From AI pilots to scale: Improving the delivery of public services

How can AI be used to deliver better public services? That’s one of the big questions for developing countries. Governments are already experimenting with AI in areas ranging from tutoring and medical diagnosis to weather forecasting and administrative tasks, and some applications are yielding tangible benefits. In Ghana, the Rori math tutor, delivered by text message on basic cell phones, produced nearly a full year of learning gains in mathematics at US$5 per student. In Bangladesh, AI-supported medical imaging increased the number of patients screened for diabetic retinopathy by 39.5 percent per day.

For most governments, the next step is the challenge: How to move from relatively small-scale experiments with AI to identifying the applications that create value and expand their rollout from a few hundred users to millions. Building that capacity requires improving systems to buy, test, evaluate, and integrate AI solutions into public service delivery – and that is what can help them move from pilots to scale. The World Development Report 2026: The Promise of Artificial Intelligence articulates a sequence of five steps.
 

Step 1: Match AI to the problem

AI is not the right answer to every problem. Putting AI on top of a poorly designed process can simply automate existing inefficiencies. Governments should start with three basic questions:

• What problem are we trying to solve?
• Would AI improve the outcome compared with simpler alternatives, and at what cost?
• What complementary changes, such as better data, redesigned workflows, skills, or connectivity, would be needed?

The right opportunity also depends on which type of AI is being considered. Revenue authorities around the world have used machine learning algorithms to analyze historical data on tax returns to uncover compliance risks that manual audits missed. Predictive AI models like this suit agencies that have large, structured data sets and clearly defined outcomes to predict. Generative AI can be better suited to language-intensive tasks, such as drafting, translating, or answering citizens’ queries.
 

Step 2: Build AI-ready data

In Brazil, the VICTOR system that converts scanned court documents into machine-readable text has reduced the amount of time it takes to determine eligibility for an appeal to the Brazilian Supreme Federal Court from 40 minutes to just 5 seconds. But this is the exception rather than the norm. Governments across developing economies identify data quality and availability as the most common barriers to using AI in their internal operations (see figure 1). Getting AI solutions off the ground will require investments to build AI-ready data: from digitizing information—drawing on censuses, tax records, birth and death registries, land records, and social programs—to improving the use of management information systems and connecting data across government agencies. 

Step 3: Make pilots a pathway to evaluating AI solutions

Pilots are a good start if they generate evidence on what should be scaled and what shouldn’t. Too often, they don’t. Of nearly 10,000 studies of AI-based health interventions, only 57 were tested with real patients and providers in low-resource settings. Only 17 of those measured causal effects on health, quality of care, or efficiency. The question of evaluation also goes beyond whether a model performs well in a technical demonstration. In Thailand, an AI tool used to screen for diabetic retinopathy rejected more than 20 percent of the images nurses took of patients because the lighting and image quality in the clinic did not match the conditions that the model had been trained on. In fact, 80 percent of governments surveyed across low- and lower-middle-income economies report having no or only basic mechanisms in place to evaluate AI performance.
 

Step 4: Move from projects to products in government procurement

Scaling AI also requires governments to rethink how they purchase technology. Traditional government procurement treats technology like a one-off project, where the agency specifies requirements, awards a contract, and closes the project once the system is delivered. But AI systems require ongoing monitoring, maintenance, testing and improvement as models change, new vulnerabilities emerge, and user needs evolve. As former U.S. Deputy Chief Technology Officer (CTO) Jennifer Pahlka has argued, most technology solutions could deliver 85 percent of required functionality at 10 percent of typical project cost if governments started smaller and improved over time. Moving towards a product model also involves assigning clear ownership and funding to carry out monitoring, evaluation and improvement beyond the initial contract. Contracts should also require:

• Data portability — so agencies are not locked into a single vendor
• Interoperability — so AI tools can connect with existing government systems
• Auditability — agencies can review how decisions are made and correct errors over time
  

Step 5: Invest in capacity to sustain the AI solution over time

Governments need people with skills to buy, deploy, maintain, and improve AI systems over time. The skills most needed to do this are understanding user workflows, evaluating vendor claims, and safely connecting government data to AI tools. That means stable funding, internal teams with the authority and technical capacity to manage the system, and partnerships with experienced practitioners from universities, research institutes, and civil society. Building AI literacy—understanding what AI can and cannot do, when its use is appropriate, and how to interpret its outputs—across the broader civil service will also matter.  

Taken together, these steps can help governments build the capacity to distinguish promising AI applications from hype, adapt them to local realities, and sustain them after the pilot team moves on. Moving from pilots to scale means moving from pilots to evidence, from evidence to institutions, and from institutions to lasting public value.

Source : World Bank