Helping LMICs Buy, Build, and Stimulate Quality AI Products and Services

Public procurement in low- and middle-income countries (LMICs) averages 13 percent of GDP—roughly $5.4 trillion annually. Although AI-enabled products—ranging from citizen-facing chatbots to back-office systems—currently represent a small share of procurement spending, it is likely to grow. Policymakers faced with procurement decisions worry about how to assess the quality of these products—and for good reason. Even before AI, technology procurement was plagued by failed deployments, inexperienced vendors, and costly vendor lock-in. AI introduces new challenges: intense private-sector hype fuels fear of being left behind, the products require new evaluation methods, and rapid technological change raises the risk of obsolescence.

LMIC governments need standards and procedures that give them confidence in procuring the best available solutions. Without strong standards to adhere to, the risk of misallocation is huge. Over the next year, CGD’s AI Initiative will consult experts in procurement, market design, government, and industry to understand their concerns and develop tools that help governments both procure and incentivize higher-quality AI products and services. We plan to publish our findings and recommendations as they emerge to prompt more discussion and adaptation.

Why now? What’s new?

There is demand and urgency from decision-makers—governments are inundated by AI pitches from vendors and nonprofits, and bureaucrats face pressure from senior leaders to keep pace with rapid advances in AI. We also think it’s worth testing whether existing procurement practices remain fit for purpose. This includes broader approaches that may sit upstream of a specific procurement decision, such as health technology assessment (HTA), and AI-specific procurement approaches such as the World Economic Forum’s AI Procurement in a Box, the UK Guidelines for AI Procurement and the UK’s Artificial Intelligence Dynamic Purchasing System, and Canada’s Artificial Intelligence Source List. We suspect that the existing literature insufficiently captures the screening needs facing LMIC decision-makers. Rather than reinventing these foundational resources, we propose to build on and complement them along three axes:

• Contemporary updates: Key resources and guidelines are now more than five years old and largely precede the emergence of large language models and agentic AI. Additionally, AI’s rapid progress raises questions about requiring continuous improvement and development from vendors.: Governments should not have to wait for the next procurement cycle to benefit from technological advances, yet specifying future upgrades is difficult when even vendors cannot predict how the technology will evolve. Updated and expanded guidance is required in view of new capabilities, a broader solution space, and larger risk surfaces.
• LMIC focus: Many existing AI procurement recommendations are designed for high-income countries and are not tailored to LMICs with thin markets and limited administrative capacity. Additionally, in many LMICs, collaboration options with development cooperation agencies and the tech-for-good and charitable sectors need to supplement a market-based perspective.
• Mechanism design: Drawing on the economics of mechanism design, contract theory, and industrial organization, we aim to develop procurement and market-shaping mechanisms that better reveal supplier capabilities, incentivize innovation, and improve procurement outcomes.

We think that procurement guidance that considers the above dimensions will help establish a credible theoretical and practical framework for LMICs to make better procurement decisions.

Designing for three procurement scenarios

We suspect that the bulk of public AI procurement can be captured through three core scenarios. While all of them can help governments procure the best-performing products, they require distinct approaches to screening and market incentivization. Before coming up with specific screening criteria to inform a procurement, governments should identify which of the following scenarios is most relevant to their needs and the available market. For example, requesting proposals for an off-the-shelf solution and designing screening criteria before a thin market has developed is likely to waste effort and lead to an unsuccessful procurement. The government may instead want to hire a vendor to build a specific solution, which requires an entirely different screening methodology. The three procurement scenarios are:

1. Buy: Governments purchase ready-made products and services.
   This refers to procuring “off-the-shelf” products and services that AI vendors pitch to governments. While many of these offerings come bundled with some implementation or maintenance services, the core distinction of this category is that the product already exists. That means vendors can often report on performance or impact metrics, and the products may have undergone some element of model, product, user, or impact evaluation (see this explainer for the distinction between these various types of evaluation) that the procurer can use to inform their decision.
2. Build: Governments hire a vendor to architect, advise, and build what they want.
   Sometimes no off-the-shelf product is available or none meets a government’s needs.. In such cases, governments often procure vendors to build solutions specific to their needs. While tenders set out the desired specification or outcomes, the product does not yet exist and therefore cannot be tested or compared at this stage. The task is instead to identify the vendor most capable of delivering the required solution at a competitive price. The challenge here is to assess capability where a vendor may not have a long track record on the specific application, where relevant skill indicators are unique, and where standard pre-qualification approaches are poorly suited to identifying capable contenders. Governments often also hire vendors to help architect, advise on, and source “off-the-shelf” components to integrate in an overall solution, instead of constructing everything from scratch. These scenarios face similar screening challenges in that there may be no standardized product or metrics to make comparisons across vendors.
3. Stimulate: Government incentivizes a domestic market.
   Sometimes, no vendors are active in a market, and no products exist even when an AI solution is technically feasible. This may reflect market failures or the fact that AI advances much faster than public services can be developed. In these cases, governments can shape markets by attracting new entrants and catalyzing investments. Existing approaches, such as pull mechanisms that reward outcomes instead of financing activities, may be needed to stimulate market entry and innovation. Methods such as the common task framework, which is a standardized evaluation benchmark that enables consistent comparison of different AI systems on a shared real-world task, can also be coupled with challenge prizes. These methods have been used successfully in promoting innovation and may be even more relevant given the speed of AI’s technical evolution.

There are also cross-cutting innovations and open questions across these three scenarios. For example, benchmarks do not have to be limited to market creation—could benchmark-based contracts reduce vendor lock-in by allowing governments to switch providers as AI evolves? Or are benchmarks too volatile, making it better to reward firms for building the internal capacity to continuously improve? Where vendors are procured before a product exists, what metrics should they be held to? Common principles also apply across scenarios, including competition, interoperability, and modularity, while newer considerations—such as technological sovereignty—are becoming increasingly important.

Ultimately, this line of work aims to help policymakers understand new directions in AI procurement, and how to design, assess, and incentivize the right offerings amid this fast-changing technology. We are exploring a convening of procurement experts, and partnering with governments, operations teams at multilateral development banks, and embedded advisory units to implement and test the recommended practices. We are currently sourcing experts from governments and multilateral development banks, and specialists in mechanism design to collaborate with. Please reach out with recommendations.