Public Procurement as an AI Governance Tool in Africa

Photo by Nik on Unsplash

Introduction

African governments are signing AI partnerships for public services, especially in areas where AI is expected to contribute to socioeconomic development. For example, in February 2026 Rwanda signed a three-year partnership with Anthropic that gives government developers access to Claude and Claude Code to improve public services, with health uses among its aims.

At the same time, many African governments are approaching AI regulation cautiously. This is seen in African AI strategies that remain largely untranslated into law, and in the African Union's Continental AI Strategy, whose first implementation phase focuses on governance structures and resource mobilisation instead of punitive legislation. It is also seen in the reception of proposed legislation. In October 2026 Kenya's Senate ICT Committee rejected the Artificial Intelligence Bill, 2026, at committee stage, finding its penalties too harsh and its obligations too heavy for a sector still in its early stages. The picture is not uniform, however, with some countries already beginning formal legislative processes towards an AI law such as Ethiopia and other countries are considering legislation of their own.

While the history of technology regulation in Africa shows the need to be cautious and the need for more grounded policy proposals, the idea of leaving AI development and deployment in a governance vacuum carries its own risks. Systems used to allocate benefits, support clinical decisions or process citizens' data can affect rights and livelihoods long before a dedicated AI law exists. Choices made at the point of adoption, such as which vendor is selected, on what terms and with what access to public data, are also difficult to reverse once systems are embedded in government operations.

This leaves a practical question. Where dedicated AI laws are still taking shape, or where governments have chosen a more gradual approach to regulation, what tools do African governments have to govern the AI they are already buying and deploying? One answer gaining attention internationally is public procurement. This post examines that answer in two parts. Part I looks at procurement as a policy instrument, first in general and then in African history. Part II turns to its use in AI: the market-shaping argument and how AI procurement differs from conventional technology procurement, drawing on policy and evidence from the United States and Europe, before considering what all of this means for African governments, including the opportunities procurement offers and the conditions on which its effectiveness depends.

Part I: Procurement as a policy tool

1. A long history of procurement linkages

The proposal to govern AI through procurement is the latest version of an old idea. Public procurement has long served as a means for governments to pursue social, economic and environmental objectives beyond the procurement of goods and services itself, and understanding how that practice has worked, and where it has fallen short, gives a better basis for judging what procurement can do for AI.

Christopher McCrudden's Buying Social Justice (2007) provides the leading account of this practice. It examines how governments in developed and developing countries use their contracting power to advance equality and reduce discrimination, with case studies including the UK, the United States and South Africa, and it describes these uses as "procurement linkages" (McCrudden, 2007). In the UK the practice goes back to nineteenth-century efforts to secure fair wages for employees of government contractors (Conley et al., 2025). In the United States, industrial policy linkages such as "Buy American" and contract compliance programmes followed in the twentieth century (McCrudden, 2009).

The European Union formalised the idea in its 2014 procurement directives. Directive 2014/24/EU sought greater inclusion of societal goals in procurement, including environmental protection, social responsibility, innovation, employment and public health (European Parliament, 2014). This shift, often called strategic procurement, treats purchasing as an instrument of public policy in its own right. The EU's experience also shows that permitting such goals is different from achieving them. A later evaluation found that the 2014 framework permitted strategic procurement but did not ensure its uptake, and the Commission's revision proposal therefore seeks to provide clearer legal bases for environmental, social and innovation requirements (European Commission, 2026). Even in a well-resourced procurement system, permitting buyers to pursue policy goals was not sufficient to ensure that they did so.

2. The case for and against using procurement to advance policy goals

The growth of strategic procurement has not settled the question of whether public purchasing should serve wider policy goals. The debate is worth setting out, because the same arguments now recur in discussions about using procurement to govern AI.

Supporters argue that procurement linkages are a legitimate and efficient means of advancing public objectives. McCrudden makes this case in relation to equality, contending that contractual conditions can secure compliance with equality norms where other methods of enforcement have proved inadequate, and that public purchasing can help to supply equality as a public good (McCrudden, 2007). At the same time, the practice has attracted sustained criticism and the main objections are grouped into four categories (McCrudden, 2007). Economists and international trade lawyers have also raised a further objection. Each objection is set out below alongside the response that supporters of linkage offer.

2.1 Whether policy goals belong in purchasing

• Objection: Procurement is commonly understood as a means of obtaining what the state needs at the best value for public money. Requirements that have no bearing on the item being bought can be seen as stretching procurement beyond that function, and as turning purchasing officers into regulators.

