Writing

Building Adapted Intelligence

why we should exist.

Lattanye is at the outset, the very beginning, of a generational project.

The more familiar path for today's startups, in seeking the affection and capital of the Valley, has been to build enterprise or consumer products for the West. We declined to take this approach.

As a result, we continue to face firm questions about the potential profitability and, indeed more fundamentally, the wisdom of our approach. Others are sceptical of what some, not incorrectly, describe as a nonexistent market for artificial intelligence systems. A market with the fastest-growing population in the world. A continent where . And with or without invitation, the numbers are simply hard to ignore.

The question is not whether intelligence comes to this continent. It is who owns it when it arrives. I have no doubt that competent leaders of government and enterprise will, over time, choose sovereign AI over generalised LLMs. That, coupled with a large untapped African market, sets up Lattanye as the pioneer of something special. An understanding of where intelligence creates value, the limits of generalised models, and the future of adapted intelligence makes this outcome not merely possible, but inevitable.

I.

Consumer artificial intelligence is a race to serve the average human being, and the average human being is a statistical figure who resembles no African. That race is essentially over, and it was won by companies whose distribution, capital and models will not be matched on this continent for a generation. Subsequently, the current consumer product becomes a thin layer stretched over a borrowed model. Africa performs the work of acquiring the user, absorbs the cost, and keeps nothing. Every technology that reaches this continent leaves us poorer. . The others took a resource; this one takes knowledge. Ultimately, building another wrapper around a foreign model is a path we cannot take.

African consumers do not have the liberty to pay exorbitant subscription fees for unnecessary wrappers in the first place. The economics that built consumer technology in the West . So we are behind. Not by a little, and not in a way that catching up on that same path will fix. . We are spectators to the most important infrastructure of our lifetimes. Yet we are subtle enough students of history to understand what becomes of those who sit out. And although it has become unpopular to concede that we live in a continent of constraints, both ideological and economic, we must build intelligence for Africa as it is, not as we wish it might be.

II.

If consumer AI is the wrong race, the question becomes which race is the right one. It is anything but certain that our most compelling strides as a continent will be defined by our place in the AGI race. And whilst the development contest between frontier large language models may suggest so, it will not be the defining moment of our future. We will not win by building a bigger model; artificial intelligence will only redefine the world order through its impact beyond the bubble and . There is no question that this has quietly become as vital to our survival, perhaps even more so, as the vanishingly thin enterprise niches the frontier startups are scrambling to squeeze into. It has simply been less fashionable to say so.

This is the case we have been making for months, and for months our views about the ways in which African intelligence could be built and what it was capable of accomplishing were derided as fringe. We were routinely told that building intelligence for the African market, let alone governments, was the surest and quickest way to run a business into the ground. Many balked at supporting us. Many were sceptical, if not outright hostile. Others insisted we should focus on catering to the needs of the West. And many more simply would not return our calls.

We have repeatedly taken what seemed to many to be enormous risks to build this, risks that, in hindsight, look almost prudent and banal. It is because of, not in spite of, this ambition, this sense of purpose, and this desire to participate in a story larger than ourselves, that we have arrived at this moment. We have grander and more idiosyncratic aims than the returns that early sceptics measure us against; popularity will never be the ultimate measure of the value, broadly defined, of what we are building.

The public understands what the sceptics did not: that impact beyond the average person is what matters, not the perception of results or the appearance of forward movement. The discussion seminar of African development, decades of conferences, frameworks and pledges, has been exposed as such to many, and the people have grown uninterested in the theatricality of the contemporary political stage. Yet it is that destructive culture that creates the gap that lets companies like ours come into existence and excel, with renewed understanding that we will be judged by what we deliver, not what we promise.

The playbooks of the past are no longer working, and even the people who had believed in them now know this. No institution will survive resistance to AI; participation is non-negotiable. As a result, the rush toward large language models, including by African institutions, has turned into a stampede. But that is not the right move, for them or their people; it is the same cycle again. This stampede, driven by a lack of nerve to build their own intelligence, to fully expose themselves to scrutiny, and to transform the institutions they run, will have unimaginable consequences.

III.

The reason intelligent public systems, not consumer apps, are the real prize for Africans comes down to what a model is and is not. The language models that have transfixed the world are incredibly capable probabilistic engines, excellent at discrete tasks, with an ability to approximate truth. They know the internet: pretrained on widely available knowledge, then post-trained on the surface of a wide variety of tasks across multiple domains. Until recently that was enough, because the moat was the model. Whoever trained the largest one owned the most capable system. That era is closing. Open-weight models now sit close enough to the closed frontier that raw capability is no longer the constraint. Behaviour is.

