The Institutional Gap Is the Feature

In discussions of AI adoption in African institutional contexts, the most common framing is one of disadvantage. African organisations face constraints — resource limitations, connectivity gaps, infrastructure deficits, shallow technical talent pools — that put them behind their counterparts in better-resourced environments. AI will be harder to adopt here. The benefits will arrive later. The gap between Africa and more technologically advanced economies will, in the short term, widen.

I think this framing is wrong. Not because the constraints are not real — they are — but because it misreads what those constraints produce. The environments most likely to build genuinely AI-complete workflows are not the best-resourced environments, where the option of adding staff to manage a process is always available and often cheaper than redesigning it. They are the environments where the gap between what needs to be done and the resources available to do it with is so wide that the process redesign question cannot be avoided.

The institutional gap is not the barrier to AI transformation. For the specific category of AI transformation that matters — the one that eliminates processes rather than optimises them — the institutional gap is the forcing function.

Where elimination is optional, it stays partial. Where it's forced, it finishes.

1. A mechanism claim, not a measured outcome, and what would undo it

This paper argues a mechanism, not a measured outcome — the distinction matters before the argument starts.

What it argues from. Four specific product case studies — PBOMaster, AccessWASH, ProcureBee, AnswerTab — described from the inside, by the company that built them. A peer-reviewed finding from innovation-management research that resource constraints have been shown to spur radical rather than incremental innovation in a different industry (manufacturing) and a different geography (not Africa-specific).¹ A well-known empirical result (M-Pesa) about what infrastructure scarcity produced in Kenya specifically.² A theoretical framework, from labour economics, for exactly what "eliminates a process" means and what can go wrong with it.³

What it does not argue from. It does not argue from a controlled comparison of AI adoption outcomes in resource-constrained versus well-resourced organisations — no such study is cited here because I did not find one that tests this specific mechanism. It does not argue that four products built by one company constitute a representative sample of how AI adoption proceeds across resource-constrained institutions generally. The Kenyan cases are illustrations of the mechanism, not a test of it.

What it does not claim. It does not claim constrained environments are better-positioned to adopt AI in general — only that, for the narrower category of AI deployment that requires eliminating a process rather than speeding it up, the absence of a staffing alternative removes an option that well-resourced organisations retain and, this paper argues, tend to exercise. §6 states directly that a serious, credentialed critic of African AI-leapfrogging optimism disputes a closely related claim, and engages her argument rather than ignoring it.

What would refute it. Given in full in §9. The sharpest version: if resource-constrained organisations are observed, at scale, defaulting to cheaper human labour over AI-complete redesign wherever that option exists at all — informal labour markets making staffing cheaper in practice than the capital and risk cost of redesign — the central mechanism this paper proposes (staffing being "not financially viable" forces redesign) would be empirically backwards, and the paper's argument would need withdrawing rather than softening.

2. The Resource Optionality Problem

There is a pattern in how well-resourced organisations respond to AI capability. When a new AI tool becomes available that could, in principle, eliminate the need for a category of human work, the organisation's default response is to treat it as an option rather than an imperative. The tool can be deployed to make the existing staff more productive. The existing process can be accelerated. The option of redesigning the process from the ground up exists but is not forced.

This optionality produces a particular kind of AI adoption: widespread deployment of AI as an improvement to existing processes, minimal deployment of AI as a replacement for existing processes. The cost of maintaining the existing process — the staff time, the management overhead, the error rate — is visible and manageable. The cost of redesigning it — the uncertainty, the implementation complexity, the change management required — is also visible and daunting. When the existing process is affordable, the default is to improve it, not redesign it.

In organisations where the existing process is not affordable — where the choice is not between "improve the process" and "redesign the process" but between "redesign the process" and "the process does not happen" — the optionality disappears. The process must either be designed to run without the resources that are not available, or it must not run. This is not a comfortable position to be in. But it is a position that forces the process-elimination question that optionality allows to be deferred indefinitely.

This pattern is not specific to AI, or to Africa. A 2013 study of resource constraints and radical innovation in the manufacturing sector, using longitudinal data from a large innovation survey and controlling for absorptive capacity, firm age and firm growth, found "full support for the hypothesis that knowledge constraints spur radical innovation and partial support for the hypothesis that financial constraints spur radical innovation" — constraints promoting, rather than impeding, the kind of innovation that does not just improve an existing process but replaces it.¹ That study is about manufacturing firms generally, not AI, and not Africa specifically; it establishes the general mechanism — constraint removing the option to merely optimise — not direct evidence for the AI-in-Africa case this paper makes. The two are related claims, not the same claim, and §6 takes up directly the question of whether the general mechanism actually transfers to this specific domain.