• Response: Where a policy goal is closely connected to how a contract is performed, such as the wages paid to the workers who deliver it, it is difficult to treat that goal as irrelevant to the purchase. On this view, how public money is spent forms part of the value it secures.

2.2 Cost

• Objection: Meeting additional conditions adds to suppliers' costs, which are likely to be passed on in their prices. Conditions may also deter some firms from bidding, weakening competition, and these costs are borne by the public and may not be weighed systematically against the benefits the policy is intended to deliver.

• Response: Higher contract prices may be warranted where they produce public benefits that would otherwise require separate public spending to achieve. The relevant comparison is between the cost of the condition and the cost of pursuing the same goal by other means.

2.3 Fairness between bidders

• Objection: Favouring certain categories of suppliers may exclude firms that offer better price or quality, which sits in tension with the equal treatment of bidders, a central principle of procurement law. Conditions that are loosely defined or unevenly applied can also create room for arbitrary or improper awards.

• Response: Concerns about fairness between bidders can be reduced through transparent, objective and consistently applied criteria that are published in advance and open to challenge through review mechanisms.

2.4 Procurement or general regulation

• Objection: If an objective is important enough to pursue, it can be argued that it should be imposed by legislation applying to every firm. Attaching it to public contracts reaches only government suppliers, and relies on procurement officials to enforce rules they may not be equipped to monitor.

• Response: Procurement conditions need not replace legislation. They can work alongside it, giving the state direct leverage over the firms it pays and allowing it to set standards for public contracts where general regulation is absent or still developing.

2.5 Competition and international trade rules

• Objection: Economists and international trade lawyers argue that preferential conditions distort competition, including between domestic and foreign firms. For countries that are party to international procurement commitments, such as the WTO Government Procurement Agreement, preferences may also raise questions of compliance with obligations not to discriminate against foreign suppliers.

• Response: International procurement rules do not exclude linkage altogether. The revised WTO Government Procurement Agreement expressly permits procuring entities to use technical specifications to conserve natural resources or protect the environment (Australian Government, 2018), allows environmental characteristics to be included among evaluation criteria (World Trade Organization, n.d.), and contains general exceptions, although commentators note that the space it leaves for social and development objectives remains limited (Thrasher, 2017). Within the European Union, Directive 2014/24/EU expressly provides for social and environmental considerations in procurement (European Parliament, 2014).

The leading account reaches a measured conclusion that whether procurement linkages are beneficial depends on how they operate in practice, although their potential is considered worth pursuing (McCrudden, 2007). Most of these objections have a parallel in the AI context. It may be asked whether procurement officers should set standards for AI systems at all, and whether AI-specific conditions will raise costs, narrow the field of vendors or disadvantage certain providers. Equally, the responses offered by supporters apply, since conditions closely tied to how an AI system performs in public service are difficult to dismiss as irrelevant to the purchase.

3. Procurement as a policy tool in Africa

For the purposes of this analysis, the history of public procurement in Africa can be understood through three overlapping approaches: reform for integrity and efficiency, procurement for redistribution, and externally financed procurement. These approaches are not sequential stages, and all three continue to operate side by side in many African procurement systems. Each has left institutional legacies, in law, in administrative practice and in relationships between governments, development partners and suppliers, that are likely to shape how African governments procure AI.

3.1 First approach: reform for integrity and efficiency (1990s onwards)

Modern procurement law in many African countries was shaped by reforms that the World Bank and other development partners supported and, in many cases, required. Donor procurement requirements gave developing countries strong incentives to reform their procurement and financial management systems (Agaba & Shipman, 2007), and the World Bank's influence was particularly strong in Africa, both because of its leadership in the field and because it financed or supported reform directly (Williams-Elegbe, 2014b). Over roughly two decades, most sub-Saharan countries undertook some form of reform, typically by passing new legislation that drew on the UNCITRAL Model Law and by creating or strengthening procurement institutions (Williams-Elegbe, 2014a). The 1993 Model Law on which much of this legislation drew was built around the principles of competition, transparency, fairness and objectivity (UNCITRAL, 1993).

These reforms laid the legal and institutional foundations on which African procurement systems still rest, including regulatory authorities, standard procedures and review mechanisms. Their focus, however, was on integrity and value for money, and their design placed strong emphasis on open competition, so that using procurement to pursue wider policy goals was not a central concern. Implementation also proved uneven, and a study of South Africa and Nigeria found that reformed systems did not always deliver more efficient, transparent and effective procurement in practice (Williams-Elegbe, 2014a). The reforms therefore created substantial legal space for procurement, while the capacity needed to make full use of that space developed more slowly than the law itself.