Behaviour is where these models fall short. They have been born into this world without any sense of its specific contours or logic, or indeed a conception of truth or basic facts, let alone the insightful knowledge of the operations of sensitive, . So models with trillions of parameters may increasingly mimic parts of a workflow, but without more, add little value to the general public. They have little innate sense of north or south, struggle on their own to solve complex analytical problems, and do not knit together the complex moving parts of an actual operation, the execution that runs across a whole institution rather than sitting inside one box.

The model alone is not that system. Such models are necessary but not sufficient. They will only be capable of truly transforming lives if they are adapted to a specific institution, its tangle of datasets, operations and personnel, and given a means of reliably and efficiently interacting with the byzantine complexity of the modern state, its idiosyncratic objects, logic, and physics. For that, two things are needed.

The first is structure. The strings of text produced by the language models are little without an architecture that can lend grammar, structure, and an action layer to the output of these probabilistic prediction engines. Look at any government and you find why. Dozens of systems, each built for its own narrow purpose, each holding its own silo of data, and the greater truth lost in the seams between them. Historically, we have relied on humans to connect those dots in their heads, and no human holds a whole state in his head. Point an LLM at hundreds of disconnected, ungoverned databases and you get a system that . For something as consequential as a nation's money, or its medicine, or its security, that is not just useless. It is dangerous.

So the model must be tethered to objects in the real world, and it is that tether that we have built: agents that watch, reason and act. They watch the signals an institution produces as they happen, reason across its departments rather than inside one, and act where they are permitted to, knitting together the complex moving parts of an actual operation. They are sovereign, built from a specific institution's objects, logic and reality. In a treasury that is Owo, our financial intelligence OS, holding a payment before it clears and handing the decision to the officer whose decision it is. This, no foreign model has ever seen.

The second is the model itself. A tether orients a generic model to our reality; it does not change what the model is. Everyone now talks about context, about feeding the model the right information at the right moment. But context supplied at the moment of execution is a thin layer over a generalised model; you cannot supply your way out of chaos. A model learns an institution in post-training, and almost everything it does after pretraining is set by the rewards it was trained on. The labs train where outcomes grade themselves: code passes its tests, a game ends. Critical institutions do not grade themselves. Whether a tender was rigged may be settled by an auditor eight months later, or never. Building those rewards honestly, in that world, is slow, specific and unglamorous work.

Until now, no model has been rewarded on the messy, unstructured data of critical African governments and enterprises, because those rewards do not exist anywhere. They cannot be scraped. They live in the records of the institutions themselves, in the judgement of their personnel, and in everyday outcomes. We are going to manufacture those rewards, then post-train an open-weight model on that work: the institution's own records, its worked cases, the decisions of the auditors, clinicians and engineers who run it. The result is not a foreign model with a local layer. It is the institution's own intelligence.

A government cannot borrow either from a lab abroad. The lab does not have the structure and cannot train the model, because neither can be built from the outside. What this requires, the institution's records and the judgement of the people who run it, should never leave the institution. As a result, the terrain for every institution that seeks to survive has shifted. Our institutions need the right intelligence stack, the underlying technical architecture that is the precondition for making artificial intelligence work for their people at all: a model trained on their own work, autonomous agents built on it, and more generally the handoff between humans and these algorithmic agents. Only the institutions that build it will survive.

IV.

Everything we build answers to one person: not the average user the frontier optimises for, but the citizen those systems were never designed to serve, the edge case by the measure of a model trained on someone else's world. Lattanye makes possible quantifiable and transformative results for that human. Our intelligence stack is built from reality, and there is, here, a dialectic between ground truth, tribal knowledge, and that edge-case person. The institutions that we serve, representing the durable, load-bearing infrastructure of our society, grasp this. They demand nothing less; they know society can afford nothing less.

We have therefore taken a different path, building a juggernaut of a company that will deliver intelligence to that person. Lattanye is in a category of its own. In the end, the route for us to something impactful, adapted intelligence, was made possible by an obsessive focus on the specific, particularised and quite idiosyncratic challenges and technical problems of our disparate set of customers, African governments.

