There is a procurement officer I think about often. She works for a county water utility, and her title — procurement officer — flatters the job by about an order of magnitude. She is, in practice, the entire purchasing function of an institution that spends hundreds of millions of shillings a year. When a treatment plant needs a replacement pump, the request reaches her as a handwritten note. She raises a requisition, walks it to three managers for signatures, types up a tender notice, waits the statutory number of days, receives bids in sealed envelopes, convenes an evaluation committee, and — if nothing is contested, which is rare — issues a purchase order some eleven weeks after the pump first failed.
I asked her once what she would do with a second version of herself. She did not hesitate. She said she would put the clone on the routine purchases — the chemicals, the spare parts, the things the utility buys every month from the same handful of suppliers at prices that barely move — so that she could spend her own time on the contracts that actually carry risk. She did not say she wanted to be replaced. She said she wanted the parts of her job that are not really judgment to stop consuming the hours she needs for the parts that are.
That sentence is the whole thesis of this paper. The most valuable thing we can build for African institutions is not a better dashboard. It is the second version of her — a worker that does the part of the work that was never really work, so that the scarce human judgment in the building can be spent where it matters. We have started calling these workers AI employees. I want to be careful about that phrase, because almost everything currently said about AI employees is being said by people describing a different continent's problem.
What you imagine when you hear "AI employee"
In the dominant telling, an AI employee is a cost-removal device. A firm in San Francisco or London has a customer-support team of forty, a sales-development team of twenty, a finance team of fifteen. Each of those people costs a salary. The promise of the AI employee is that some fraction of them become unnecessary — that the org chart can be redrawn with software where headcount used to be. The framing is replacement, the metric is heads removed, and the moral temperature of the conversation is correspondingly anxious. When a Western commentator says "AI employee," the unstated second half of the sentence is "and what happens to the human who had that job."
This framing is not wrong for the context that produced it. Western firms are, by the standards of the rest of the world, almost grotesquely well-staffed. They spent five decades building out the back office — the ERP systems, the SaaS stack, the layers of analysts and coordinators and operations managers — that turns a business into a legible, governable machine. The AI employee arrives into a fully furnished house and asks which furniture can be removed.
The African institution is not a furnished house. In most of the buildings I have worked in, the furniture was never delivered.
The counterfactual test
Here is a single question that, asked honestly, separates the two continents' versions of this technology. When an AI employee takes on a task, what was happening to that task before?
In the Western firm, the answer is: a person was doing it. The AI employee's arrival displaces a salaried human, and the economic calculation is a comparison between two costs — the marginal cost of the software against the marginal cost of the worker it replaces. This is why the Western debate is fundamentally a labour debate.
In the African institution, the honest answer is almost always one of two things. Either the task was being done badly, by someone heroically improvising with a paper ledger and their own memory — the manager who needs twenty-two minutes to find yesterday's water-production figure, a scene I have described elsewhere as the signature of an institution that operates blind. Or — and this is the case that the Western framing has no category for — the task was not being done at all.
The constituency office that cannot tell a citizen whether their borehole project has been tendered is not running an inefficient project-tracking function. It is running no project-tracking function. The role does not exist. The sacco treasurer who cannot produce a same-day liquidity position is not staffing that analysis poorly; the analysis is simply never performed until the month closes. Across the institutions I have spent my career inside, the binding constraint is not that work is being done expensively. It is that enormous quantities of necessary work are not being done by anyone, because no one could be hired to do them and no system was ever built to do them instead.
This inverts the economics completely. The Western AI employee competes against a salary. The African AI employee competes against zero — against a task that was falling on the floor. Its counterfactual is not a displaced worker drawing a wage. Its counterfactual is the leak that was never located, the drug stock that was never counted, the supplier who overcharged for three years because no one was watching the routine line items. We are not removing labour from these institutions. We are introducing it for the first time.
The Western AI employee is a subtraction from a staffed organisation. The African AI employee is an addition to an understaffed one. Confusing the two is the single most common error in how this technology is discussed on the continent.
This is why I am impatient with the imported anxiety. African small and medium institutions are not nervously protecting incumbents from automation. They are hungry. They have been waiting two decades for a workforce they could never afford, and they are about to be able to afford it — not as humans on a payroll, but as roles instantiated in software, available at the marginal cost of compute.
From the operating layer to the operating workforce
I have argued before that the central failure of African institutions is a missing layer — the absence of the operational data infrastructure that would let an institution see its own functioning in real time. That argument was about visibility. A utility that cannot see its non-revenue water cannot manage it; a health facility that cannot see its stock cannot defend it from leakage. The first generation of the work my lab does is building that layer: giving institutions, for the first time, a continuous machine-readable record of what they are actually doing.
But visibility is necessary, not sufficient. Seeing the work is not the same as doing it. An institution can be given a perfect real-time picture of its procurement backlog and still drown in it, because the bottleneck was never the seeing — it was the doing, and there were never enough hands. The dashboard tells the procurement officer she has eleven routine purchases pending. It does not make them happen. The missing layer, fully built, produces an institution that can finally watch itself fail in high resolution.