3.2 Second approach: procurement for redistribution (late 1990s onwards)

A second approach has used procurement explicitly to pursue social goals, and South Africa offers the most developed example. Section 217 of the Constitution permits preferential procurement at the award stage, and preferential procurement was formally initiated in 1998 to favour historically disadvantaged individuals (Smallwood et al., 2011). The 1997 Green Paper on Public Sector Procurement Reform led to the Public Finance Management Act of 1999 and the Preferential Procurement Policy Framework Act of 2000 (World Bank, 2003), and the most recent development in this framework is the Public Procurement Act 28 of 2024. South Africa is widely cited internationally as a leading case of procurement being used to advance equality (McCrudden, 2009). Its courts have both enforced the regime, holding that public bodies must investigate credible allegations of fronting (Viking Pony Africa Pumps (Pty) Ltd v Hidro-Tech Systems (Pty) Ltd, 2010), and set limits on it, invalidating the 2017 Preferential Procurement Regulations on the ground that the Minister lacked the power to make them (Minister of Finance v Afribusiness NPC, 2022).

Kenya adopted a set-aside model through its Access to Government Procurement Opportunities (AGPO) programme, which requires procuring entities to reserve 30% of procurement spend for enterprises owned by youth, women and persons with disabilities. The programme was introduced through regulations in 2013 (Hivos, n.d.) and later carried into the Public Procurement and Asset Disposal Act 2015, but its implementation remains incomplete. For instance, the Auditor-General found that Kenya Power reserved only 11% of its procurement budget for these groups in the year to June 2025 (Mutua, 2025).

The experience of South Africa and Kenya shows that African procurement law can be used to advance policy objectives, and that governments have been willing to use it in this way. It also reveals a gap between setting a goal in law and achieving it in practice, because requirements such as ownership quotas are straightforward to write into legislation but difficult to monitor and enforce. In South Africa, the preferential regime has faced persistent concerns about fronting, where a firm lists eligible owners on paper while control and benefits remain with others, and the B-BBEE Commission has investigated a number of such cases (South African Government, 2020). In Kenya, some public entities continue to fall short of the AGPO targets.

3.3 Third approach: externally financed procurement

A third approach is externally financed procurement, in which loans and grants from foreign governments and their agencies determine or shape the choice of supplier. The practice is not new, since it predates the reform era through aid and export credit tied to the suppliers of donor countries, but it has expanded significantly since the 2000s, particularly through Chinese concessional finance. Such financing has enabled infrastructure investment that many governments could not otherwise afford, often on concessional terms and without the policy conditions attached to some Western aid (Warner & Ajibade, 2024). At the same time, it can narrow the choice of supplier, and a study of Uganda found that political interference and Uganda-China economic bilateralism shifted preferential schemes toward foreign companies instead of local ones (Wanyama, 2016).

In some countries this approach also rests on an express legal basis. In Kenya, section 4(2)(f) of the Public Procurement and Asset Disposal Act 2015 excludes from the Act procurement under bilateral or multilateral agreements between Kenya and a foreign government, agency or multilateral body (Bowmans, 2020). The 2020 Regulations qualify this exemption by providing that, where such an agreement is financed through a negotiated loan, the Act does not apply if the agreement specifies its own procurement procedures, and by requiring procurement professionals to be involved in negotiating these agreements so that Kenya's procurement interests are considered (Republic of Kenya, 2020). As a result, significant decisions about suppliers can be made outside the ordinary procurement system, and the safeguards that apply depend largely on what is negotiated in the financing agreement.

Taken as a whole, the three approaches show that African procurement systems have the legal framework needed to pursue policy goals, and that governments have been willing to use it. The same challenges, however, appear across all three, namely that compliance is difficult to verify, enforcement capacity is still developing and a share of strategic acquisitions is made outside ordinary procurement procedures.

4. Five conditions for procurement to deliver policy goals

Drawing on the global and African experience set out above, this post proposes that procurement is most likely to achieve policy goals when five conditions are met. These conditions provide the analytical framework for Part II.

• Leverage: The public buyer must be significant enough to the supplier to influence its conduct. A government that purchases a large share of a product can shape what is offered, as NASA did with early integrated circuits (discussed in section 5), whereas a government that is a minor customer of a global firm has limited influence.

• Specifiability: The policy goal must be capable of being expressed as concrete requirements in a tender or contract. A requirement to pay a minimum wage, of the kind found in early fair wages clauses, can be specified precisely, while a requirement of "fairness" is considerably more difficult to define.

• Verifiability: The buyer must be able to assess whether the requirements are met, both when bids are evaluated and throughout the performance of the contract. The difficulty of verifying ownership under South Africa's preferential regime and Kenya's AGPO programme shows how much depends on this condition.