3. The Kenyan Cases

PBOMaster was built for an environment where the compliance lawyer is not a standard part of a community-based organisation's operating infrastructure. The choice was not between "organisations hire lawyers" and "organisations use PBOMaster." The choice was between "most organisations remain non-compliant" and "build the software that closes the gap." The gap was not a constraint on what could be built. It was the argument for building it.

AccessWASH was built for water utilities operating in counties where the data analyst, the GIS specialist, and the reporting officer are positions that exist on the organisational chart but are unfilled because the budget does not support them. The choice was not between "utilities hire analysts" and "utilities use AccessWASH." The choice was between "utilities continue operating in operational blindness" and "build the platform that provides visibility without requiring the staffing it would require in a better-resourced environment." The operational data infrastructure that well-resourced utilities have through staff was built, in the AccessWASH context, through software that works with the staff that is actually present.⁴

M-Pesa is the most cited example of this pattern, and its lessons have been so thoroughly absorbed into development economics discourse that they risk being treated as historical rather than instructive. A 2016 study in Science, using panel data tracking Kenyan households over six years, found that access to M-Pesa agents increased per-capita consumption and lifted an estimated 194,000 households — about 2% of Kenyan households — out of extreme poverty, with effects concentrated among female-headed households and driven by improved financial resilience (the ability to absorb a shock without cutting consumption) and increased savings.² That is a measured outcome, not a mechanism claim — the study does not argue that Kenya's banking-infrastructure scarcity forced a better design than a well-resourced alternative would have produced; it measures what the mobile-money product that scarcity gave rise to actually did for the households that adopted it. The lesson of M-Pesa — that the constraint of an under-developed banking infrastructure produced a categorically different infrastructure, not a delayed version of the one that existed elsewhere — is an interpretation of that history, documented more broadly in the economic-history literature on technology adoption and resistance, which finds that incumbents with an existing infrastructure to protect are frequently the ones most resistant to a redesign that would obsolete it.⁵ It applies directly to the AI workflow argument. It is being underused.

4. What the AI version of this looks like

The argument I am making is an extension of the leapfrog thesis, but with a specific mechanism: the institutional gap is the forcing function for AI-complete workflow design, because AI-complete workflows are built in environments where the alternative — staffing the workflow — is not financially viable.

In the compliance infrastructure domain, this is already visible. The PBO Act's requirements cannot be met by most Kenyan CBOs without either legal assistance or a system that replaces legal assistance for the compliance-checking tasks that are amenable to automation. PBOMaster was built because the gap demanded it. The product that was built is, arguably, better than the alternative it replaced: more systematic, more consistent, more accessible to rural organisations than a Nairobi legal market would ever be.

In the procurement domain, ProcureBee was built for companies operating in markets where the finance officer, the procurement manager, and the compliance reviewer are often the same person — or the same person wearing three hats. The agentic procurement workflow that eliminates the approval chain was not built as an improvement on an existing process that was working. It was built for contexts where the existing process was either absent or so burdensome that it was producing non-compliance rather than compliance.

In the knowledge infrastructure domain — the work AnswerTab is doing — the gap is between the volume of information that organisations need to process and the reading capacity available to process it. In organisations with large research teams and dedicated analysts, AI-assisted reading is an improvement on what already exists. In organisations with one programme officer covering three thematic areas and managing four active grants, AI-complete reading is not an improvement. It is the difference between the information being processed and not being processed.

5. The Implication for Where AI Development Happens

The global AI investment conversation is overwhelmingly focused on the highest-resource environments: Silicon Valley, London, Beijing. This is not surprising. The frontier AI models are being built where the frontier AI researchers are, which is where they can be paid the most, which is where the most capital is.

But the frontier of AI deployment — the edge of what it means to build software that genuinely replaces human processes rather than augmenting them — is not the same geography. It is the geography where the process-elimination question is forced rather than optional. Where the alternative to AI-complete workflows is not a less efficient human process. It is no process.

I am not making a prediction about where AI frontier research will be done. I am making a claim about where the most consequential AI deployments will come from — the ones that change institutional structures rather than making existing ones faster. Those deployments will come disproportionately from environments where the institutional gap makes the process-redesign question non-optional.

Nairobi is one of those environments. Not because of any technological advantage, but because of a specific combination: a sophisticated enough technical infrastructure to build and deploy serious software, and institutional gaps wide enough to make the process-elimination question unavoidable. The compliance officer that a Kenyan CBO cannot afford to hire is the argument for PBOMaster. The procurement team that a Kenyan startup cannot staff is the argument for ProcureBee. The reading capacity that a Kenyan programme officer doesn't have is the argument for AnswerTab.

