Ghana's AI Strategy Needs a Working-Language Test

Strategy is not localization

Ghana launched its National Artificial Intelligence Strategy on April 24, 2026, with a clear policy ambition: Ghana should not remain a passive consumer of AI, but should design, govern, and deploy it for national transformation.¹ President Mahama also put the language question in the foreground, calling for local data ecosystems, indigenous-language integration, and context-aware systems.²

That is the right test. AI that works only in English is useful, but it will not be a national public utility. The working-language test is whether the system can support the languages of actual administration, commerce, health, agriculture, education, and worship. The question is not whether Ghana has an AI strategy. It is whether the strategy can hear Ghana.

Localization is not a paragraph in a strategy. It is a product requirement, a procurement rule, and a national data-building program.

The local-language stack

The minimum stack has five layers: language data, speech recognition, translation, text-to-speech, and task agents that can use those abilities in real work. Ghana has credible activity at each layer, but the system is uneven. GhanaNLP maintains open resources for Ghanaian-language NLP, including datasets and models for languages such as Akan and Dagbani, while gaps remain for others.⁴⁵

A 2026 GhanaNLP parallel-corpora paper reports 41,513 parallel sentence pairs across Twi, Fante, Ewe, Ga, and Kusaal, and notes deployment into real applications such as the Khaya AI translation engine.⁶ The strategy should scale this public-good data work: corpora, evaluation, and working services before demos.

Creators to watch

Khaya AI is one of the strongest production examples. Its automatic speech recognition lists support for Akuapem Twi, Asante Twi, Dagbani, Ewe, Fante, Ga, Hausa, and other languages, and the company frames its work around African-language translation, ASR, and TTS.⁷⁸

GhanaNLP remains important because open language infrastructure is a public good. The strategy should treat GhanaNLP-style work as core infrastructure, the same way a transport plan treats bridges and feeder roads.

Mobobi and Abena AI show what small creators can do when language support is built into a product, not bolted on later. Mobobi describes Abena AI as an offline voice assistant fluent in Twi and Ghanaian Pidgin, with mobile-money actions, translation, TTS, ASR, and more than 500,000 downloads.⁹¹⁰

AdwumaTech AI is another useful signal: an Accra-based AI data-infrastructure company building African-language training data, RLHF, evaluation, and annotation services across dozens of languages.¹¹¹² Ghana's AI strategy should make room for these small and mid-sized builders, not only hyperscalers and national centers.

The procurement test

The strategy documents and dashboards should publish a language-readiness score for each AI project. A chatbot with English-only support can be acceptable for a narrow back-office use. A citizen-facing agriculture, health, education, or justice tool should not pass procurement without a language plan.

Xcelerator additions

The Ghana Xcelerators list should add profiles for the builders who are making language work in public: Paul Azunre and Lawrence Asamoah Adu-Gyamfi for Khaya AI and GhanaNLP, Nana Kwame Ghartey for Mobobi and Abena AI, and the AdwumaTech AI team for African-language data infrastructure. These are not side projects. They are national infrastructure candidates.

A serious localization strategy should fund the builders, publish the benchmarks, procure working systems, and make the language layer reusable across ministries, schools, clinics, merchants, and local governments. Ghana's AI strategy passes only when it works in the words Ghanaians already use to work.