AI · Harare

Econet's AI speaks Shona. How good is 80%?

Econet AI says its Shona 2.0 model lets the Yamurai chatbot understand 80% of customer queries, including mixed Shona and English.

Kapture · · 2 min read

A mobile network mast in Zimbabwe
Photo: Nyasha L Guzha, CC BY-SA 4.0 via Wikimedia Commons

Econet AI, the artificial intelligence unit Econet Wireless Zimbabwe set up this year, has launched Shona 2.0, a new version of the model behind its Yamurai customer-support chatbot. The company says Yamurai can now understand and answer 80% of customer queries in Shona.

That is up from more than 70% only a few weeks earlier, Techzim reports. Econet says the model also copes with people who switch between Shona and English mid-sentence, which is how much of Zimbabwe actually talks.

What Yamurai does

Yamurai is not new. Econet launched it in 2020 during the COVID lockdown to handle routine problems such as PIN resets and airtime and bundle issues, passing harder cases to human agents. Econet's chief executive, Navdeep Kapur, has admitted the early version felt like talking to a robot, Techzim notes.

The roadmap is ambitious. Econet says it is training the system on six Shona dialects so Yamurai can reply in the customer's own dialect, and it is building an Ndebele version. It calls Shona 2.0 a first for an African language model at this level.

What 80% does not tell you

Techzim's caution is fair. Econet has not said what the 80% measures: all calls, only Shona calls, or a test set. Understanding a question is not the same as solving the problem, and it is unclear how often Yamurai fixes an issue rather than handing it to a person. There is also no shared benchmark for African languages, so the "first" claim is hard to check. Techzim plans its own tests in straight Shona, mixed Shona and English, and with a real problem such as a failed EcoCash transaction.

There is a jobs question too. Past Econet releases said the bots work alongside about 600 contact-centre staff. A chatbot that handles most Shona queries changes what those people do.

Econet is not alone. Six large operators, including MTN and Airtel, are building African language models through the GSMA. Whoever gets local-language AI right first will own a powerful advantage: customers who can be served at scale, in their own words.

For creators and businesses, the opportunity is bigger than customer care. A model that understands Shona and code-switching could power voice notes, captions, search and content tools that finally work for the way Zimbabweans speak.

Code-switching is the hard part. Models trained mostly on formal text tend to stumble when a sentence starts in Shona and ends in English slang, which is how many customers actually type and speak. If Econet has cracked that, it matters more than the headline number.

Our read: real local-language AI is a prize worth chasing, and Econet is right to chase it. The proof will not be a percentage in a press release. It will be in the PIN resets.

Written by the Kapture newsroom.

More from Kapture