AI adoption is a behaviour problem
We keep improving what AI can do. I am increasingly interested in a different question: how does that capability become part of ordinary life without asking people to reorganize themselves around it?
The default explanation for slow AI adoption is that the technology is not capable enough yet. Better models will arrive, accuracy will improve, and everyone will eventually use them. I think that explanation is incomplete.
For a growing number of tasks, the capability already exists. The harder problem is behavioural, economic, and institutional. People do not wake up wanting to use AI. They want to finish work, understand something, make a decision, speak to someone they trust, or get through a difficult day. AI has to enter those activities without becoming another piece of work.
Products compete with muscle memory
Every new app asks a person to form a ritual: remember it exists, open it, learn its language, move context into it, judge its output, and then return to the original task. Even a powerful product can lose to a familiar, slightly worse way of doing things.
This is why adoption is not simply a product capability problem. It is a distribution and workflow problem. A new behaviour usually wins because the old one is painful enough to abandon, or because the new capability disappears into something people already do.
AI should be a layer, not a place
The most useful AI may not look like an AI product at all. It may sit inside the browser, operating system, hospital workflow, company process, or conversation where the work already happens. People should not have to become dedicated AI users. The environments and people they already rely on should become more capable.
Invisible during the work. Visible at the point of consequence.
That last part matters. Invisible AI should not mean unaccountable AI. When money moves, care changes, access is granted, or an important decision is made, people should be able to see what happened, what evidence was used, how confident the system was, and who remains responsible.
Efficiency is not the only thing people value
Many AI products are designed as if every step a person takes is friction. But some steps create trust. Comparing alternatives can be part of understanding a purchase. Asking a question can preserve a relationship. Exploring can be enjoyable. Human interaction can be the service, not a defect in it.
Good automation removes burden without erasing agency, evidence, exploration, or meaning. The better product question is not only, “Can this be automated?” It is also, “Which parts should disappear, and which parts does the person still want to experience or control?”
The real intelligence is appropriate assistance
Complete autonomy is not always the immediate goal. Some decisions can be automated. Some need approval. Some should produce a recommendation. Some require escalation to a person. Sometimes the smartest action is to stay quiet.
Before adding AI to a workflow, I find these questions more useful than asking what the model can do:
- 01What is the person actually trying to accomplish?
- 02Where does this already happen, and what behaviour is already natural?
- 03Which part requires judgment, and which part is merely repetitive?
- 04What evidence or control does the person need before trusting the result?
- 05Should the system act, ask, recommend, escalate, or stay quiet?
Layered models are one possible solution
One possible solution is a layered architecture. Small models could provide fast, private, offline, and context-aware assistance close to the person. Larger models could be called when a task needs deeper reasoning, broader knowledge, or more compute. People would still retain control, judgment, and responsibility.
This approach would not solve every adoption problem, and it is not the only architecture worth exploring. But it changes the trade-offs around cost, latency, privacy, and access. It could also make room for systems that understand local languages and context without sending every interaction somewhere far away.
Africa makes the problem impossible to ignore
An AI product designed around constant connectivity, expensive devices, a single formal language, individual subscriptions, and high institutional trust will not simply become African because it is made available here.
Useful AI has to fit the realities of devices, bandwidth, cost, code-switching, informal work, existing institutions, and the people who already carry trust in a community. In many cases, distribution may happen through a clinician, teacher, business owner, cooperative, or public service rather than through another consumer app.
I have been researching the intersection of AI and Nigerian healthcare. Through that work, I have learned that the opportunity is not to place a generic chatbot beside an existing system. Useful AI has to begin with how care is actually delivered, then become native to that environment while preserving clinical judgment and accountability.
Capability is only the beginning
Better models will matter. But capability alone does not create adoption. Useful systems will also need to fit existing behaviour, preserve meaningful control, and work within the economic and institutional conditions around them.
I am still testing this thesis. But I increasingly believe the winning AI products will not be the ones that demand the most attention. They will be the ones that become part of natural behaviour, improve judgment, and know when their involvement is useful.
One possible shape: small models for presence, large models for depth, and humans for control, judgment, and meaning.