AI Tax Rules in East Africa: What Should Firms Pay?

 The question is not simply whether AI should pay tax

Bill Gates has revived an old argument in a new form. In his 26 August 2026 essay, “The turbulent AI era is here. The choices we make now are critical,” he argues that governments should rebalance the tax treatment of labour and capital and considers taxes on AI tokens and robots as part of the response to worker displacement. The concern is understandable. If a business produces the same output with fewer employees, wage based tax collections may weaken at the same time that governments face more pressure to fund retraining, income support and public services. But for East Africa, the first legal question is more basic: what exactly would an “AI tax” tax?

Artificial intelligence is not a single taxable object. A company may buy software from a foreign supplier, subscribe to a cloud service, run an open source model on its own servers, purchase a robot, automate an accounting process, or use an AI assistant without reducing a single job. Each transaction has a different legal character. A tax aimed simply at “AI” could therefore catch ordinary software investment while missing the activity that policymakers actually care about: substitution of labour, concentration of profits, or cross border digital income. The design problem comes before the rate.

East Africa already taxes parts of the AI economy

The region is not starting from a tax vacuum. Kenya’s Income Tax Act contains a significant economic presence tax for certain non resident persons earning income from services supplied over the internet or an electronic network, including through a digital marketplace, where the user is in Kenya. The current law taxes the deemed taxable profit at the statutory rate rather than creating a special category called “artificial intelligence.” Uganda’s Income Tax Act similarly taxes income of non residents providing digital services to customers in Uganda and expressly includes cloud computing, data services, online marketplaces and other internet delivered services. Rwanda’s online services VAT framework, implemented through a 2026 Ministerial Order, requires foreign suppliers of taxable online goods and services to register for VAT when supplying customers in Rwanda, and Rwanda Revenue Authority explains that the standard VAT rate is 18 percent.

Those rules already reach many commercial routes through which AI is sold. An overseas model provider may therefore face a digital income tax or VAT obligation even though no AI specific levy exists. A local company using AI may already pay corporate income tax on profits, employment taxes for workers, VAT on taxable purchases, customs duties on qualifying imported equipment, and other sector charges. Before adding another levy, a government should identify the revenue gap that existing rules do not address. Otherwise an AI tax may become an additional charge on the same economic activity rather than a carefully targeted response to automation.

Five possible tax bases produce five different laws

A tax on the AI provider would resemble a digital services or significant economic presence tax. It is administratively attractive because the supplier can often be identified, but it does not necessarily track job displacement. A foreign AI company could sell a medical diagnostic service that expands access without replacing workers, yet still pay the tax. Conversely, a local employer could use a freely available model to reduce staff and create no new provider side liability.

A tax on the business user would be closer to an automation levy. The difficulty is defining the taxable event. Is the tax triggered when software performs a task previously done by an employee, when headcount falls, when labour costs decline, or when an automated process exceeds a particular value? Each approach invites disputes about causation. A business may reduce staff because demand fell, reorganise departments, outsource work and introduce AI at the same time. Tax administrators would then have to decide how much of the change was caused by AI.

A tax on compute or AI tokens is more measurable in some settings, but it risks becoming a tax on use rather than harm. High compute consumption can support research, translation, fraud detection or pharmaceutical development without displacing employees. It can also be purchased outside the country. A tax on robots is clearer where physical machinery is imported or depreciated, but it covers only one part of modern automation. A tax on profits is the most familiar route and may capture gains from higher productivity, yet ordinary corporate tax does not by itself distinguish profits created by AI from profits created by better management, new capital or stronger demand. The choice of base therefore determines the policy.

The informal economy changes the economics of an AI levy

East African labour markets make the issue different from the advanced economies in which many robot tax debates began. Rwanda’s National Institute of Statistics reported that informal employment accounted for 90.4 percent of total employment in 2024. The World Bank’s 27th Uganda Economic Update, published in 2026, states that nine out of ten employed Ugandans work in the informal sector without the security or benefits associated with formal jobs. These figures matter because a tax attached to formal payroll substitution may operate on a relatively narrow part of the labour market while leaving much informal production untouched.