• Legal and institutional space: Procurement law, international trade commitments and financing agreements must permit the buyer to impose the condition, and the acquisition must in fact pass through the procurement system. The WTO rules on preferences and Kenya's exemption for bilateral and multilateral agreements illustrate how this space can be both opened and limited.

• Capacity and integrity: Procurement officials must have the expertise to design and enforce conditions, and the system must be sufficiently free from capture that preferences are not diverted to favoured interests. The experience of the reform era, in which legal frameworks developed faster than the capacity to implement them, underlines the importance of this condition.

AI places pressure on each of these conditions, and several of them are still developing in many African contexts. Part II applies this framework to explain how.

Part II: Procurement and AI

5. The market-shaping power of procurement

One of the most ambitious versions of the procurement argument holds that, through deliberate AI procurement, governments can influence the wider AI market as well as the systems they buy, producing positive effects across the industry (Ben Dor & Coglianese, 2022). Whether this is plausible depends largely on the first condition set out in Part I, namely leverage, and the history of an earlier technology helps to show what that condition requires in practice.

5.1 The classic case of integrated circuits

One of the best-known examples of procurement shaping a technology market is the early semiconductor industry. In 1962 NASA announced that its prototype Apollo guidance computer would use integrated circuits from Fairchild, and the Air Force soon adopted them for the Minuteman II missile (Goldstein, 2018). By the mid-1960s NASA was reportedly purchasing around 60% of the integrated circuits produced in the United States (Goldstein, 2018). These large orders secured supply for the programme, and NASA also asked several other manufacturers to act as second sources (Leibson, 2024).

Three mechanisms were at work in this case:

• The government acted as a lead customer, buying an immature technology before most commercial buyers were prepared to do so.

• It imposed demanding quality and reliability standards, which suppliers then carried into their commercial products.

• It used second-sourcing to keep the supply base competitive.

Each of these mechanisms depended on the government being the dominant buyer of a new and largely standardised product, and it is on this point that the comparison with AI turns.

5.2 Why AI presents a more difficult case

General-purpose AI differs from early integrated circuits on several of the features that made that case successful.

• Government share of demand: In the 1960s the US government was the dominant buyer of a nascent product, whereas today governments are significant and growing customers for AI but only one group among many, alongside large consumer and enterprise markets.

• The nature of the product: Integrated circuits were standardised components built to specification, while general-purpose AI models are developed for global markets and generally offered on standard terms that individual buyers have limited ability to change.

• Who sets the standards: NASA and the Air Force imposed their own reliability requirements on suppliers, whereas standards for AI are set largely by developers and industry bodies and, increasingly, by regulators in some jurisdictions.

• The structure of the supply base: Several manufacturers produced integrated circuits, which made second-sourcing possible, while frontier AI development is concentrated among a small number of developers, alongside a wider market of smaller and open models.

For frontier and general-purpose models, individual governments therefore have limited leverage, and this is particularly true of smaller states. Market shaping is nonetheless more plausible in two narrower settings:

• The first is sector-specific systems for which governments are the main or only buyer, such as tax administration, identity management, social protection and public health.

• The second is aggregated demand, where many public buyers adopt common terms, and the European Union offers the clearest example. Standard clauses for procuring algorithmic systems, developed by the City of Amsterdam in 2018 (European Commission, 2023), were adapted into EU model contractual clauses for AI (MCC-AI), first proposed in September 2023 and updated in March 2025 to align with the AI Act, with separate templates for high-risk and other systems (Cooper et al., 2025). As the clauses are non-binding and public authorities are only encouraged to report their use, however, there is no reliable measure of how widely they have been adopted (Cooper et al., 2025). In the United States, OMB Memorandum M-25-22 (Driving Efficient Acquisition of Artificial Intelligence in Government) similarly aims to bring greater uniformity to federal AI acquisition (Bourne & Howe, 2025).

5.3 Market shaping as industrial policy

Market shaping can also mean building a domestic supply base. M-25-22, for instance, directs federal agencies to prioritise AI developed in the United States (OECD.AI, 2026), following the same logic as Buy American policies and NASA's role as a lead customer, in which government demand is used to grow national industry. This objective is distinct from managing AI risk, and although the two can be complementary, they may also conflict where the preferred domestic supplier is not the safest or most accountable option available.

5.4 The African position

For African governments, leverage over the largest global AI developers is limited, since individual African governments are unlikely to be significant customers for frontier laboratories. Direct partnerships, such as Rwanda's agreement with Anthropic, may give some governments a degree of influence, although the terms of such arrangements are not usually made public, which makes that influence difficult to assess.

