AI Productivity in East Africa: What Evidence Fits?

 A productivity estimate is not a passport

Daron Acemoglu’s work on artificial intelligence is valuable partly because it is more cautious than the loudest claims about an approaching productivity boom. In his 2024 National Bureau of Economic Research paper, The Simple Macroeconomics of AI, Acemoglu estimates that task level cost savings from AI could produce a total factor productivity increase of no more than about 0.66 percent over ten years under the assumptions he examines. He also warns that early studies may overstate future gains because they often focus on tasks that are easier for machines to learn and evaluate.

That argument should not be turned into another universal number. The legal and policy question for East Africa is whether an estimate built principally from evidence and occupational structures in richer economies can be carried into decisions about East African tax incentives, labour policy, digital infrastructure, public procurement or education. The answer is not that Acemoglu is wrong. It is that economic evidence travels only after its assumptions have been checked against the place where the law will operate.

Technical exposure is different from economic adoption

Many AI studies ask whether a task could technically be performed or assisted by a model. A business owner asks a different question: is it cheaper, reliable enough and practical to deploy here? That gap is especially important in East Africa. The International Labour Organization’s 2025 refined global index of occupational exposure finds that exposure to generative AI rises with country income and that clerical and highly digitised occupations remain among the most exposed. The ILO also emphasises that most exposed jobs are more likely to be transformed than eliminated because human input continues to matter.

For East African policy, an exposure score should therefore be treated as the beginning of inquiry rather than evidence of actual adoption. A task can be technically automatable and commercially unattractive. Electricity reliability, internet quality, cloud costs, access to payment systems, cybersecurity, language support, data quality and the availability of staff who can supervise the system all affect whether a firm will use AI. The ILO’s wider AI work expressly recognises that infrastructure bottlenecks, skills deficiencies and technology costs can widen productivity gaps between countries and between larger firms and small or micro enterprises.

Low wages can change the business case for automation

The price of labour is another assumption that often disappears when global AI predictions reach local policy debates. Where wages are high, replacing a repetitive task with software may quickly generate savings. Where wages are low, the same subscription, cloud bill or imported machine may not be financially attractive. This does not mean low wage countries are protected from automation. It means that the threshold at which automation makes commercial sense can be different.

The composition of employment matters too. Rwanda’s 2024 Labour Force Survey reported that 90.4 percent of employment was informal, with market oriented agriculture accounting for a large share of informal work. The World Bank reported in 2026 that nine out of ten employed Ugandans work in the informal sector. A model based heavily on formal office occupations may therefore describe only a limited part of the labour market. An AI system that drafts contracts, analyses spreadsheets or automates customer service may have a large effect on a bank, insurer or professional firm while having little immediate effect on a farmer, street trader or informal construction worker whose tasks are physical, variable and embedded in local relationships.

Small firms face a different productivity equation

Large firms can spread the fixed costs of AI across many employees and customers. They may have internal data teams, lawyers, cybersecurity staff and procurement systems. A small enterprise may rely on one accountant, one laptop and intermittent connectivity. Even where a general purpose AI service is inexpensive, the hidden costs of checking outputs, protecting customer data, training staff and correcting errors can consume the apparent productivity gain.

This is why East African productivity policy should distinguish frontier capability from firm level absorption. A government may be tempted to offer tax incentives for “AI adoption” as if buying technology were the same as increasing productivity. A better legal design would ask for evidence of the business function being improved, the cost of adoption, the effect on output or service quality, and any impact on workers. Incentives for digital transformation should reward measurable productive use rather than the presence of an AI label in a procurement invoice.

Productivity gains do not automatically become worker gains

Acemoglu and Simon Johnson make a second point that is particularly important for employment policy: higher average productivity does not guarantee higher wages or better jobs. Their historical work on machinery and labour argues that automation can raise wages when it creates new tasks that increase workers’ marginal productivity or when complementary sectors expand, but productivity gains can coexist with weak wages, surveillance and poorer job quality when workers have little bargaining power.

That distinction matters for East African law. A company may become more productive by using AI to monitor drivers, score call centre workers, allocate delivery tasks or evaluate staff. Output may rise, but the legal questions then include privacy, discrimination, workload, disciplinary fairness and the transparency of automated decisions. The success of an AI policy should therefore not be measured only by output per worker. A credible public policy should also ask who receives the gain, whether the work becomes safer or more precarious, and whether workers have realistic routes to challenge harmful automated decisions.

Imported predictions should not become automatic legal presumptions

The temptation to use a headline number becomes strongest when governments are making law. A legislature may assume that AI will remove a fixed percentage of jobs and create a tax. A ministry may assume that AI will raise productivity and subsidise adoption. A university may redesign training around occupations said to be “at risk.” A public body may procure an automated system because international studies describe the technology as efficient. Each decision can be rational, but none should treat an external exposure estimate as conclusive evidence of local effect.

The safer approach is an East African adoption test. Before a major policy relies on predicted AI productivity, officials should examine the actual local task, the sector’s degree of formality, the wage and technology cost relationship, available digital infrastructure, the skills needed to supervise the system, the quality of data, and the capacity of smaller firms to adopt it. The evidence should distinguish a demonstration from sustained use. It should also separate time saved from value created: a task completed faster does not improve productivity if staff spend the saved time correcting unreliable output or if customers lose trust.

The Common Market gives the region a reason to build its own evidence

The East African Community Common Market Protocol already commits Partner States to programmes promoting employment creation, vocational and technical training, social protection and social dialogue. Those objectives provide a legal and institutional reason to develop regional evidence about technological change. Comparable surveys of AI adoption, task changes, wages, business size and productivity would be more useful than simply importing a single estimate from Europe or North America.

The practical conclusion is modest but important. Acemoglu’s productivity argument should be used in East Africa as a method of asking questions, not as a number to copy. His task based approach encourages policymakers to identify which tasks are affected and what cost savings are actually achieved. East Africa should add its own filters: informality, wage structures, infrastructure, small firm capacity, language, skills and local demand. Good law is not made by rejecting international research. It is made by knowing which parts travel, which assumptions do not, and what evidence must be collected before a prediction becomes a rule.

Source note. This article is based on Daron Acemoglu, The Simple Macroeconomics of AI, NBER Working Paper 32487 (2024); Daron Acemoglu and Simon Johnson, Learning from Ricardo and Thompson: Machinery and Labor in the Early Industrial Revolution, and in the Age of AI, NBER Working Paper 32416 (2024); the International Labour Organization, Generative AI and Jobs: A Refined Global Index of Occupational Exposure (2025), Generative AI and Jobs: A 2025 Update, and Artificial Intelligence Adoption and Its Impact on Jobs (2025); the National Institute of Statistics of Rwanda, Labour Force Survey Annual Report 2024; the World Bank, 27th Uganda Economic Update (2026); and Article 39 of the East African Community Common Market Protocol.

Suggested citation: 

Ronald Serwanga, “AI Productivity in East Africa: What Evidence Fits?” East Africa Legal Insight (5 September 2026).