That changes what Lattanye is at the technical level. Our asset is the environments, graders and corrections that exist only inside the institutions we serve, and the agents we build on them to watch, reason and act. An environment is where a model practises before it is trusted: a frozen copy of a real institution at a point in time, run through the exact tools the production agent uses, so that what it learns in training holds in production, and so that no lab abroad can build it.

Adapted intelligence is everything between an open checkpoint and an agent an institution trusts to act, and only some of it is weights. The base is replaceable: we start from open-weight models, and when a better one ships we retrain onto it. Facts live in the institution's own record, not in the weights; they change daily and need provenance, so the agent reaches them through tools and cites them, and the weights hold what does not change daily: how to investigate, what counts as evidence, when to stop, when to escalate. Behaviour comes from post-training on worked cases written by the institution's best professionals, on every approval and correction a reviewer makes, and in environments built from the institution's own history. What ships is small, and it belongs to the institution.

Every correction in production becomes training data. A flag an auditor overturns, a fault an engineer reclassifies, a payment a reviewer releases: each is a correction with its rationale and evidence, and each is the next release's training data. No release reaches production without beating the last one on a test it has never seen, passing a regression set built from every past failure, and running in shadow beside the live model until it matches or beats it on live traffic. The data outlives the weights. When a better base appears, every model is retrained from that store. The store is the asset, not the checkpoint.

The corrections are made by people. The auditors, clinicians and engineers who review the agents' work are not a cost of running the system; they are the source of its training data, and they are paid for the corrections they make. That workforce is African, it works inside the institutions it serves, and it is the part of this that no lab can hire from outside.

Our position compounds. The longer we operate, the better our agents become, and the harder they are to replace. The labs will keep releasing better open models, and every one of them makes us stronger: each improvement lifts every model we retrain onto it, while the environments, the graders and the corrections remain ours. What they cannot release is the reward for reconciling a payroll that has never been reconciled, for finding the ore that disappears between pit and port, or for holding back an accusation the evidence does not carry.

There is really nobody making the case that the African market should have second-rate AI capabilities or lacklustre software. If we are going to have agents operating within consequential public systems on behalf of our governments and people, they deserve the best. We stand on the walls, sentinels of the inner sanctum, against the assault of AI slop. And it has been our essentially aesthetic instincts, our willingness to treat taste as a discipline rather than an afterthought, that have shaped how this intelligence is built.

Some on the outside would find it striking how we continue to build this company. We are less a conventional startup than an engineering commune pointed at a single conviction. A remarkable amount of delusion and grandiosity is permitted, even encouraged, because the work is too hard and the stakes too high for deference. The raw power of authority is wielded sparingly. It is that tolerance for dissent, for ambition, for rogue ideas about what a government could be, that produced the architecture in the first place.

This is at once a hard-headed and a romantic enterprise. We will get there only through a fierce pragmatism, a nearly ruthless focus on real-world impact at the expense of the performative and the theatrical. What we are building is, perhaps more importantly, a business built around solving problems, not around legions of well-credentialed consultants with their presentations and wise counsel benefitting from the problem continuing to exist. And yet we have arrived at this conviction as much through a fierce idealism, a refusal to accept the existing limits on what intelligence can do for low- and middle-income countries.

V.

To build this, we first have to believe we are allowed to. We have, as a continent, and perhaps as a civilisation, grown uncertain of what we believe and far too fearful of embracing any degree of specificity of our own, deferring instead to systems and standards built elsewhere, for someone else.

But what exactly are we deferring to? A reticence, or perhaps an incapacity, to pronounce and to prefer, beyond the shallow and ritualistic deference to whatever was built abroad, has had costs. A tolerance of everything, a shallow embrace of all systems and standards as equally valid for us, devolves in the end into a belief in nothing of our own.

Our governing class is not merely adrift, as many have claimed, but has lost a sense of confidence, self-possession, and internal resolve. It is not quite clear what much of it even holds dear, beyond its own continuation in office. And so the question presses: is the African state nothing more than an economic zone, an empty administrative vessel, through which a self-imposed elite and its foreign vendors travel and enrich themselves? We have, in too many capitals, abandoned the work of deciding what we are for, and we are in retreat, so fearful of common endeavour and identity that all that remains is the management of appearances. It is that instinct, to prize the management of our own image above the building of anything real, that makes possible the borrowed system, the imported dashboard, and the fabricated sense of progress.