Leverage is stronger in two other respects. First, in many African countries the state is among the largest domestic buyers of technology, and it has been argued that the public sector is the largest market for AI solutions in the developing world (Nagitta et al., 2022), which gives governments meaningful influence over local AI firms and systems integrators. Second, leverage could be increased through aggregation, for example through common contractual clauses developed at regional level.

The most realistic market-shaping role for African procurement is therefore twofold:

• It can set standards for local and regional vendors and integrators.

• It can support the growth of local AI supply through preferences of the kind Kenya's AI Strategy contemplates in emphasising local content in procurement and partnerships (Ministry of Information, Communications and the Digital Economy, 2025).

The experience described in section 3 indicates, however, that preference schemes require effective verification, an area in which AGPO and South Africa's preferential procurement regime continue to face challenges. Procurement conditions can also carry costs for the local firms that preferences are meant to support. Requirements for testing, documentation and audit rights may be easier for large international vendors to meet than for smaller African suppliers, so conditions would need to be proportionate to the risk of each system if they are not to work against the goal of growing local supply.

6. AI procurement compared with traditional technology procurement

Public procurement systems were designed mainly to acquire goods, works and services whose characteristics are largely settled at the point of purchase, such as a road, a vehicle or a licence for a defined piece of software. The argument for governing AI through procurement assumes, at least implicitly, that AI can be acquired in a similar way, with the relevant conditions written into the tender and the contract. Several features of AI make this assumption harder to sustain. Some of these features are shared with modern cloud-based software, so the differences are often of degree, but they are more pronounced for AI, as the following six differences illustrate.

• Stability after purchase: Traditional technology behaves largely as it did at delivery, with changes introduced through versioned releases, whereas an AI system's behaviour can change when the vendor or an upstream model provider retrains, updates or retires a model. This mainly affects verifiability.

• Supply chain: With traditional technology the buyer usually contracts with the party that built the product, whereas an AI vendor often builds on a model supplied by a third party with which the buyer has no direct contract. This affects both leverage and specifiability.

• Data: Traditional systems process the buyer's data to deliver the contracted service, whereas with AI the provenance of training data also matters, and the buyer's data may be retained or used to improve models. This affects specifiability and legal space.

• Performance: Traditional technology is tested against a defined specification, whereas AI performance is probabilistic and context-dependent, and may vary across tasks and populations. This affects verifiability and capacity.

• Commercial model: The price of traditional technology is generally agreed at signature, whereas AI is often priced by consumption, with costs continuing to accrue after the contract is signed. This mainly affects capacity.

• Entry route: Traditional technology usually enters through a discrete purchase decision, whereas AI features may be added through updates and changes to the terms of products already in use. This affects legal and institutional space.

Commentary from procurement consultancies and practitioners identifies each of these differences, and research supports most of them.

• Changes outside the customer's view: Practitioners warn that an AI product can change its behaviour for reasons the customer cannot see, for example when the provider of the underlying model retrains it or withdraws a version, and that many vendors build their products on models supplied by third parties (Elevate Consult, 2026). Systematic testing confirms that such changes can be substantial. In a study published in the Harvard Data Science Review, Chen, Zaharia and Zou compared versions of GPT-4 released three months apart in 2023 and found that its accuracy on one task fell from 84% to 51%, while its willingness to answer sensitive questions and its ability to follow user instructions also declined (Chen et al., 2024). Some researchers have questioned how far these results reflect a decline in capability, arguing that they show changes in the model's default behaviour (Narayanan & Kapoor, 2023), but on either reading the study shows that a product's behaviour can shift significantly without any change to the customer's contract.

• Use of the buyer's data: Buyers are advised to confirm whether their data will be kept and used to improve models (Redress Compliance, n.d.). A Stanford analysis of the privacy policies of six US frontier AI developers found that all six appear to use users' chat data to train their models by default, that some retain chat data indefinitely, and that the policies often lack essential information about these practices (King et al., 2025). The study concerns consumer services, and enterprise and government contracts may carry different terms, which is precisely why these terms need to be settled through procurement.

• Variable performance across populations and languages: Practitioners caution that the performance of AI tools can vary between uses and types of data (Redress Compliance, n.d.), and independent evaluations show that this variation often disadvantages African users. The US National Institute of Standards and Technology tested 189 face recognition algorithms from 99 developers and found that false positive rates often differed across demographic groups by factors of 10 to more than 100, with differentials varying widely between algorithms, and that, for high-quality photographs, they were highest for West and East African and East Asian people (Grother et al., 2019). For language models, the IrokoBench benchmark evaluated 16 open and proprietary models in 17 African languages and found a significant performance gap between high-resource languages such as English and French and the African languages tested (Adelani et al., 2025).

