“What do I do?” The real-life example in the title is not the kind of prompt that AI labs optimise for, but it’s the kind that the Canadian company Viamo hears. The questions arrive by phone call, often from the cheapest handsets on the market, in languages that, until recently, no major AI model officially supported.
Last month, MTN Rwanda launched Miss Baza, a voice assistant built on the same approach. Callers dial 845 on an MTN line, ask a question in Kinyarwanda, and get a spoken answer. Viamo provides the voice platform, Google’s Gemini handles the conversation and Rwandan firm Pindo contributes the language infrastructure. No smartphone, data bundle or app necessary.
TelcoForge’s recent coverage of the digital divide has focused on the price of the device and the skills to use it. Viamo’s argument is that, to bridge the AI divide, neither a smartphone nor special skills are necessary. “We can turn all of those devices that are already out there into smart devices for these audiences,” says Kellen Eilerts, Viamo’s CMO.
Users on the Menu
Viamo began as a way for development agencies, governments and some large companies to contact people. “People would be using radio, but you don’t know who you’ve reached and what they’ve retained and how they’ve acted on that,” Eilerts observed. Viamo’s telephony-based service offered precise information.
Originally, Viamo supplied menus of local-language audio content on a range of topics, which callers navigated with their keypad, and in return the operator gave users a large allowance of free access. “We wanted it to be user-driven, rather than us pushing something out and saying we think you should hear this,” he said. “We knew that these populations had limited means of paying.”
An early GSMA grant to analyse usage data showed real demand for that kind of content. Over a dozen years the service spread across 20 countries, mostly in Africa and some of South Asia. Last year, Eilerts says, around 27 million people made more than half a billion searches on it.
“What we didn’t realise in building that platform out was that at the same time we were creating what is now a distribution layer for access to AI over a phone call in local languages, as those languages become supported by AI.”
Once ChatGPT appeared, Viamo started testing whether callers could speak instead of pressing buttons, and get a personalised answer instead of the same recording everyone hears. Serious investment followed when Google took an interest. “What you actually have here is search from populations who aren’t searching on Google,” is how Eilerts sums up Google’s reasoning.
Together they tested whether it would work in poorly supported languages, over ordinary operator telephony, and for people with no professional use for AI. “They’re not going to use it to edit their emails,” he says. Nor, he adds, are they “ever going to pay for a Gemini Pro account.” After around 18 months, Viamo is confident enough to start scaling the service across its whole footprint.
“What Is This?”
However, the early results were not encouraging. Viamo counts a call with AI a success when the caller says something that shows they know they are talking to an AI and are using it for something an AI can do. At the start, only about 20% of callers managed that. Many said nothing at all; others asked “Hello, what is this?”
Over the past year, that success figure has risen to around 70%. Some of the gain came from teaching people through their mistakes. If a caller is silent, the system explains how to hold the phone and ask a question, with examples. If the system can’t make out what they said, it plays back its best guess, so the caller understands it is trying to listen and can clarify.
But Eilerts is clear that the interface was not the main lever. “The biggest single driver of successful behaviour is the value proposition that you present to them,” he says. One that works especially well is offering a private confidant. “You can use this service to talk about things that you wouldn’t feel comfortable talking about with another person.” For example, questions about sexual health or help with money problems.
A reason to call matters, because the experience isn’t always straightforward. “People are using the cheapest possible phone in a noisy environment, there’s a goat bleating in the background, so the language model doesn’t fully capture what they say,” Eilerts explained. A strong enough reason gets people through that, and once they stay, they “go on and explore many other ways that they can use the tool.”
It’s a lesson that applies well beyond feature phones. People will put up with friction for something they want; few will learn a new interface just because it exists.
Privacy Please
Viamo has now handled close to 3 million of these AI question-and-answer exchanges. Nigeria and Rwanda are the two markets where it runs at scale. Of Viamo’s roughly 1.8 million monthly users in Nigeria, a large share live in the north, an area Eilerts calls “very obviously isolated and resource-poor.”
The confidant model is visible in what they ask. “We have a huge amount of questions related to different health topics, sexual reproductive health in particular,” Eilerts says, along with questions about conflict and security. “There’s a lot of those very sensitive topics coming through.”
That shapes the design. On health, the service is “informational rather than diagnostic,” and quick to refer callers to real services.
The next step is to go from answering questions to connecting people with help. “If someone asks a question where they need to go to the clinic, well, can we actually tell them where the nearest clinic is to them in northern Nigeria?” Eilerts asks.
Talk Is Cheap
Eilerts grumbles that the AI industry is looking for growth in the wrong place. AI investment, he argues, is pouring into ever-better English models while ignoring what can already be delivered to the rest of the world, on networks and handsets that already exist. “We don’t have to wait for the satellites or the balloons or whatever else is out there,” he commented.
Even those who want to deliver AI services in these different languages reach instinctively for the wrong channel. Viamo has been approached by organisations that want to make AI available to everyone over SMS. “Why would we do it over SMS? We don’t need to do it over SMS,” he sighed.
“Even the operators [aren’t] thinking about this as an opportunity, much less the governments who could be providing constituent services over voice.”
Voice isn’t automatically easy: Viamo’s 20% starting point shows that talking to a machine has to be learned. But a phone call is a channel nearly everyone already knows, on a device nearly everyone already has, and it requires no additional data plan, app store or password.
That makes Viamo a bellwether for more than emerging markets. In Japan, the most determined smartphone-training effort in the rich world still leaves around three in ten people in their early 80s without a smartphone, even as analogue public services are withdrawn. A number to call, answered by an AI in the caller’s own language, is a plausible way to reach them – provided it comes with a clear reason to call and gives feedback when callers go wrong, as Viamo’s numbers suggest.
Caller ID
Once AI moves from answering questions to acting on them – booking the clinic, moving money, claiming a benefit – the service needs to know who is calling, which gets into questions of identity.
Banks have spent the past few years watching AI clones pass voice ID checks, and analysts now expect institutions to move away from voice as a trusted primary factor. A noisy call from a cheap handset is about the worst input a voiceprint could get, and in markets where families share one phone, the device doesn’t identify the person either.
Identifying callers also cuts against what makes services like Viamo’s work. People ask about sexual health and armed groups precisely because nobody is checking who they are. Bolt identity onto the whole service and that confidence could disappear.
However, it would be possible to create a tiered design for AI-based voice services – anonymous by default for information, verified only when a caller needs a service that relies on identity to complete an action. Operators already hold some alternative tools to enable this kind of approach. Network APIs such as CAMARA’s SIM Swap or Number Verification can confirm that a number hasn’t recently changed hands, while KYC rules could tie the number to a user without asking the caller to do anything.
Some analysts have recently been talking about “next-generation voice” as an opportunity for the telecoms sector. Eilerts would likely argue that current-generation voice is the overlooked opportunity.
Image courtesy of Africa12 on Pixabay