We should, and indeed must, return to a shared continental experience, an embrace of common identity that by definition puts forward certain ideas, values, and ways of building at the exclusion of others. It is and was a mistake to casually proclaim that every imported system serves us as well as one built for our own reality. Some tools are wondrous and generative. Others are destructive and deeply regressive when dropped onto ground they were never designed for. And the greater mistake still is to believe we must adopt the world's defaults to be taken seriously.

And yet we have, as a culture, resisted, some actively and strenuously, the degree of conviction in our own reality that would make such building possible. Adapted intelligence is, by definition, particular. It puts forward a specific reality, our objects, our logic, our priorities, and necessarily rejects the generic. The same instinct that makes us hesitant to assert who we are is the instinct that would have us rent our intelligence rather than own it. It is that pursuit of something of our own, and the rejection of a borrowed and generic default, that will ensure our continued strength and survival.

VI.

This is no longer a question of efficiency alone. AI is becoming infrastructure of extraordinary strategic consequence, increasingly embedded in systems of finance, energy, and defence, making it imperative that governments begin to treat it as such. The choice is not whether to let this technology into the machinery of the state. It is already coming. The world is in the midst of an AI revolution that is reshaping industries and economies, and the choice that remains is .

A gulf is emerging between those who perceive these preconditions for harnessing the power of artificial intelligence and others who are far more vulnerable and adrift. The world of AI adoption is increasingly divided between haves and have-nots. Some nations are thriving; others have been left naked on the beach. Africa does not get to assume which side it lands on.

The way onto the right side of that gulf is plain. It should be uncontroversial that the single most effective means of getting artificial intelligence to work for these economies is to make everyday public systems intelligent. Governments cannot continue to hide behind a purely oppositional identity to avoid the revolution and, now, the efficiency that intelligent public systems make possible. And that is the system, singular and without peer, we believe we are building.

The rewards from a real embrace of AI here will be immense, both in terms of human liberation and economic prosperity. The construction of such a platform, one that reflects our commitment to risk-taking and innovative ambitions, should be a rallying cry for progressives and critical thinkers across the political spectrum who profess to be interested in advancing human interests. And yet many take the easier path, blithely dismissing technology as the enemy and thereby deprive themselves of any role in crafting our most capable weapon.

Indeed, if we fail to act, to formulate an affirmative view of what technology can accomplish, the undirected logic of the market will decide for us. It will fill the chasm that has opened as a result of our contemporary skittishness to define who we are as a people and what we aspire to create.

VII.

The next chapter belongs not to the critics and bystanders, those at the margins of creative empires, but to those who build them. The future belongs not to those who style themselves as creators, or even to the merely brave, but to those who actually build something substantial and lasting. The sceptics and cynics, those who doubt without putting anything at risk, and those who deride the work of others without even feigning to construct something of their own, will be swept aside.

It is a mistake to think of the act of creation, of building something from nothing, as belonging only to the West. Far from it. Young Nigerian, Kenyan, and Rwandan coders are creators in the thickest sense of the word, and we should celebrate the risks they take and the things they build. And yet the chattering class, the merely critical, have lulled this continent into a kind of sleep, into believing that commentary and rhetoric alone, without any corollary or yardstick in the physical world, are sufficient. It is an uncomfortable fact that those most bearish on their own ability to 'develop' often turn instead to organising, to position, to the politics of the conference room. We have declined to bend our way of working to a world in which success is whatever passes for success. And it would be prudent to prepare for a world in which that critical class finds its power over the culture diminished and in retreat.

There will be incumbents, the consultancies and the vendors of expensive nothing, who would prefer we did not exist, and there will be officials who pray for our failure because our success exposes theirs. The temptation is to cast them as caricatures, to refuse them any strength or any humanity, and that temptation is a trap. A maniacal obsession with one's own virtue, and the insidiousness of one's enemies, is blinding. The most withering portrait of an opponent is the one that finds the drop of humanity in him, because it is the one that understands him well enough to defeat him. A degree of grace toward those who have wronged us, even those who would pray for our demise, is not weakness. It is the lethal advantage of seeing clearly while they do not.

We are still at the very start of things. But we are already far behind, and that gap is only growing. With continued execution, a focus on what matters, and a near-complete disinterest in what does not, we believe that adapted intelligence will be the dominant model of the future. And the market is now waking up to this reality. We have already built it.

Tatenda Ncube-Muchandibaya

Founder and CEO