6. Objections I take seriously

"A credentialed, Africa-based technology-policy researcher has argued the opposite of this paper's core premise." This is the most serious objection and I want to engage it directly rather than cite it in passing. Rose Mutiso — research director at the Energy for Growth Hub, writing on African technology policy for over a decade — published a reflection in Science in 2025 explicitly warning against treating leapfrogging as a default expectation, arguing that a technology only qualifies as a viable leapfrog if it is "already proven in mature markets," offers "comparable functionality," and genuinely "enables workarounds to expensive or obsolete infrastructure" — not merely because resources are scarce.⁶ In a separate piece specifically about AI, she goes further: "there is simply no 'leapfrogging'" around the foundational prerequisites — reliable power, connectivity, usable data — that AI integration actually requires, and she argues governments should act as "enablers, rather than orchestrators" rather than pursue sweeping digital-transformation strategies.⁷ This is a direct challenge to this paper's optimism, from someone whose professional focus is exactly this question, and it deserves an honest answer rather than a dismissal.

My answer is narrower than a rebuttal. Mutiso's argument is about infrastructure-dependent leapfrogging — mobile money, grid electricity substitutes, large-model AI that needs power and connectivity at scale — where she is right that scarcity of the prerequisite resource does not somehow conjure the resource into existence; you cannot leapfrog around needing electricity. This paper's claim is narrower and does not depend on bypassing that kind of infrastructure constraint: PBOMaster, AccessWASH, ProcureBee and AnswerTab are process-redesign tools, not infrastructure substitutes, and the specific resource their absence forces redesign around is staff, not power or bandwidth. Mutiso's caution is correct and important for the infrastructure-leapfrogging claims she is responding to; it does not, on its own terms, refute a claim about process-elimination AI specifically. But her broader point — that resource scarcity is not automatically generative, and treating it as such is a pattern worth being skeptical of — is one this paper should take as a standing caution rather than a settled rebuttal, and I do not have a controlled comparison that would settle which of us is right at the level of outcomes rather than mechanism.

"Eliminating a process with AI is not costless even when staffing is unaffordable — it can simply mean the process does not happen well, rather than happening differently." Acemoglu and Restrepo's labour-economics framework distinguishes a displacement effect, where automation takes over a task previously done by labour, from a reinstatement effect, where new tasks emerge that create demand for labour elsewhere; they find that recent decades of slower employment growth in the US are explained by an accelerating displacement effect paired with a weakening reinstatement effect.³ Applied here: an AI system that eliminates the compliance-review process at a CBO that could never have afforded a compliance officer is pure reinstatement from the CBO's point of view — it creates capability that did not exist, rather than displacing an existing worker — but if "forced AI-complete redesign" becomes the default response to any staffing gap, including gaps that would have been better filled by a trained person once resources allowed, the same framework warns that built-in elimination can also foreclose a future reinstatement effect that hiring would have produced. This paper does not have evidence distinguishing these two cases at scale; it is a real, underexamined risk the enthusiasm in §4 does not fully address.

"Four case studies built by the company making this argument is a conflict of interest, not evidence." Flux built PBOMaster, AccessWASH, ProcureBee and AnswerTab, and has a direct commercial interest in the thesis that resource-constrained AI deployment produces better, more complete products than well-resourced alternatives — that is close to Flux's own sales pitch. This is a real conflict and I am not going to pretend the case studies are a neutral sample; they were selected because they are the clearest available illustrations, from the only pipeline of products I have direct access to describe from the inside. The honest mitigation is narrow: the mechanism argument in §2 does not depend on these four products being representative — it depends on the Keupp and Gassmann finding in manufacturing and the M-Pesa history in Kenyan financial services, both independent of Flux, as evidence that the general mechanism (constraint removing an optionality that otherwise gets exercised) operates outside Flux's own product line. The four case studies illustrate the mechanism; they do not establish it.

"The leapfrog thesis itself has a decades-long history of being used to oversell Africa-specific technology narratives, and this paper risks repeating that pattern with 'AI-complete workflow' substituted for the previous decade's buzzword." I think this is a fair historical pattern to be wary of, and I do not have a strong defence beyond naming the four falsification conditions in §9 as genuine commitments rather than rhetorical gestures. If the pattern this paper describes turns out not to hold up against those tests, the right response is to say so, not to find a softer version of the same claim.

7. The Broader Argument

I want to end with the claim I began with, stated precisely, and bounded by what the preceding sections actually established rather than restated at its most ambitious.

The most constrained environments are plausible — not proven — fertile ground for AI-complete workflow design, because the process-elimination question is forced rather than optional where staffing a process is not financially viable. When staffing the process is not viable, the design question is not "how can AI help our people do this?" It is "how can we build a process that does not require our people to do this?"

The solutions produced in constrained environments are, on the evidence of the four cases described here, structurally simpler, because they are designed without the assumption of the supporting infrastructure that well-resourced environments take for granted. They work with the organisation that is actually present, not the one that should be present. Whether that pattern holds beyond these four cases, and beyond one company's product line, is exactly the open question §9 states as a falsification condition rather than an assumed yes.