The second generation — the one this paper is about — is the operating workforce. If the operating layer is the institution's nervous system, the operating workforce is its hands. These are not tools that make a human faster at a task. They are agents that own a task end to end: that receive the input, apply the institution's own rules, take the action, produce the record, and escalate only the genuine exceptions to a human. The progression is deliberate and ordered. First you give the institution eyes. Then you give it hands. The hands are only trustworthy because the eyes came first — an AI employee is safe to deploy precisely because every action it takes is logged into the same operational record that the operating layer made visible.
Let me be precise about what I mean by an AI employee, because the term is being stretched to cover things that are not it. A chatbot is not an AI employee; it answers, it does not act. A copilot is not an AI employee; it accelerates a human who remains the doer of the work — and I have written about why most of what is sold as AI transformation is exactly this: the acceleration of a process that should have been eliminated. A robotic-process-automation script is not an AI employee; it executes a fixed sequence and breaks the moment reality deviates from the recording.
An AI employee is defined by three properties together. It is role-scoped: it occupies a named function — the purchaser, the bookkeeper, the collections officer — rather than a generic capability. It is agentic: it takes real actions in the world, against real systems, with real consequences, not merely producing text for a human to act on. And it is accountable: every action it takes is bound to the institution's rules and written into an auditable record, so that the institution can answer, at any moment, what the AI employee did and why. Role-scoped, agentic, accountable. Remove any one of the three and you have something less — a feature, a toy, or a liability.
Why the first one is a buyer
We did not choose procurement as the first AI employee because it is easy. We chose it because it is the place where the African institution's dysfunction is most acute, most expensive, and most legible — which makes it the place where an honest AI employee can prove itself.
I have written a longer argument about why institutional software procurement in Kenya is structurally broken — how a legal framework that treats software as a specifiable commodity has produced a graveyard of failed national systems and, worse, a pattern of vendor capture in which a single supplier ends up holding operational control of infrastructure the public cannot do without. That essay was about how institutions buy the big systems. But the same epistemic failure plays out a hundred times smaller, every day, in the routine purchasing that never makes the newspapers: the eleven-week pump, the chemicals bought at last year's inflated price because no one re-tendered, the supplier whose invoices have crept upward unnoticed because the person who would notice does not have the hours.
Procurement is the right first hire for four reasons. It is universal — every institution does it, which means one well-built buyer serves a water utility, a sacco, a clinic, and a constituency office with the same core. It is rules-bound — purchasing runs on thresholds, approval ladders, and supplier lists, which is to say it runs on exactly the kind of explicit logic an accountable agent can be made to follow and to prove it followed. It is high-pain — it is where money leaks, and an AI employee that closes a leak pays for itself in a way no productivity tool can. And it is auditable — every action leaves a document, which means the institution can verify the AI employee's work against the same standard it would hold a human to.
There is a deeper reason, too, and it connects to something I have written about the buyer side of software. An agent that purchases does not experience the persuasion architecture that human buyers are subjected to — it does not anchor on a high number, feel a "most popular" badge, or accept an ambiguous seat definition; it reads price and terms as structured data and discards the theatre. An institutional buyer built as an AI employee is, for the first time, a purchaser that cannot be socially engineered, cannot be worn down by a vendor's retention script, and cannot quietly let a recurring charge drift upward. In a procurement culture where so much loss is the product of inattention and pressure, an agent's indifference to both is not a limitation. It is the feature.
This is what ProcureBee is: the first AI employee, scoped to the one role where African institutions lose the most and can verify the most. It is the low-hanging fruit. It is also, deliberately, the first of a long line.
The fleet
The reason to start with one well-chosen role is that the second role is cheaper than the first, and the tenth is cheaper than the second. The hard part is not the procurement logic; it is everything underneath it — the connectivity that drops, the staff who are not power users, the data that arrives as a photograph of a handwritten form, the requirement that every action survive a regulator's audit. That foundation, once built for the buyer, is the same foundation the next employee stands on. We are not building a product. We are building a workforce, one role at a time, on a shared substrate.
The line that follows the buyer is not speculative; it is dictated by the same question that justified the first hire — which necessary role is no one filling? The bookkeeper that reconciles the day's transactions against the ledger before the month can hide them. The collections officer that knows, today, which members are in arrears and acts before the exposure compounds. The compliance officer that prepares the regulatory return continuously instead of in a panic each quarter — the one role whose absence I watched a utility CEO describe when he admitted that the NRW figure he reports to the regulator is, honestly, an estimate. The monitoring-and-evaluation officer that a donor-funded programme is required to have and almost never can afford. The grants officer, the stock controller, the records clerk. Each of these is a role that the Western firm staffs as a matter of course and the African institution has simply done without — not from negligence, but because the salary was never in the budget and the system was never built.
