Azizi accidentally built Africa's voice internet — and now we're leaning in

Africa's abundant AI future will unlock the next billion creators — and leapfrog literacy, even the computer

Meet Nancy

Nancy deep-fries cassava for school kids in Western Kenya. She doesn't read or write well. This morning she called a number, spoke Swahili to a voice that answered back in Swahili, asked if there was stock, got a price, described her location in landmarks (nobody out here is sending you a Google Maps pin, my friend), had her order read back, and paid with an M-PESA prompt. Five minutes. No app, no field agent, no QR code she'd need her neighbour to explain, no form, no literacy tax.

She wasn't talking to a chatbot. She was talking to a marketplace that happens to be a coach and a business partner — one that's already done the generic advice and so doesn't waste her time with "have you considered diversifying your revenue streams." It just answers. Stock is there, pay this much, driver will call when he's near the chemist by the primary school. Done.

That's Azizi.

So what was the problem again

If you've spent any time building for rural African users, you know the shape of this. Organisations pour money into selling and training and "engaging" end users across the continent, and most of that money goes to field forces because somebody has to be the bridge between HQ and the guy with the groundnuts. Field forces are expensive, hard to manage, and prone to every human failure mode from fake reporting to broken-telephone training updates. So some smart person says "let's give the users an app." The app gets built, and it gets used by maybe fifteen percent of the people it was built for, because everyone else either can't read it, can't afford the data, or is busy frying cassava.

The companies that actually care about this try to fix it with custom platforms and in-house tech teams and "innovation" units, and sometimes they pull it off; but by the time a lesson learned in April reaches the customer in production, it's August, and by August the customer has moved on, or worse, quietly churned to a competitor who was simply more reachable. "Focus" gets invoked a lot in these rooms. Focus is the word organisations use when they're about to stop doing the thing that would actually help.

The unlock is this: the users can't read the tools, and the builders can't hear the users. Everything else is a symptom.

Voice fixes it. But voice alone isn't enough — not the version you're imagining, anyway. You also need WhatsApp — so the call gets a text summary that somebody's grandson can read back later, with a photo of the merchandise. And USSD — so it works on the cousin's not-smartphone. And:

• Analytics — so the organisation can actually act on what she said.
• Safety engineering — so a beneficiary can't get jailbroken into hearing something dangerous.
• African-accented voice processing — so she gets understood and not "I'm sorry I didn't catch that."
• Local language — so the whole interaction feels like a human and not a badly dubbed telenovela.

Built separately, these things are a nightmare that have toppled the most bullish founders in Africa. Built together, they're Azizi.

Why right now, specifically

Two years ago, this company would have written off "Ag" and wound down.

Voice models weren't good enough — the latency made conversations feel like you were calling from the moon, and the accents got mangled enough that rural users would hang up in the first thirty seconds. African language data was locked up in academia or didn't exist. And the funding environment still rewarded "let's hire more field officers and field managers" as a credible growth strategy.

All three shifted.

Voice quality crossed the threshold in late 2024, and 2025 was the year the latency got boring enough to hold a real five-minute conversation. Recently, huge language corpuses have been dropped by Google and Microsoft. The data moat that would have protected a would-be competitor has evaporated. What's left as defensible is distribution, accent and dialect engineering tuned for real African phone-line conditions, and the conversational data that only comes from live deployment. Azizi has all three.

And the economics. Post-2023, nobody has money to throw at field forces. Organisations that used to fight bottom-line automation are now calling us.

None of this was true in 2023. All of it is true now. The window is eighteen months, maybe twenty-four. After that it's a scramble, and scrambles favour whoever already has community and distribution.

"So are you an AgriTech company or an 'AI' company?"

Both. And before you accuse us of being AI-hype tech bros trying to ride the wave, know that we moved to rural Western Kenya in 2024 and spent over a year there. In a village. Not a field office we fly into for workshops. An actual office with a lease where actual Nancies walked in and told us our first version was terrible. We crossed into Uganda for produce because the East African Community trade agreement is a real thing if you try to use it rather than just cite it in a pitch deck — as long as you don't take eyes off the produce from the aggregator, and follow the truck across the border. That's another story.

We built the omnichannel marketplace first. Voice came later, as the last bet we placed on the marketplace's usability for the users we wanted. Turns out that voice is what every other organisation in our ecosystem was quietly desperate for: weather services, farm hardware companies, agricultural education programmes, nonprofits with coaching curricula they can't scale, even banks and power-grid companies. So we started selling it to them. Now we're an enterprise AI company that happens to run a marketplace, or a marketplace that happens to sell AI to enterprises, depending on which cap we're wearing on which Tuesday. The technically correct term is "UX infrastructure flywheel," but it mostly means we're too useful to both sides to pick one.