• Unpredictable costs: AI contracts often charge a rate per unit of use, such as per request, which makes the total cost harder to predict (Bower, 2026). In a 2025 IDC survey of 318 senior decision-makers in large organisations in the United States, the United Kingdom and Ireland, 96% of those deploying generative AI reported costs that were higher or much higher than expected, and 71% said they had limited ability to tell where their AI costs originated (IDC, 2025). The survey was commissioned by an AI vendor and covers private organisations in three Western countries, so it is best read as indicative.

• Entry through updates and informal routes: AI features may enter an organisation through updates to products it already uses, accompanied by changes to the terms of service (Vota, 2026). In a study based on interviews with 19 city employees responsible for AI procurement across seven US cities, Johnson and colleagues found that AI acquired by cities often does not pass through a conventional procurement process at all, which makes oversight more difficult (Johnson et al., 2024). They also found that long-standing procurement laws and norms shape which AI is purchased and who holds decision-making power over it (Johnson et al., 2025).

Taken together, the practitioner commentary and the research indicate that these risks are real and, in several cases, measurable, although systematic evidence on how often they arise specifically in public procurement remains limited. Policy has begun to respond. M-25-22, for example, addresses the entry-route problem by requiring federal agencies to consider AI used by contractors in performing government contracts, including where AI is not being deliberately acquired (Bourne & Howe, 2025).

6.1 From a single purchase to a continuing relationship

What these differences have in common is that they unfold over time. Traditional procurement places most of its controls at the award stage, through specification, evaluation, selection and the signing of the contract, whereas with AI many of the matters that require oversight arise after the contract has been signed. As commentary from the development sector notes, a model's behaviour is assessed at a particular moment while the product itself continues to change (Vota, 2026). Governing AI through procurement therefore relies increasingly on contract management throughout the life of the contract, including monitoring performance, controlling updates, auditing the use of data and enforcing the right to exit.

Recent policy instruments reflect this shift towards contract management. The EU's model contractual clauses for high-risk AI provide that suppliers must, on request, test the system in the public organisation's own environment, and must keep their risk management process under regular review (European Commission, 2025). In the United States, M-25-22 similarly requires contracts for high-impact AI to include performance monitoring and regular risk assessment (OECD.AI, 2026).

6.2 The critique of regulation by contract

These differences give particular force to the critique advanced by Albert Sanchez-Graells. His 2024 book argues that relying on regulation by contract creates a false sense of security and leaves the public sector exposed to the pressures surrounding new and heavily promoted technologies (Sanchez-Graells, 2024), and a related article described the expectation of responsibly buying AI as a "regulatory hallucination" (Lynskey, 2024). Three of his arguments correspond closely to the differences set out above.

• Procurement cannot easily discipline both parties: Regulation by contract assumes that procurement can constrain both the vendor and the public body that uses the AI system, and he argues that this assumption does not hold for digital technologies (Lynskey, 2024). Procurement officials have limited control over how a ministry later deploys a system, and because AI continues to change, its use after the award is where much of the risk lies.

• Private standards fill gaps in public guidance: Where governments have not translated AI principles into detailed guidance, public buyers tend to rely on private standards (Lynskey, 2024). The layered supply chain reinforces this, since many of the most important terms are set by upstream model providers.

• Public sector capability is limited: A growing digital capability gap in the public sector constrains buyers' ability to specify, evaluate and enforce technical requirements (Sanchez-Graells, 2022), which matters all the more because AI governance increasingly depends on continuing oversight after the contract is awarded.

In response, Sanchez-Graells proposes an institutional remedy that combines independent oversight, mandatory standards and a dedicated regulator for public sector AI (Lynskey, 2024), on the view that the public sector needs to move beyond regulation by contract (University of Bristol, 2024).

Empirical evidence from Europe is consistent with these concerns and with the US findings discussed above, most notably a 2026 study by Emma Kallina and colleagues titled "It's Just a Wild, Wild West". Based on interviews with public sector buyers and procurement experts in the EU and the UK, the study concludes that the potential of procurement as an AI governance mechanism is not yet realised there (Kallina et al., 2026). It finds that AI systems often enter public bodies through routes outside formal procurement, that AI-specific procurement practices remain fragmented and nascent, and that large technology suppliers exert strong influence over both what is procured and the requirements attached to it (Kallina et al., 2026). The study also identifies six promising practices for strengthening AI procurement, including interdisciplinary procurement teams, greater AI literacy among buyers and knowledge-sharing between public bodies (Hudig, 2026), and its findings are informing the Council of Europe's work on AI procurement guidelines, due by the end of 2026 (ITU, 2026).

