Making AI work: From promise to practice

AI is here.  Over 3 billion people globally use generative AI (genAI) like ChatGPT or Baidu’s Erni Bot monthly. AI helped researchers discover new proteins (and get a Nobel prize) and AI tutors enabled students in Nigeria catch up 2 years’ of learning progress in only 6 weeks. 

Yet so far AI’s impact on aggregate jobs or growth has been limited. Past waves of technology led to enormous progress but also left many behind. Whether AI follows suit depends on what foundations are in place. The 5th annual AI in Action conference brought together leading researchers and policymakers to identify what it actually takes to make AI work for firms and workers.

Why does AI take time to pay off?

Previous waves of technology, from the steam engine to the computer, led to enormous productivity gains, but these took decades to materialize. It took 40 years for electricity to show up in productivity data because it required a complete factory reorganization, not just a new energy source. The binding constraint was not access to technology, but the capacity to reorganize.

AI is following a similar pattern. Raffaella Sadun (Harvard Business School) calls this the J-curve. When firms adopt technologies like AI, they first incur adjustment costs from redesigning and reallocating workflows. Productivity may initially dip before gains materialize. While AI already delivers large productivity gains for narrow tasks like writing emails or summarizing documents, aggregate growth requires reorganizing how humans and AI work together.

What skills are needed in the age of AI?

Recent job cuts in high-profile firms have been blamed on AI, raising fears of widespread displacement. However, rigorous causal evidence is limited.  

Entry-level workers are feeling the effects first. Customer service workers aged 22 to 25 in the US saw a 10% decline in employment since ChatGPT launched in 2022 (Figure 1). Employment is falling in roles where AI automates tasks like software development and accounting, but rising in roles where it augments workers, like management, nursing, and maintenance. What happens at the first rung of the career ladder shapes the long-term pipeline of skilled workers.

Figure 1: Slowdown in customer services jobs for early career workers, since the launch of ChatGPTSource: Brynjolfsson et al. (2026), “Canaries in the Coal Mine: Six Facts about the Recent Employment Effects of Artificial Intelligence” Stanford Digital Economy Lab. 

To help workers adapt, embedding AI in education can narrow skills gaps. In one experiment, AI tools raised student achievement by the equivalent of pushing a B+ student into the A range, with the largest gains for lower-performing students. The trade-off is that students who relied on AI performed worse in subsequent courses without it, developing AI-specific skills at the expense of some basic human capital.

It is costly but necessary for workers to experiment with AI and learn to reorganize workflows around it. At Procter & Gamble, AI-enabled teams achieved comparable quality results to fully human teams in a fraction of the time. Yet when similar tools were introduced at a Colombian bank, veteran employees resisted. AI works best as a thinking partner, augmenting judgment and initiative. Making this shift requires time and managerial support.

AI for the few or the many?

AI models are developed in few countries, driven by massive investments in compute. But, smaller, cheaper models are converging towards frontier performance. While most frontier large language models are trained by a handful of companies in advanced economies, developing countries have an opportunity to develop smaller models or fine-tune open-source models for local contexts.

Data is often referred to as the oil of the AI economy.  Unlike oil, data creates market power through network effects where more users generate better data, improving models and attracting further users. Developing countries risk the reverse: sparse data leads to irrelevant models, discouraging adoption (see our third annual conference). Clear data governance can help break this cycle. One experiment showed that storing data under EU privacy protections largely assuaged privacy concerns, encouraging adoption.

Even within multinationals, adoption depends on organizational capacity. Most implement AI only in subsidiaries already using complementary digital technologies like enterprise resource planning systems (ERPs) and cloud computing.  

Developing countries have benefited hugely from IT or business process offshoring, but AI is increasingly automating these services.  Evidence from an online job platform shows that AI reduced offshore employment but also shifted offshoring towards higher-value, more complex tasks.

Making AI work: What can policymakers do?

First, support the skills and retraining necessary to reorganize. AI demands complex soft skills like experimentation, initiative, and judgment.  Good management is crucial for helping workers reorganize tasks and effectively adopt AI. Support should encompass investments in reorganization and training, not just technology access.

Second, remove policies that constrain reorganization. Countries can unlock greater value by opening input and product markets so ideas, capital, and labor flow to where they are most productive.

Third, access to data and models matter. Government support for open-source policies and accessible, public data can facilitate AI adaptation in local contexts.

Fourth, good institutions shape adoption. Clear, credible governance frameworks that manage risks and protect data encourage AI uptake while leaving space for experimentation and innovation.

Technology creates capacity, but reorganization ultimately determines the gains.

Source : World Bank