Read that list again and notice what it is not. It is not a list of humans to be made redundant. There is no incumbent bookkeeper at the county water utility to displace; the books are reconciled, if at all, by the same overstretched accountant who is also doing four other jobs. The fleet does not arrive to thin a workforce. It arrives to assemble one that never existed.
The incumbents, and why no one built this
It is worth being honest about why this workforce is missing, because the answer indicts a specific status quo. The institutions are not served by no software. They are served, badly, by two kinds of incumbent.
The first is legacy monolithic software, often procured a decade ago through exactly the broken process I described, frequently controlled by a single vendor who has no incentive to improve it and every incentive to keep the institution dependent. This is the eCitizen pattern at small scale: systems that were sold as solutions and have become rents. They do not produce a workforce. They produce lock-in.
The second incumbent is more insidious because we do not usually call it software at all. It is people being used as software. The clerk whose entire job is to retype numbers from one form into another system is a human being performing an integration that should be a line of code. The committee that exists to manually route an approval is a workflow engine made of meetings. A vast amount of what looks like employment in African institutions is in fact people standing in for the automation that was never built — humans absorbing, with their hours and their patience, the absence of the operating layer. When I say an AI employee competes against zero, this is the nuance: it does not compete against the judgment of these people, which is real and scarce. It competes against the part of their day where they are being asked to act like a machine because no machine was provided.
No one built the alternative for reasons I have come to know well. Venture capital chases consumer markets with geometric growth, and a county utility's purchasing function is not that. Development agencies fund systems on grant cycles, specified by consultants who visited for two days and gone before the first staff change, so the software is abandoned the moment the project closes. Both failure modes share a root: a refusal to make the long, unglamorous, present commitment that institutional software requires. The work that matters here is translation — taking a genuine capability and fitting it to the actual connectivity, the actual staff, the actual procurement and financial structures of the institution — and translation cannot be done by a vendor who is not staying. It is the work neither the VC nor the donor was ever structured to do.
What this means for African institutions
If we build this well, three things follow, and they are larger than any single product.
The first is a leapfrog of the back office. Kenya already produced the canonical example of skipping a generation of infrastructure: M-Pesa did not build a faster bank branch, it built the payment system a branchless population actually needed. The operating workforce is the same move, aimed at a different layer. The West spent fifty years and untold billions assembling the human-and-software back office that makes an institution governable. African institutions are not going to retrace that path. They are going to skip it — to acquire, in a few years and at the marginal cost of software, the operational workforce it took the incumbents half a century to staff. The institutions that do this will not be slightly more efficient versions of their current selves. They will be a different kind of institution: one that can see itself and act on what it sees.
The second is sovereignty, and it is the reason this cannot be imported wholesale. An AI employee built in San Francisco for a San Francisco firm assumes reliable connectivity, clean structured data, English-fluent power users, and a payments rail that does not exist here. Ported over, it fails in exactly the ways the donor systems failed. But more than that: a workforce is too important to rent from a foreign monopoly that can raise the price or revoke access. The whole argument of the procurement essay was that institutions lose control when they let a single distant vendor hold the keys. To build the operating workforce African-institution-first — designed for the real context, owned and stewarded close to the ground — is not merely a technical preference. It is the difference between a continent that hires its new workforce and one that leases it back from the people who built it for someone else.
The third is a warning, and I would be dishonest to omit it. The danger is that we use this technology to do the wrong thing faster. An AI employee bolted onto a broken process does not fix the process; it industrialises it, and makes the dysfunction permanent by making it cheap. The county that automates its eleven-step approval ritual instead of asking why eleven steps exist has not hired a procurement officer — it has hired a faster bureaucrat. The entire value of an AI employee is that it forces the institution to make its rules explicit, and the moment the rules are explicit, most of them turn out to be indefensible. The right way to deploy this workforce is to let it eliminate the work that should never have existed, not to preserve that work in software. Visibility first, then honest hands. Build it the other way around and you have spent real money to encode a bad institution more durably than before.
The throughline
So here is the line that runs through everything, stated plainly.
African institutions fail not for lack of money but for lack of a layer — the operational nervous system that lets them see themselves. Build that layer, and they can finally watch their own work. But watching is not doing, and the doing was always the deeper shortage, because the workforce that does the routine, necessary, rules-bound labour of a functioning institution was never hired — not from negligence, but because it could not be afforded and was never built. The technology to instantiate that workforce in software now exists, and for the first time it can be built for the African institution rather than ported, broken, from somewhere it was never meant to fit. We are starting with the buyer, because procurement is where the loss is largest and the proof is clearest. The buyer is the first of a fleet — a line of role-scoped, agentic, accountable employees, each one a hire the institution could never make. Done with discipline, this does not thin these institutions. It staffs them, for the first time, with the workforce the rest of the world spent fifty years assembling and they are about to acquire in a fraction of that.
The second version of the procurement officer is coming. She asked for it herself. Our work is to make sure that when it arrives, it takes the machine-work off her desk and leaves the judgment — and that it was built, from the first line, for the institution she actually works in.