The critique and the evidence do not imply that procurement has no role. Instruments such as the EU model clauses and M-25-22 show that contracts can carry meaningful obligations on testing, monitoring and risk management, and procurement can also work alongside other forms of oversight. However, no published evaluation has yet shown that such conditions have changed how procured AI systems behave in practice, and evidence on their uptake and effect has not been systematically collected in either the United States or the European Union. The broader implication is that AI shifts much of the weight of procurement governance from the award stage, where legal rules are most developed, to contract management, where institutional capacity matters most. Where that capacity is weak, procurement conditions may therefore exist on paper while having limited effect in practice.

7. Governing AI through procurement in African contexts

7.1 Opportunities and considerations

In many African countries, the conditions that make procurement an attractive AI governance tool also affect how much it can achieve. Procurement is attractive because dedicated AI laws are still taking shape or are being approached gradually, as the rejection of Kenya's AI Bill illustrates, and because it can be used within existing legal frameworks without new legislation. At the same time, procurement is most effective when the five conditions in Part I are met, and several of these conditions are still developing in many African procurement systems. Technical evaluation capacity is being built, many AI standards are developed outside the continent, and some strategic technology is acquired through routes other than ordinary procurement. Procurement therefore takes on greater importance where dedicated AI law is still emerging, while its effectiveness depends on continued investment in the institutions that support it.

Two further considerations shape this picture. The first concerns the balance between adoption and governance. Kenya's AI Strategy aims to accelerate AI adoption and position the country as a regional hub, and views on the AI Bill were divided. Some critics argued that its proposed regulation could slow the growth of a nascent sector (Mutuku & Otieno, 2026), while others, including trade union commentators, argued that it needed to do more to protect workers affected by algorithmic management (Otiato, 2026). The second concerns the balance between sovereignty and partnership. African governments seek local content and control over data, while much of their technology is financed or supplied through partnerships with foreign governments and companies, which often bring their own suppliers and terms.

The first of these considerations also presents an opportunity. Since procurement applies only to the state's own use of AI and does not impose obligations on private developers more generally, it may offer a form of governance that is compatible with both responsible adoption and a measured approach to AI regulation.

7.2 The five conditions in African contexts

The five conditions can be applied to African contexts as follows, together with what each implies for AI procurement.

• Leverage: African governments have limited influence over global AI developers but greater influence over local vendors and integrators (Nagitta et al., 2022). This points to a focus on local supply chains and on aggregating demand at regional level.

• Specifiability: AI standards are largely developed outside the continent, and law firms have already proposed adapting the EU MCC-AI to South Africa's Protection of Personal Information Act (Cliffe Dekker Hofmeyr, 2025). International clauses are therefore useful starting points, but they need adapting to African law, languages and use cases.

• Verifiability: Public capacity to evaluate models is still being built, and Kenya's Strategy proposes a technical centre for evaluation and certification (Ministry of Information, Communications and the Digital Economy, 2025). Since systems developed elsewhere can perform less well for African populations and languages, as the NIST and IrokoBench evaluations discussed in section 6 show, shared national or regional evaluation capacity, including testing on local populations and languages, is a precondition for effective AI procurement.

• Legal and institutional space: Some technology is acquired through donations, financing agreements and government-to-government arrangements, so procurement rules and guidance need to reach acquisition routes outside open tenders.

• Capacity and integrity: Procurement capacity is still developing (Williams-Elegbe, 2014a), and preference schemes face verification challenges. Contract management and verification are therefore as important as tender rules.

7.3 How technology enters government

Of the five conditions, the fourth, legal and institutional space, merits particular attention, because public technology in Africa has often been acquired through arrangements that sit outside ordinary procurement procedures. Some systems have been introduced through donations of equipment and donor-funded pilot projects, which can help governments adopt new technology quickly and at little upfront cost. Research on biometric identification systems in Africa has noted that the impetus to adopt such systems often comes from external funders, and that this can create reliance on foreign technology providers over time (Roberts, 2026). Other systems have been acquired through concessional loans and financing packages linked to particular suppliers, which have made substantial infrastructure investment possible while also determining the choice of supplier at the financing stage (Warner & Ajibade, 2024). Still others have been acquired through direct procurement or government-to-government agreements, often in response to tight statutory deadlines, where the time available for competitive tendering and detailed contract negotiation is limited.