There is also a competition problem. A compliant medium sized company that records employees, pays taxes and buys licensed AI could become an easy target for an automation levy. A smaller informal competitor using similar technology might remain outside the system. The result could be a tax that unintentionally penalises formalisation. The International Labour Organization has repeatedly warned that technology can interact with informality in different ways; digitalisation does not automatically make work formal. East African tax design therefore has to consider enforcement capacity and the incentives it creates for firms to stay inside the formal economy.

Retraining revenue must pass through public finance law

Gates’s strongest argument is not that robots deserve a tax bill. It is that a technological transition may reduce labour tax revenue while increasing the need for worker support. East African governments can legitimately decide that additional revenue should finance retraining or social protection, but calling a tax “for workers” does not itself place the money in a worker fund. Public finance rules still govern collection and spending.

Kenya illustrates the point clearly. Article 206 of the Constitution places money raised or received by the national government in the Consolidated Fund unless an Act of Parliament validly excludes it for another public fund, and withdrawals generally require lawful appropriation. Uganda’s Constitution follows the same basic discipline in Articles 153 and 154: government revenues ordinarily enter the Consolidated Fund and expenditure must have constitutional or legislative authority. In practice, therefore, an AI levy that is meant to finance reskilling needs two legal designs, not one. The tax statute must define who pays and on what base; the public finance framework must lawfully direct and account for the expenditure. A political promise to earmark revenue is weaker than a transparent statutory fund or an annual appropriation tied to measurable programmes.

A better East African approach is to tax the problem, not the label

A defensible approach would begin with evidence of the problem a country is trying to solve. If the concern is that foreign digital firms earn substantial local revenue without adequate tax nexus, significant economic presence and VAT rules are the more direct tools. If the concern is unusually rapid labour displacement in a particular formal sector, government could study a temporary sector contribution or adjustment mechanism linked to measurable reductions in payroll rather than to the mere purchase of software. If the concern is financing retraining, the cleaner solution may be to strengthen ordinary revenue collection and make explicit budget appropriations for skills and social protection.

Tax incentives also deserve attention. A system that allows generous deductions for automation investment while heavily taxing labour can create a real bias, but the remedy need not be a new AI tax. Depreciation rules, investment allowances, payroll charges and training credits can be reviewed together. A government could, for example, condition certain automation incentives on investment in worker training or redeployment. That approach connects the fiscal benefit to transition costs without requiring officials to decide whether every spreadsheet assistant, chatbot or machine qualifies as artificial intelligence.

Regional coordination matters before the tax base fragments

East African businesses increasingly operate across several markets, while AI services are often supplied remotely from outside the region. If each country creates a different definition of an AI token, robot, automated process or displaced worker, compliance could become more expensive than the revenue collected. The East African Community Common Market Protocol already requires Partner States to coordinate social policies, promote employment creation and vocational training, expand social protection and promote social dialogue. Those commitments do not create an AI tax, but they offer a reason to discuss the transition regionally rather than through five unrelated national experiments.

The practical conclusion is that East Africa does not need to accept or reject Bill Gates’s proposal as a package. His warning usefully identifies a fiscal tension, but the region should separate the policy objective from the tax label. AI can already fall within corporate tax, digital service tax, significant economic presence rules and VAT. A new levy should only be considered after a government can state the missing taxable event, explain why existing taxes do not reach it, show that the measure will not punish formal businesses disproportionately, and create a lawful route for any promised worker support. The strongest tax is not the one with the newest name. It is the one that can be defined, administered, justified and spent transparently.

Source note. This article is based on Bill Gates, “The turbulent AI era is here. The choices we make now are critical” (Gates Notes, 26 August 2026); Kenya’s Income Tax Act, including section 12E on significant economic presence tax; Uganda’s Income Tax Act, including section 86 on non resident digital service providers; Rwanda’s 2026 Ministerial Order governing VAT on online goods and services and Rwanda Revenue Authority guidance; the 2024 Rwanda Labour Force Survey; the World Bank, 27th Uganda Economic Update (2026); International Labour Organization materials on artificial intelligence, digitalisation and work; Articles 206 and related provisions of the Constitution of Kenya; Articles 153 and 154 of the Constitution of Uganda; and Article 39 of the East African Community Common Market Protocol.

Suggested citation: 

Ronald Serwanga, “AI Tax Rules in East Africa: What Should Firms Pay?” East Africa Legal Insight (5 September 2026).