By 
By Ross Moyo
For years, talking to a machine in Shona felt like talking to a robot. Now Econet says its chatbot understands 8 out of 10 queries, handles Shona-English mixing, and will soon tell a Zezuru speaker from a Karanga speaker – and answer back in their own dialect.
Econet Ai – a wholly owned subsidiary of Econet Wireless Zimbabwe – has launched Shona 2.0, the latest version of its Shona Artificial Intelligence (AI) model.
AI models power Yamurai, Econet’s multilingual customer support chatbot, with the intelligence that enables it to increasingly handle most of the company’s customer inquiries.
Econet said the model now has the capability to comprehend and answer 80% of customer queries, putting it on par with international languages, and marking the first time an African AI language model has reached this level of comprehension.
That last claim matters. If it holds up to independent testing, it is not just an Econet upgrade. It is a continental milestone.
For decades, African languages have been classified in AI research as “low-resource” – meaning there is not enough clean, labelled text, speech and conversation data to train models to the same standard as English, French or Mandarin. African languages remain underrepresented in natural language processing (NLP), with most corpora limited to formal registers that fail to capture the vibrancy of everyday communication.
Econet is now saying: we have closed that gap, at least for customer service in Shona.
What Shona 2.0 actually does
According to Econet, Shona 2.0 can now understand customers who mix Shona and English when speaking, a common practice that often confuses AI Models.
Anyone who has lived in Harare, Bulawayo, Mutare or Masvingo knows that practice. It is not “bad Shona” or “bad English.” It is how Zimbabweans actually speak:
“Ndoda ku-checker data balance yangu, but bundle rangu rapera?”
“Ndaka-recharger airtime but haisi ku-reflecter, mubatsirei?”
“PUK number yangu ndaikanganwa, mungan-di-send-er here?”
That code-mixing – Shona-English slang, verb borrowings, English nouns with Shona prefixes – is exactly what breaks most global models. They were trained on formal corpora, struggling with slang and code-mixed inputs.
A 2025 study of Shona-English slang found the same problem and built a dataset from anonymised social media conversations, annotated for intent, sentiment, dialogue acts, code-mixing, and tone, to prove it could be solved. That work fine-tuned a multilingual DistilBERT classifier for intent recognition, achieving 96.4% accuracy and 96.3% F1-score.
Econet’s claim is different in scale. It is not a research demo. It is a production system answering real customers, on WhatsApp, 24 hours a day, about real money – airtime, bundles, PUKs, SIM swaps, roaming.
Yamurai is the WhatsApp Chatbot Virtual Assistant from Econet Wireless, reachable by sending messages to the WhatsApp number +263 781 222 000.
This multi-lingual assistant can communicate in Shona, Ndebele, English, and other recognised languages. It offers guided interactions through predefined buttons, making it easy for customers to access automated services 24/7. It helps with tasks such as SIM replacement, international prepaid roaming, checking data bundle balances, retrieving PIN & PUK codes, addressing over-scratched cards, managing out-of-bundle browsing, reporting network issues, and transferring airtime.
If Shona 2.0 can truly handle 80% of those queries in Shona, in Shona-English mix, without handing over to a human, that is a transformation in service delivery for a company that dominates Zimbabwean telecoms.
The dialect breakthrough
Econet Ai has also begun to train the model in six different Shona dialects. It will soon be able to identify the dialect a customer is using and automatically respond in that dialect.
This is the most culturally significant part of the announcement – and the most technically difficult.
The Shona language is spoken in five dialectal clusters, which are ChiKaranga, ChiManyika, ChiNdau, ChiZezuru, and ChiKorekore. ChiKaranga is spoken in areas surrounding Masvingo town in the south, ChiManyika in areas around Mutare in the east, ChiNdau in Chipinge in the south east, ChiKorekore in Mutoko in the north east, and Chizezuru in northcentral Zimbabwe.
It should be noted that Shona dialects are mutually intelligible and are, thus, dialects of the same language. Indeed, 80 to 90 per cent of the vocabulary is common to the whole area of Shona speaking communities.
But that 10% to 20% difference is where dignity lives.
For decades, school curricula, textbooks, examinations, and teacher training privileged Standard Shona (Zezuru/Karanga core). Students from Manyika, Ndau, or Korekore backgrounds who used their home lexical items, phonological features, or morphological patterns were marked wrong.
A diglossic situation emerged. Zezuru occupied the highest prestige position, Karanga an intermediate but still elevated status, and Manyika, Ndau, and Korekore lower status. Speakers of the latter were often ridiculed for “incorrect,” “rural,” or “funny-sounding” speech.
Econet is now promising to reverse that logic in a chatbot: instead of forcing the customer to speak “Standard Shona” to be understood, the machine will learn to hear the customer as they are – and answer back in that dialect.