Each of these routes has served legitimate purposes, but each also places important decisions about technology outside the procedures through which procurement conditions are usually set. AI is likely to enter government through similar routes, including free trials of enterprise tools, AI features added to existing software and cloud services, donor-funded pilots in sectors such as health and agriculture, and AI components within larger financed infrastructure projects. The partnerships described in the introduction represent a further route, since they are generally concluded as memoranda or framework agreements and their detailed terms are not always made public. For these reasons, the finding from the US and European studies discussed in section 6, that a significant share of AI enters government outside formal procurement, is likely to apply in many African contexts as well.

7.4 Data protection law as an existing anchor

Existing data protection frameworks offer a further foundation on which AI procurement can build, and they go some way towards providing the independent oversight that Sanchez-Graells recommends. While most African states have not yet adopted dedicated AI laws, many have enacted data protection legislation and established independent data protection authorities, whose mandates typically extend to public bodies and include requirements such as data protection impact assessments and contracts with data processors. Courts have also applied these frameworks to public technology projects. In October 2021, for example, Kenya's High Court required the government to conduct a data protection impact assessment under section 31 of the Data Protection Act before further processing data for the Huduma Namba digital identity programme (People Daily, 2021). Since that assessment was required after enrolment had begun, the case also illustrates the value of applying such requirements earlier. Linking procurement to existing data protection obligations would allow these assessments to take place before a system is acquired, giving data protection authorities and procuring entities a shared point at which to consider the risks of AI systems that process personal data.

7.5 Strengthening the role of procurement

The analysis above suggests seven measures that could help procurement contribute more effectively to the governance of AI in the public sector.

• Defining acquisition broadly: Procurement rules and guidance could cover the full range of routes through which AI enters government, including free trials, donations, donor-funded pilots, AI features added to existing contracts and the use of AI by contractors, an approach reflected in M-25-22.

• Agreeing terms at the financing stage: Where loans or grants shape the choice of supplier, terms on data, audit, exit and hosting could be agreed within the financing agreement itself, with the involvement of the national treasury, the relevant line ministry and procurement professionals. Kenya's 2020 Regulations already provide for the involvement of procurement professionals in such negotiations.

• Integrating data protection requirements: A data protection impact assessment and an appropriate data processing agreement could be required before any AI system that processes personal data is acquired or piloted, giving data protection authorities a clear role in the process.

• Publishing information on AI systems in use: A public register of government AI systems, as recommended in Kenya's AI Strategy, would allow parliaments, courts, the media and civil society to see which systems have been acquired and on what terms.

• Sharing capacity and contractual terms: National procurement regulators or regional bodies could develop shared evaluation capacity, including the testing of systems on local populations and in local languages, together with model contractual clauses adapted to African legal frameworks. This would strengthen the position of public buyers and complement international standards with locally developed ones.

• Investing in contract management: Given that much of AI governance takes place over the life of a contract, procuring entities would benefit from resources to monitor updates, performance and data use, and from contract terms that secure exit and data portability rights.

• Building verification into preference schemes: Where local-content preferences are used to support the growth of African AI suppliers, they could be accompanied by verification mechanisms that draw on the lessons of AGPO and South Africa's preferential procurement regime, with procurement conditions set in proportion to the risk of each system so that smaller local suppliers are not excluded.

Conclusion

Using procurement to govern AI builds on a long tradition of procurement linkages, from fair wages clauses in the nineteenth century to South Africa's preferential procurement regime and Kenya's AGPO programme. That tradition suggests that procurement can help to advance policy goals when five conditions are met, namely leverage, specifiability, verifiability, legal and institutional space, and capacity and integrity.

AI raises new considerations for each of these conditions. Individual governments have limited influence over the developers of general-purpose AI, and because AI systems can change after purchase, rely on upstream providers outside the contract and may enter government through updates and trials, much of the work of governance shifts from the tender stage to contract management. Experience in the United States and the European Union reflects these considerations. Policy instruments for AI procurement are developing quickly, while research shows that some AI enters government outside formal procurement and that established practices and large suppliers influence outcomes. Evidence on how far procurement conditions have shaped the behaviour of procured AI systems is still being gathered.

For African governments, procurement offers a practical way to guide public sector use of AI while dedicated AI legislation is still taking shape or is being approached gradually, and it does so without placing new obligations on private developers more generally. Its effectiveness will depend on continued investment in evaluation capacity, the development of locally adapted standards, and attention to the full range of routes through which technology is acquired, including partnerships and externally financed projects. Procurement is therefore likely to be most effective as one part of a wider governance framework, working alongside existing data protection laws and authorities, shared technical capacity, public registers of AI systems, and the oversight of courts and parliaments.