The app was the easy half

  

In 2020, with the country locked down, a team I led launched a telemedicine platform that let Nigerians speak to a doctor from their phone. It worked. Over three million people used it. We could have stopped there, written the case study and called it a success.

Then we looked properly at who was not calling.

Mobile signal reached most of the country. Smartphones did not. Pew Research found that in 2018, thirty-nine per cent of Nigerian adults owned a smartphone, forty-four per cent owned a basic phone only, and seventeen per cent owned no phone at all. Roughly six in ten adults could not have opened our app if they had wanted to.

Think about what that means for a health service. Had we stopped at the app, we would have built telemedicine for the people who already found it easiest to see a doctor. The young, the urban, the connected. Everyone else, which is to say most people, would have been outside it. Not excluded by anyone’s intention. Excluded by an assumption nobody checked.

So we built something else alongside it. A phone line, on all four networks, so that anybody holding the most basic handset could call and speak to a practitioner. No smartphone. No data bundle. No internet. Just a call.

We also left the building. We went into four communities in Osun State and worked through the traditional rulers, because in those communities trust does not arrive through a download link. It arrives through someone the community already trusts saying this is legitimate. That outreach reached around two thousand residents and it is still running.

The phone line and the outreach were harder than the app. They were slower, less fundable and far less impressive in a pitch deck. They are also the only reason the service reached the people it was built for.

I am writing about a 2020 health platform because we are about to repeat the same mistake, at speed, with artificial intelligence.

*The gap is not coverage. It is use.*

Here is the number that should be on the wall of every ministry and every product team in Lagos. The GSMA estimates that around 140 million Nigerians lived within mobile broadband coverage in 2025 and did not use mobile internet.

Not 140 million people beyond the reach of a mast. One hundred and forty million people under a mast, not online.

And that gap is widening. The equivalent figure for 2023 was 120 million. Across Africa, the GSMA and its Partnership for Digital Access, launched during the United Nations General Assembly in September, count 906 million people in the usage gap against 122 million with no coverage at all. That is roughly seven people who could connect and do not for every one person who cannot. The reason is mostly the handset. In early 2025 the GSMA reported that around sixty per cent of Nigerians with 4G coverage had never used mobile internet, and that more than four in ten of those non-users simply lacked an internet-enabled device.

This is where the optimism needs testing. You will have seen the headline from September that Nigerian smartphone ownership has risen to seventy-five per cent. It is a real finding and a welcome one. It is also a survey of 13,251 respondents across twelve major cities. It is an urban number. Read as a national one, it will shape products and policies for a country that does not exist.

The same applies to the figures our regulator publishes. Over 150 million internet subscriptions does not mean 150 million people online. It means SIMs and lines, many of them second and third subscriptions held by the same person. And affordability moved the wrong way. When the tariff increase of fifty per cent was approved in January 2025, the floor price of a gigabyte went from 287.50 naira to 431.25. Analysis of what people actually paid found the average across operators roughly doubled. Usage dipped, then recovered to record levels, which tells you demand is strong and also that households absorbed a real cut elsewhere to pay for it. The GSMA now warns that memory chip prices, which more than doubled between late 2025 and early this year and rose a further eighty to ninety per cent in the second quarter, threaten its plan to pilot forty-dollar smartphones in Nigeria.

The device is the gate. It is getting more expensive, not less.

*Who gets left out of an AI-first service*

Now layer artificial intelligence on top of that and ask who is designed out.

The person without a smartphone. The person with a smartphone but no reliable data. The woman, because the mobile internet gender gap in Nigeria stands at twenty-six per cent, among the widest anywhere, and only about a third of women in sub-Saharan Africa own a smartphone at all. The person whose power supply decides whether they are online this evening. And, more quietly than any of these, the person who does not do business in English.

That last exclusion is measurable, and the measurements are worse than most people assume.

IrokoBench, a benchmark covering seventeen African languages published at a leading computational linguistics conference last year, found an average gap of around forty-five per cent between high-resource languages and African languages across every model tested. A separate benchmark found that one leading commercial model performs more than twenty-five points better in English than in African languages.

Then there is the cost. Large language models charge by the token, and African languages are tokenised far less efficiently than English. Research presented this year across sixteen African languages found that higher token counts predict lower accuracy, and that doubling the token count roughly quadruples cost and latency. A study published in June puts the median premium for African languages at nearly double English on the tokeniser behind the newest frontier models, rising to eight or nine times for some languages, which also shrinks the usable context window to a fraction of what an English speaker gets.

Put plainly, a Yoruba or Hausa speaker pays roughly twice as much per unit of meaning, waits longer, and receives a less accurate answer. If a bank, a hospital or a ministry deploys an AI assistant with English as the default and a per-token cost model, that penalty is passed straight to the person least able to absorb it.

We already know the fix works, because it has been tested. An evaluation of a phone-based information service among displaced people in Burkina Faso found that those using the local-language version improved their malaria knowledge more than those using the French one. Same channel, same content, different language, measurably different outcome.

*What it would take*

I want to be fair to the counter-argument, because it is a serious one. Urban smartphone ownership is rising fast. Nigerian data is cheap by regional standards. Building parallel non-digital channels costs money, requires translation and retranslation, and the evidence base for voice services is still thin in places. When we ran our own phone-based screening for a separate health programme, we had to retranslate the prompts in one state even after pre-testing them.

All true. None of it changes who is standing outside the building while we decide.

Four things would change the outcome.

Measure adoption at the edges rather than at the centre. A service that performs well among the already-connected has told you almost nothing. Publish usage among non-smartphone households, among women, and in the three major languages besides English.

Before any AI feature ships, name three specific users it quietly excludes. If a team cannot name them, the team has not finished designing. This takes an afternoon and it is the single cheapest intervention available.

Build the non-digital channel alongside the digital one rather than promising it for phase two. Phase two is where inclusion goes to die. Voice and USSD are not consolation prizes; for a hundred and forty million people they are the only door.

And treat language as infrastructure, not as localisation. Nigeria’s own multilingual model, N-ATLAS, covering Yoruba, Hausa, Igbo and Nigerian-accented English, exists precisely because the frontier models handle our languages badly. Public procurement should require that any AI system bought with public money is evaluated in the languages the public actually speaks, with the cost penalty disclosed.

We have a habit in this country of measuring digital transformation by counting the people who were always going to be fine. It produces excellent dashboards. It also produces services that work beautifully for the thirty per cent and not at all for everyone else.

Talent is everywhere in Nigeria. Opportunity is not. Whether artificial intelligence widens that gap or closes it will not be decided by the models. It will be decided by who we remember to design for, and whether anyone is counting them.

Abisola Areola is a certified senior project manager who has delivered national-scale digital programmes in Nigeria. She is the founder of Skills Co-op, a cooperative social enterprise delivering free digital and AI skills training, and writes on AI, workforce intelligence, and the human layer that determines whether transformation works.