A company born from a gap between two numbers
We develop and run Lazao, an artificial intelligence assistant for the phone, which collects payment through mobile money. Sainillar was founded in 2026 to develop it, run it and sell it.
Why this company exists
We started by reading a table. On one side, a continent that solved its payment problem through the phone and moves 1.4 trillion dollars a year on that rail. On the other, an artificial intelligence industry that can only collect payment by international card, held by three adults in a hundred.
Between the two, nothing. No technical obstacle, no lack of demand: Kenya is the world's leading country for AI assistant use. Nobody was doing the integration work, long and to be redone in every country: one mobile money operator and one tax regime per country, languages absent from the training sets. We took it on.
The name says the same thing. Saina, intelligence; pillar, a pillar. We don't train the models. We develop the payment, the routing, the languages and the monitoring that let them reach a phone in Kisumu or Antsirabe.
What we do
We develop software for the phone. An installable web app, an iOS app, and a server that streams answers, collects mobile money and lays out Word, Excel and PowerPoint documents.
We do applied research in artificial intelligence. Evaluating models on Swahili and Malagasy, writing the per-request routing, the fallback between providers, web search detection and request classification.
We put parties in touch. Advertisers and ad networks on the free tier, buyers of aggregated data under k-anonymity, payment operators in each country.
What we make public about the company
The legal form, country of incorporation and the identity of the director appear in the legal notice, as the law requires. The company is funded from its own resources, with no outside investor. We do not communicate beyond that: not on capital, not on ongoing discussions, not on individuals.
What we publish can be checked: the models in service, incidents and their duration, the subprocessors and what each one sees, the open countries. The rest, capital, ongoing discussions, people, stays private.
Mission
The world is accelerating with artificial intelligence. An entire continent cannot pay for it.
People want to use it, and know how to use it. They cannot pay, because the subscription requires an international card and the continent pays by phone. Everything we develop starts there.
Accounts are spreading. Cards are not.
In ten years, sub-Saharan Africa has solved much of its financial inclusion problem. The share of adults with an account rose from 34 % in 2014 to 58 % in 2024. That is a considerable shift, and it happened almost entirely outside banks: through the phone.
But artificial intelligence is not sold through a phone. It is sold by subscription, on an international card. And there, the curve barely moves: debit card ownership went from 17.9 % to 22.2 % over the decade, and credit cards, the ones most subscription services require, sit at 3.1 %.
In other words: ninety-six adults out of a hundred in sub-Saharan Africa cannot subscribe to an AI service, even when they have the use for it, the appetite and the money. They have the money on their phone and no way to get it into a bank card form.
3.1 %
58.2 %
22.2 %
40 %
Demand is already there
It would be easy to conclude that the market does not exist. The figures say otherwise, spectacularly enough to be worth pausing on.
In July 2025, Kenya ranked first in the world for ChatGPT use among internet users: 42.1 % of connected Kenyans aged sixteen and over reported using it in the previous thirty days. Ahead of the United Arab Emirates. Ahead of Israel. More than double the United States, at 19.1 %.
A country where 4.4 % of adults hold a credit card is the world's heaviest user of artificial intelligence, in proportion. What those 42 % use, they use without being able to pay. That is what we are changing.
What the continent already pays, and how
The payment rail exists. By value, it is the largest in the world. It simply is not the one AI companies accept.
$1.4 tn
347 M
87.5 %
under 1 %
Two countries first
The same gap everywhere, not only in Kenya
Kenya is the most spectacular case, but it is not an exception: across the continent, mobile money is widespread and credit cards stay marginal. It is a continental pattern, and that is what makes the answer transferable from one country to the next.
Madagascar sits at the other end of the scale, with one adult in four holding an account. That is also why it opens at the same time: a service that only worked in Kenya would have proved very little.
- Kenya
- Madagascar
Share of adults, %. Sources: World Bank, Global Findex 2025 (2024 data).
So we took the work nobody wanted
There was no new payment method to invent, and no card to wait for. We developed the integration nobody was doing: M-Pesa and MVola directly, an aggregator for the other operators, each country's VAT and digital tax, and the evaluation of models on languages absent from their training sets.
Slow, repetitive, to be redone country by country. That is precisely why it was free to take.
Four commitments that hold the rest
Four rules, each written into the code and each with a cost we pay.
The language of the question
An assistant that answers only in English and French does not serve the continent; it serves the part of it already served. Our platforms answer in the language of the question wherever the model can, and native speakers review each language before we offer it. A language loses its listing if review fails twice in a row.
The country's currency and rail
Prices are shown in local currency, in round amounts, and settle through the payment method people already use. We absorb the mobile money operator's commission, the payment aggregator's fee and the currency conversion cost. What you see is what is debited.
Honest billing
When a model fails or a request does not go through technically, the quota is not counted. That sounds obvious; it is not, and it costs us real money, since the provider still bills us for the call. We carry it because charging for a failure is the fastest way to lose the trust of someone who counts their units.
The continent as the main market
Developed for an entry-level Android phone on a 3G network: streamed answers, light pages, a web app that installs without an app store, quotas that free up over 24 hours rather than at midnight.
Approach
We don't train models: we route them and we measure them
We don't train models. We run several, from several labs, behind three names; on every answer we measure the billed cost, the time to first token and the failure rate, and we pay the payment commissions on your behalf.
Why three names and one routing rule
On everyday questions, understanding a text, drafting a message, translating, summarising, an inexpensive model answers as well as a far more expensive one. For a service sold at four dollars a month, paying for the expensive model where nobody would see the difference would end up in your price.
So Mwezi serves short questions, Jua homework, contracts and code, Nyota long reasoning and documents, and a classifier decides on each request. We review that routing with the month's measurements: billed cost per answer and time to first token.
Sainillar is not affiliated with any model publisher, nor sponsored or endorsed by them; any brands mentioned belong to their respective owners.
What we measure, continuously
Real cost per answer
Not an average estimate: the amount the provider bills for each answer, tracked per user and per day. That figure decides which model serves which use.
Time to first token
On an unreliable network, a perfect answer that takes eight seconds to start is a bad answer. We measure the time to first token of every answer; when a model becomes too slow at the median on short questions, we take it out of the routing, and we switch off reasoning mode wherever it does not improve the answer.
Margin per subscriber
The price is fixed; usage is not. We track margin per subscriber per month and adjust quotas rather than price, because a price that moves is a price nobody trusts.
Fallback between providers
Each model name hides several engines at several providers. At the first timeout, the request goes to the next one, and we route to the fastest host measured over the day. You see an answer, not an error.
What we absorb, and what we cannot
The price shown on our platforms is the amount that will be debited. We absorb the mobile money operator's commission, the payment aggregator's fee, and the cost of converting between the local currency and our billing currency. The rate is the same everywhere and converted into round amounts in each currency, so the price stays readable and memorable.
What we do not control is what your own operator or bank charges you on their side: agent withdrawal fees, account maintenance fees, the exchange margin your bank applies to a card. Those amounts fall under your contract with them and we have no way to touch them.
We would rather write that here, in the same paragraph and the same type size as the promise, than in a footnote.
What runs on its own
An internal dashboard brings together cash, cost per answer and per country, quotas, abuse and alerts.
The security review runs after every deployment, usage reports produce themselves, the fallback between providers switches on its own. What stays manual: the decisions, the languages, the people.