Econet has not yet published the list of the six dialects it is training on. The five canonical clusters are Zezuru, Karanga, Manyika, Ndau and Korekore. The sixth may be a sub-dialect – such as Budya, or a distinct urban Harare variety – or Kalanga, which some researchers treat as part of the Shona cluster. This newspaper has asked Econet Ai to confirm the six.
What matters is the principle: dialect is not an error to be corrected. It is data to be learned.
“Trained just like children”
“AI models are trained just like children; they learn and continually improve from listening to our customers. So, you find that in the beginning Yamurai in Shona was like speaking to a robot, but today it is getting much more like a person. It is getting better all the time,” said Econet Ai CEO Mr Navdeep Kapur.
That metaphor is more accurate than it sounds.
Large language models improve not by being programmed with grammar rules, but by being exposed to millions of real conversations, corrected, and retrained. Econet says there is a full-time engineering team working on continually training and improving the Yamurai language models.
That team is part of a much larger build-out. Econet AI recruited more than 100 engineers and specialists. Its partnership with Cassava AI provides access to NVIDIA accelerated computing infrastructure, enabling practical AI solutions across telecommunications, finance, agriculture, insurance, mining, healthcare, retail, logistics, education and government.
In other words, Yamurai is not a side project. It is the front door of a new AI division.
Ndebele is next
Ndebele 1.0 was launched last month, and work is at an advanced stage to get Ndebele 2.0 out before the end of the month.
That timeline is aggressive – two major language upgrades in less than 60 days – and it signals Econet’s sequencing: Shona first, because it is the largest language; Ndebele second, because it is the second largest; then, according to earlier Econet statements, Tonga, Kalanga and other languages spoken in Zimbabwe.
Econet is currently the only mobile operator in Africa with a multilingual AI helpline offering support in local languages. Its Yamurai service answers questions in English, Shona and now Ndebele, with Tonga, Kalanga and other languages spoken in Zimbabwe set to be added soon – a critical inclusion strategy for rural and elderly users.
That claim – “only operator in Africa” – has not been independently verified by this newspaper, but no other African operator has publicly demonstrated a production chatbot handling two indigenous languages at this scale.
The company said Yamurai had drastically improved Econet’s ability to respond to customer queries and significantly reduced customer waiting times on the helpline, ensuring that all queries are addressed.
The scale of that improvement has been reported before: Econet AI’s customer service platform, Yamurai, has cut response times from between eight and ten minutes to just three seconds, effectively giving each customer a personalised digital assistant.
Earlier reporting also noted that Econet’s Shona-speaking chatbot, Yamurai, can already handle more than 70 per cent of customer queries, alongside about 600 human contact-centre staff.
Econet’s new figure – 80% – therefore represents a jump from 70%+ to 80% in comprehension. In a call centre that serves millions, a 10-percentage-point jump means hundreds of thousands fewer people waiting on hold.
“Real light at the end of the tunnel”
Econet is continuously expanding Yamurai’s capabilities as it seeks to improve the quality of its service delivery. Through Econet Ai, the company is tackling problems many operators in Africa have struggled with for years without any real solution.
“Now there is real light at the end of the tunnel, as AI opens solutions at all levels of telco operations. It takes time to get them all deployed but we are confident that we are only months away from having the best service in Africa,” said Mr Kapur.
Econet said the development of multilingual and agentic chatbots forms part of its broader strategy to use AI to improve customer service, increase accessibility and provide faster support across its platforms.
“Agentic” is the key word there. It means a chatbot that does not just answer – “Your balance is X” – but acts – “I have recharged you, I have blocked the lost SIM, I have escalated the network fault to the team in your area.”
That is the next frontier for Yamurai: from answering to doing.
Econet Ai was launched earlier this year to deliver advanced AI solutions to the Zimbabwean market.
Econet Ai works closely with Cassava AI, the artificial intelligence business of Cassava Technologies, to bring advanced AI solutions to Zimbabwe. The collaboration connects Econet Ai’s local expertise with Cassava AI’s global technology partnerships and enterprise AI capabilities, supporting the development and delivery of solutions tailored to the Zimbabwean market.
Cassava Technologies president and chief executive officer, Mr Hardy Pemhiwa, said the launch represented a structural shift for Zimbabwe’s economy and digital future. Cassava had invested in Africa’s first AI Compute Centre to serve the region and confirmed that Econet AI would now access world-class GPU computing power and AI cloud services locally.
Econet AI will roll out Cassava AiCloud, a platform that allows Zimbabwean enterprises to run data-intensive workloads using high-performance computing powered by NVIDIA GPUs. The infrastructure is hosted at the Cassava AI Factory in South Africa, currently the only facility of its kind on the continent, enabling firms to deploy AI tools without building their own systems.
That is why Shona 2.0 is not just a chatbot upgrade. It is a proof point: if Zimbabwe can train a model in Shona, with dialects, with code-mixing, to 80% comprehension, then it can train models for agriculture, health, education, mining and government – in Shona, in Ndebele, in Tonga, in Kalanga.
Comments