Europe and Central Asia has a striking artificial-intelligence gap. The technology can materially affect the tasks performed by roughly one in five workers, according to the World Bank, yet fewer than one in ten firms in the region use it. That is about half the share recorded in the European Union.
Exposure is not the same as adoption, and neither measure is a forecast of job losses. Exposure identifies occupations in which generative AI could automate or accelerate a meaningful share of tasks. Productivity rises only when employers integrate the tools into workflows, prepare their data and change who makes which decisions.
Broad access has produced shallow adoption
Most current use cases are narrow: drafting documents, searching internal knowledge and answering customer questions. They can save time for an individual employee while leaving approvals, quality control and handoffs unchanged. The World Bank describes this pattern as broad but shallow adoption, with little measurable productivity impact at firm level.
A faster first draft does not help much when staff still move information between incompatible systems, wait for several approvals or correct errors manually. The larger gain comes from redesigning the complete process, including which outputs can be trusted, who verifies them and how faster work changes capacity elsewhere in the organisation.
OECD data illustrate why firm size matters. In 2025, 20.2% of businesses across OECD countries reported using AI, up from 14.2% in 2024. Adoption reached 52% among large firms but only 17.4% among small firms. Bigger companies can spread the cost of data preparation, security reviews, integration and specialist staff across more transactions.
Complementary assets determine the return
Data quality is the most common internal constraint. Information may be fragmented across systems, stored in inconsistent formats or contain unresolved conflicts. A model can produce an answer quickly in that environment, but it cannot guarantee that the answer is reliable. Firms therefore worry about data leakage, factual errors and an uncertain return on investment.
Management capability matters just as much as technical skill. A useful project begins with a process that has measurable costs, cycle times, error rates or revenue outcomes. Without a baseline, companies cannot tell whether AI reduced cost, increased sales or merely changed the way employees perform the same work.
Demand for AI-related workers in the region has nearly quadrupled. Engineers alone are not enough. Managers must redesign processes, domain specialists must verify outputs, and frontline employees need to recognise poor results. Training therefore has to extend beyond software teams.
Compute access reinforces the regional divide
The World Bank estimates that data-centre capacity in the region averages 2.1 megawatts per million people, compared with 31.6 megawatts in high-income countries outside the region. The gap is roughly fifteen-fold.
Most firms do not need to own a data centre, but limited local capacity can raise cloud costs, increase dependence on foreign providers and complicate work with sensitive data. For smaller economies, shared infrastructure, compatible public clouds and workable cross-border data rules may be more efficient than trying to reproduce the full technology stack nationally.
Jobs are changing more slowly than tasks
An International Labour Organization review finds real productivity gains in some occupations, but limited evidence of large-scale displacement. Time savings are often only a few percentage points and do not automatically become higher output, earnings or employment.
The effects are uneven. AI exposure is concentrated in office-based occupations and is relatively high among younger, highly educated and female workers. Companies are likely to change task composition and hiring requirements before they reduce total headcount. Entry-level roles deserve particular attention because routine tasks that once provided early career experience may be automated first.
Scaling requires a business case, not another pilot
Firms should start with a specific process, define measures for time, quality, cost and risk, and use the pilot to identify required changes in data and decision rights. Scaling should come after those changes, not before them.
Governments can lower shared costs through digital infrastructure, machine-readable public data, training and clear liability rules. Practical examples already exist. Türkiye uses AI in monthly value-added-tax fraud risk scoring, while Ukraine is integrating the technology into its national digital-services platform.