calls, radio, and WhatsApp in multiple local languages. The new AI voice chatbot upgrade draws on local meteorological data and agroecological knowledge to deliver personalized guidance on crop varieties, soil health, and pesticide reduction.

Here, the constraint is not only literacy but the sheer ratio of farmers to extension agents: one agent for every 15,000 producers in some areas. Voice becomes the only scalable channel. The system succeeds when it speaks the farmer's language and respects local growing conditions.

Capitec Bank: Conversational AI in South African Fintech

Fintech offers a parallel window into customer-facing reality. Capitec Bank in South Africa now routes half of its 15 million annual client conversations through WhatsApp using conversational AI. Customers check balances, dispute charges, or apply for loans without downloading another app or waiting on hold. The system handles asynchronous exchanges that mix voice notes, rapid-fire text, and follow-up queries. This human-centric automation has pushed first-contact resolution to 71%.[6]

Masakhane: Building Open-Source African Language Infrastructure

None of these tools would exist without foundational work on African languages. Masakhane, the pan-African grassroots collective, has spent years building open-source machine translation models, named-entity recognition datasets, and speech corpora for dozens of languages spoken by millions.[7][8] Their community-driven approach directly feeds the transcription and translation layers that UlangiziAI and NAFOORE rely on.

The collective's recent push into multimodal datasets spanning text, speech, and culturally grounded visuals reinforces the same lesson: language infrastructure is not a preprocessing step, it is the prerequisite.

The Hidden Infrastructure Problem

These breakthroughs run on thin margins and "noisy" data. While a lab-trained AI expects a quiet room and a single language, African reality is a crowded Lagos market where Automatic Speech Recognition (ASR) must contend with overlapping voices, background traffic, and rapid-fire code-switching. High-quality datasets for these languages remain scarce,[9] and because dialects shift by village rather than by border, a model that understands one user may be deaf to their neighbor.

The missing link remains a participatory infrastructure. We need continent-scale datasets that prioritize the 'grit' of real-world usage over the sanitized sterility of lab recordings. Until we bridge this gap, AI deployments will continue to lean on 'hidden scaffolding' like local agents and manual transcriptionists who quietly rescue models when they fail to parse a chaotic voice note. The industry should stop dismissing linguistic nuances as edge cases and start recognizing them as the primary dataset.

Reframing the Baseline

Global AI products are built on a fragile set of assumptions: stable text-first input, high-literacy users, and monolingual environments. African realities expose these assumptions as brittle. In regions where mobile phone adoption has outpaced formal schooling, high literacy is a luxury that design cannot afford to take for granted.

Successful systems in these markets reframe the entire design problem through three core shifts:

• Dynamic Multilinguality: AI must follow the user’s code-switch in real-time rather than forcing a single, rigid language choice.
• Behavior-Driven Interfaces: Systems must be built around existing messaging and voice habits rather than forcing users into polished, unfamiliar chat windows.
• Voice as Primary: Voice and image-based messaging become the foundational channels, not just graceful fallbacks for "low-bandwidth" environments.

When product teams treat these conditions as the baseline rather than the exception, they build something more robust: systems designed for the way the world actually communicates, suited for any high-diversity, mobile-first market on the planet.

Conclusion

Africa is not a test market or a special-needs appendix to global AI strategy. It is an early indicator of what the rest of the world will soon face. The deployments that work today in Malawi, Senegal, and South Africa demonstrate that multilingual, multimodal AI succeeds when it is engineered from the constraints outward (i.e., voice-native, messaging-first, locally grounded).

The lesson for anyone building global products is straightforward: treat language diversity and modality fusion as core architecture decisions, not localization afterthoughts. The tools that win will be those designed for the way people actually speak, share, and solve problems.

About LILT

LILT is a multilingual applied research lab partnering with researchers to design custom evaluations, benchmarks, and RL environments that measure real model behavior across 200+ languages. As a founding member of CLEAR (the Coalition for Language Equity in AI Research), LILT works alongside academics, industry leaders, and global nonprofits to improve representation of underrepresented languages in AI systems, building the open datasets, benchmarking standards, and community engagement playbooks that the next generation of multilingual infrastructure demands.[10]

References

[1] Askyazi. (2025). WhatsApp usage across Africa: Key statistics & insights for 2025. https://www.askyazi.com/...

[2] DataReportal (via Askyazi). (2025). WhatsApp statistics by country (2025 estimates). https://www.askyazi.com/...

[3] Lelapa AI. (2024, October 25). InkubaLM: A small language model for low-resource African languages. https://lelapa.ai/blog/inkubalm-small-language-model

[4] RobotsMali. (n.d.). AI, illiteracy, and written language in Mali. https://robotsmali.org/en/ai-illiteracy-and-written-language-in-mali/

[5] Pew Research Center. (2018, October 9). Majorities in sub-Saharan Africa own mobile phones, but smartphone adoption is modest. https://www.pewresearch.org/…

[6] LivePerson, "Capitec: Pioneering a Banking Revolution with Conversational AI" (Case Study, 2024). Published by LivePerson, Inc. Available at: https://www.liveperson.com/resources/success-stories/capitec-conversational-banking/

[7] Masakhane. (n.d.). Masakhane – Machine translation for Africa. https://www.masakhane.io/

[8] Orife, I., et al. (2020). Masakhane – Machine translation for Africa. arXiv. https://arxiv.org/abs/2003.11529

[9] Statista. (2026, January 26). Africa: Number of living languages by country 2022. https://www.statista.com/…

[10] LILT Team. (2026, February 10). Introducing CLEAR: The Coalition for Language Equity in AI Research. https://lilt.com/blog/clear-the-coalition-for-language-equity-in-ai-research