Focus, as mentioned, is for companies that have the luxury of doing only one thing. We don't. And anyway, both things are the same thing: the same phone line Nancy calls is the one an NGO deploys for its beneficiaries. The difference is who's paying, and what language they understand best.

Everything is a product. The culture. The jobs.

We don't have employees. Before you raise your HR eyebrow, this isn't a gig economy cost-cutting thing. It's the opposite. Every piece of work at Azizi is a Bounty: a specified unit of work with AI-hardened acceptance criteria — edge cases, documentation requirements, and proof that real users have tried it and tried it again. Meet the criteria, get paid. Doesn't matter if you shipped it in an hour or a week. Doesn't matter if you're a human, an AI, or a human with an AI agent doing most of the typing.

What we govern is the shape of the work. Bounties have to be small enough to finish before the strategy pivots. They deliberately leave room for what users actually need instead of what the exec team imagined last quarter. The executive function is ensuring a roadmap of ever-meaningful bounties. Bounties that don't target job replacement, but rather treasure maps for our software pirates to chase uncharted value, each expedition ending with a real rural user delighting in tech. To work here is to brave adventure, replete with setbacks and close shaves.

Humanity is in the middle of a generational re-think about what constitutes work. The industrial factory model is dying. A lot of people who trusted it have been hurt, and "just learn to ship end-to-end" is not a universally helpful response to that, nor are handouts. Fair. What we can say is that our pirates have more fun, higher stakes, ship faster, and own their work in a way traditional software careers rarely allow. If we had more money we'd make bigger treasure maps. That's on the roadmap too.

We've extended the same logic to sales. A substantial commission on the first $500,000 of voice credits a new customer buys — paid to whoever closed it. Could be a partner, an existing customer, or an investor who wants more than passive equity. Works because our call costs are about $0.50/min including the phone lines, plus the less-than-five-minute tangible value conversion, plus the tech support — versus the $0.05/min + $10,000-a-month toll-free call-centre alternative that also needs a swarm of trainers, content recorders and hardware for each cell provider to keep up with the next Swahili groundnut-seller product launch.

Commission-productised sales is how we scale revenue without hiring a sales team we can't afford and don't want. The users love talking to the model at their own convenience, not a sales team paid to stalk and pretend to know them.

What we're actually building

The enterprise business pays for the conversation rails while reducing the cost and complexities of operations and go-to-market, and in return, stays connected to their customers. The rails are not the company.

The company is the voice-internet. The multilingual voice app store.

Eighteen months from now, the same infrastructure Nancy uses to order cassava is open to anyone who wants to build on it. A pastoralist teenager in Turkana registers as a developer — over a voice call, not a web form, because the whole point is that you don't need a laptop or a formal education to participate — and ships an app that predicts rain and pasture availability in Ng'aturkana. Other pastoralists pay to use it. We take an app store cut. The app is not in English. It doesn't need to be. It doesn't even need to exist in text.

A fellow in Kisumu launches a Luo-language podcast on Azizi, sponsored by the local savings cooperative, with smartly placed engaging ads voiced by our tech on the fly. His listeners are seventy-year-olds on 2G phones who've never used Spotify and never will. They end the call with their vote on next week's topic. The cooperative gets brand lift. The podcaster gets paid. Azizi gets paid.

A kindergarten teacher who's a great storyteller in Nairobi publishes a bedtime app in Sheng, because the kids who speak Sheng don't have Sheng content, because the content industry doesn't know Sheng is a market. Until it does.

A university partners with us on use-case hackathons, and anonymised conversational data research in languages WAXAL didn't cover. Kalenjin dialects, Borana, the dozen variants of Maa that linguists argue about — for maternal-health research, or climate adaptation work, or whatever the grant is about this year. That's revenue we don't have to build ourselves, on data the community generated by using the platform.

This is the real bet: the next billion people online don't show up as consumers of an internet built in San Francisco. They show up as creators of their own, in their own languages, over voice, earning. Africa will have the world's largest young workforce by 2050. Almost all of them won't get seats at whatever the Ivy Leagues have morphed into by then — but all of them will have phones, will likely be doing some sort of small-scale agricultural pursuit, and all of them will have something to contribute. Whether they say it on a platform owned by somebody who's never been to Turkana, or on a model fine-tuned here, using win-win data their own communities generated, earning in their own economies — that's the challenge the next eighteen months addresses.

We think the response to that challenge is obvious. We're just the company building it like we believe it.

How to join in making history

Enterprise lead drowning in last-mile coordination? Reach out.

Developer who wants to simplify your UX and APIs by adding a phone number that speaks in a language Silicon Valley has written off? Really reach out.

Investor? Let us empower your portfolio. Monetize your networks. Remember, that 5% commission applies to you too.

Join us in building AI abundance for those who care about the global south, today.

Come find us — hello@aziziafrica.com