And the market this argument describes is larger than the institutional AI conversation currently assumes, if the mechanism is real. The population of organisations that cannot afford the staffing alternative is not a niche. The compliance infrastructure gap described in Kenya is the same gap, in different legal clothing, in every African country. The procurement workflow problem is the same problem for every under-resourced organisation that has a spending policy and no mechanism to enforce it except human attention. But this paper has not measured that population, and Appendix B registers that absence directly rather than assuming the scale claim follows from the mechanism claim.

The companies that understand this — that the institutional gap is a design brief, not only a constraint — may build some of the most consequential AI products of the next decade, in Nairobi, Lagos, Kampala, Accra, not because those cities have an advantage over San Francisco in AI research, but because they have a specific kind of problem San Francisco mostly does not: the alternative to building software that works without the staff is, in a meaningful number of cases, no software that works at all.

That is not, on this paper's argument, simply a disadvantage. But whether it is a reliable enough pattern to call a feature, at the scale this paper's closing paragraphs imply, is the thing §9's falsification conditions are actually testing.

8. The limits of a mechanism argument

No controlled comparison between resource-constrained and well-resourced organisations' AI adoption outcomes is cited above, because none was found that tests this specific mechanism.

It does not claim the four Kenyan case studies are a representative sample of how AI adoption proceeds in resource-constrained institutions; §6 states plainly that they were selected as the clearest available illustrations from one company's own product line, not sampled.

It does not claim Rose Mutiso's scepticism about African AI-leapfrogging is wrong; §6 argues this paper's narrower process-elimination claim is not directly refuted by her infrastructure-focused argument, which is a different and more limited defence than disproving her position.

It does not claim that AI-driven process elimination is costless or risk-free even where staffing genuinely is not viable; §6's second objection, grounded in the displacement/reinstatement framework, names a specific, unresolved risk.

It does not claim to have measured the size of the market this thesis implies exists; Appendix B registers this as an open question rather than an assumed scale.

And it does not claim this is a general theory of African technology adoption; it is a claim about one category — AI deployment that eliminates rather than augments a process — under one specific condition — staffing the process is not financially viable — and it does not extend further than that without more evidence than this paper has.

9. What would falsify this

Stated as commitments, in decreasing order of how much damage each would do to this paper's claims.

Resource-constrained organisations are observed, at a scale larger than four case studies, defaulting to cheap informal human labour over AI-complete redesign wherever that labour is available at all — i.e., the "staffing is not financially viable" premise is empirically false for a meaningful share of the organisations this paper's argument is about. That would invert the paper's central mechanism, and the right response would be withdrawal, not qualification.

A controlled or large-sample comparison finds AI adoption in well-resourced organisations produces process-elimination outcomes at a rate equal to or exceeding resource-constrained organisations, once controlling for sector and problem type. That would falsify the Resource Optionality Problem in §2 as stated — the claim that optionality specifically suppresses elimination-oriented AI adoption in well-resourced settings.

Mutiso's infrastructure-prerequisite argument (§6) is shown to apply to process-elimination AI as directly as it applies to infrastructure-substitution leapfrogging — for instance, if PBOMaster, AccessWASH, ProcureBee or AnswerTab are found to depend on power or connectivity reliability that the organisations they serve do not actually have, undermining the "staff is the binding constraint, not infrastructure" distinction this paper's rebuttal in §6 rests on. That would collapse this paper's narrower defence into the broader critique it is trying to distinguish itself from.

A longitudinal study of AI-complete workflows built under resource constraint finds a measurable displacement effect with no offsetting reinstatement — people who would otherwise have been hired as the organisation grew are never hired, and no new category of work is created in their place — at a rate that looks like net harm rather than net capability creation. That would substantiate §6's second objection as the dominant case rather than an unresolved risk, and would require this paper's closing optimism to be walked back substantially.

Appendix A — Evidence table

Appendix B — Citation register

CITATION-NEEDED — scale of the addressable population. Appendix A row 7. §7's claim that this is "not a niche" is an assertion, not a measurement; no source quantifying the number of African organisations for which staffing is genuinely unaffordable (versus merely preferred not to pay for) was found.

CITATION-NEEDED — labour-substitution behaviour under constraint. Appendix A row 8. The paper's central mechanism assumes organisations facing a staffing gap turn to AI-complete redesign rather than informal cheap labour; no study directly testing this choice was found, and §9 names it as the sharpest available falsification test.

CITATION-NEEDED — displacement/reinstatement balance for constraint-driven AI. Appendix A row 9. Acemoglu and Restrepo's framework is general labour economics, not applied to this paper's specific case; whether constraint-driven AI deployment in African institutions nets out as capability creation or net job loss has not been measured.