Two weeks after we turned it on, I was staring at a session graph that was pointing the wrong way.
We had just shipped AI powered game recommendations on the trivia platform I look after at VDL Technologies. It was the kind of feature that reviews well in a planning meeting. Every time a player opened the app, we would look at what they had played, what they had won, what time of day it was, and serve them the game they were most likely to enjoy next. Smarter app. Happier players. Longer sessions.
Sessions got shorter.
Not by a little. Enough that I stopped what I was doing and went looking. The model was not wrong. Its recommendations were, by every offline measure we had, better than the static list they replaced. The problem was that most of our players never saw them. The call went out on app open, and on a mid tier Android phone on a congested mobile network in Lagos, the screen sat there and did nothing for several seconds. People do not wait several seconds to play a trivia game. They close the app.
We had built something intelligent that our own users could not reach.
The infrastructure gap is the product problem
Most of what gets written about AI product management assumes things that are simply not true here. It assumes cheap abundant compute. It assumes an API that answers quickly. It assumes users on fast connections with recent devices and no anxiety about what a feature costs them in data.
Nigeria is a different room. Data costs money that people feel. Device spread is enormous, so the same screen has to work on a sixty thousand naira Tecno and on an iPhone. Power is not guaranteed, which means a user might not get to finish what they started. None of this is an edge case. This is the median experience of the market.
So every AI decision carries weight it would not carry in San Francisco. A feature that makes three model calls per interaction might be brilliant and completely unusable for someone on a slow connection in Abuja. That is not a technical footnote. That is the product failing.
The best AI feature is the one your user can actually finish using.
We rebuilt the recommendation engine as a batch job. It ran overnight, computed recommendations for every active player, and served them as a small static payload the app could fetch instantly or fall back on if the network was bad. On paper it was a downgrade. Less real time, less responsive to what happened in the last five minutes, less clever.
Engagement recovered and then went past where it had been. The platform went on to reach around fifty thousand users. The lesson has stayed with me since: I had been optimising the model when the thing that needed optimising was the distance between the model and the person.
What actually gets used
The clearest version of this lesson came from healthcare, not gaming.
I worked on an electronic medical records platform that eventually ran across sixty hospitals. Going in, I had opinions about what intelligent features the system should have. Then I watched nurses use it. They were moving between patients with almost no slack in their day and even less patience for a system that wanted them to learn something new. Any feature that added a step, however smart that step was, would quietly not get used. Staff would work around it and the record would end up incomplete, which is worse than no feature at all.
What worked was small and unglamorous. Pulling patient details in from voice rather than making someone type them. Flagging a dosage that looked out of range. Surfacing the next likely action so the person did not have to hunt for it. None of that is technically impressive. All of it removed a step instead of adding one, and that is why it survived contact with a real ward.
That is the opportunity here, and I do not think enough people are naming it. Nigeria has more than two hundred million people, a great many of them navigating healthcare, tax, pensions and credit with almost no digital support. AI is not a luxury layer on top of a working system. In a lot of cases it is the only realistic way to give personalised help to people who will never be able to pay for a human advisor. The job is not building the most sophisticated model. It is building the most useful bridge between a capability and somebody's actual circumstances.
What the role really asks of you
The foundation is the same as anywhere. Research properly. Frame the problem honestly. Prioritise ruthlessly. Lead people who do not report to you. On top of that, working here adds a few things:
- Cost sensitivity down at the API level. Not because you are building the model, but because you are deciding whether a feature's unit economics survive contact with this market. Token costs, caching, batch versus real time. These are product decisions wearing engineering clothes.
- Offline as a first class state. Users will lose connection mid flow. If your AI feature has no answer for that, it fails loudly and publicly, and people remember.
- Trust as a metric, not a nice to have. Nigerian users have been burned enough by unreliable services and outright fraud that accuracy is not a performance number here, it is a trust number. A feature that is wrong even occasionally gets abandoned faster than a slower one that is always right.
- Language and context that match the room. Pidgin, Yoruba, Igbo, Hausa. A product that only works in formal English is not really a product for this country, and models trained mostly on Western English need deliberate work before they behave.
On certification, and what it is actually good for
I hold IBM's Enterprise AI certification. The most useful thing it gave me was not technical. It was a framework for assessing whether an organisation is ready for AI at all.
Most companies I talk to here are not. The data is messy or scattered. Nothing is labelled. There is no internal capability to judge whether a model's output is any good, which means nobody would catch it when the thing starts drifting. Selling them an AI roadmap in that state is not ambitious, it is negligent.
So a real part of this job is being willing to say: fix your data hygiene first. Build the manual process that generates clean training data as a by product. Come back to this in eighteen months. That is not a failed engagement. That is the work, and saying it costs you the quick win but buys you a client who trusts you.
Where I think this goes
The next few years of technology in this region will be decided by which products work out how to deliver intelligence to people who do not have the infrastructure that most AI products quietly assume. That is a design problem and a judgement problem more than a modelling one.
The product people who will matter in that window are the ones who can hold both things at once. Real understanding of what these systems can do, and real understanding of the person on the other end of the screen, including how much data they have left this month.
That is the job. That is the one I want.
If you are building in this space, or looking for a senior PM who has done it, let's talk.
