AI Productivity Gains in G7 Economies | OECD Analysis

Why AI Could Become the Next General Purpose Technology

Artificial intelligence, particularly the emergence of generative AI and large language models since late 2022, has reignited a fundamental question in economics: can a single technology meaningfully shift aggregate productivity growth? The OECD’s June 2025 paper, “Macroeconomic Productivity Gains from Artificial Intelligence in G7 Economies,” authored by Francesco Filippucci, Peter Gal, Katharina Laengle, and Matthias Schief, provides one of the most rigorous attempts to answer this question with empirical data and structured economic modelling.

The concept of a general purpose technology (GPT) — not to be confused with OpenAI’s product naming — describes innovations that permeate entire economies, spawn complementary innovations, and create sustained productivity improvements over decades. Historical examples include the steam engine, electricity, and the internet. Each of these technologies transformed not just the sectors that created them but fundamentally reshaped how all industries operate. The OECD’s analysis examines whether AI possesses these characteristics and quantifies the magnitude of potential gains across the world’s seven largest advanced economies.

What distinguishes this report from earlier projections is its grounding in three measurable dimensions: micro-level performance improvements documented in controlled experiments, sector-by-sector task exposure analysis across G7 economies, and harmonised firm-level adoption data from national surveys. This triple foundation allows the researchers to project macroeconomic outcomes while being transparent about uncertainty at each layer. For organisations exploring how AI transforms document engagement and knowledge work, understanding these macroeconomic dynamics provides essential context for investment decisions.

OECD Micro-to-Macro AI Productivity Framework Explained

The analytical framework developed by the OECD follows a structured micro-to-macro approach that converts task-level productivity evidence into aggregate economic projections. This methodology is particularly valuable because it avoids the common trap of extrapolating from a single dimension — such as assuming that because AI can write code faster, entire economies will grow proportionally faster.

The framework operates through three sequential layers. First, the micro-level layer quantifies how much AI improves individual worker or task productivity, drawing on controlled experimental studies. The paper uses a conservative baseline of approximately 30% total factor productivity (TFP) uplift, derived from averaging results across multiple published experiments with generative AI tools.

Second, the sectoral exposure layer maps which tasks within each industry can be meaningfully augmented or automated by AI. This builds on the OECD’s earlier work with the OECD AI Policy Observatory, which has developed detailed task-level exposure assessments across occupations. The analysis distinguishes between baseline AI capabilities (current LLM performance) and expanded capabilities (LLMs combined with additional software and integrations).

Third, the adoption layer models how quickly firms actually integrate AI into their core business functions — not just experiment with it. The OECD harmonises national survey data from all seven G7 countries to create comparable adoption estimates, then projects future adoption using S-shaped logistic curves calibrated against historical technology diffusion patterns. The final output combines all three layers into annual labour productivity growth contributions under multiple scenarios, providing policymakers with a range of plausible outcomes rather than a single point estimate.

Worker-Level AI Productivity Evidence From Experimental Studies

The micro-level evidence supporting the OECD’s productivity projections draws from a rapidly growing body of controlled experiments and field studies. These studies measure how access to generative AI tools affects worker output, quality, and efficiency across different tasks and skill levels.

Customer service applications have shown productivity improvements of approximately 14% when workers use AI-assisted tools, with the gains concentrated among less experienced workers who benefit most from AI-generated suggestions. Software development studies report more dramatic results, with some experiments documenting productivity gains of up to 56% for specific coding tasks when developers use AI pair-programming tools like GitHub Copilot.

The OECD adopts a conservative baseline of 30% as the representative micro-level TFP uplift for their modelling, acknowledging that individual study results vary widely depending on the task, worker skill level, and AI tool quality. This conservatism is deliberate — the researchers note that many published studies examine narrow, well-defined tasks in controlled settings, and the gains may not fully translate to complex, multi-step business processes where AI integration requires significant workflow redesign.

An important methodological nuance is whether micro studies measure true total factor productivity improvements or merely labour cost reductions. If the gains primarily represent workers completing the same output in less time rather than producing more output with the same resources, the macro-level implications differ. The paper notes that converting labour productivity gains to TFP may require scaling by the labour share of income (approximately 60%), which would reduce the projected macro-level effects. This distinction matters significantly for understanding how AI reshapes knowledge work across entire organisations.

Sectoral AI Exposure Across G7 Industries

Not all sectors of the economy face equal exposure to AI augmentation. The OECD’s sectoral analysis reveals stark differences in how deeply AI can penetrate different industries, creating a landscape where productivity gains will be concentrated in specific parts of the economy while leaving others largely unaffected.

Knowledge-intensive service sectors show the highest AI task exposure. Finance and insurance leads with approximately 70–80% of tasks potentially augmented by expanded AI capabilities. ICT services, including software development and data processing, show similar levels of exposure. Publishing, media, and professional services — encompassing consulting, legal, and accounting firms — round out the most-exposed sectors with 50–70% of tasks within the scope of AI augmentation.

At the other end of the spectrum, sectors that rely heavily on physical activity, environmental variability, and manual dexterity show much lower exposure. Agriculture, mining, and construction have only 10–30% of their tasks exposed to AI, primarily in planning, logistics, and administrative functions rather than core physical operations. Manufacturing occupies a middle position, with significant variation depending on the specific subsector and the degree of automation already in place.

The economic significance of these exposure differences becomes clear when examining the value-added share of highly exposed sectors across G7 countries. The United States has approximately 25% of its economic value added in the top five most AI-exposed sectors, reflecting its large financial services, technology, and professional services industries. The United Kingdom follows closely at 23%. In contrast, Japan, Italy, and Germany — countries with relatively larger manufacturing, agriculture, and construction sectors — have approximately 15% of value added in the most AI-exposed sectors.

This structural composition explains much of the projected variation in AI productivity gains across countries. Nations with economies skewed toward knowledge-intensive services are positioned to capture disproportionately larger benefits from AI, a finding with significant implications for industrial policy and international competitiveness.

AI Adoption Rates Across G7 Countries in 2024

Measuring AI adoption at the firm level is one of the most challenging empirical tasks in contemporary economics. The OECD’s contribution to this measurement challenge is a harmonised framework that makes adoption estimates comparable across G7 countries despite differences in national survey methodologies and definitions.

The preferred estimates focus specifically on high-intensity adoption — firms that have integrated AI into core business functions, not merely experimented with it or used it for peripheral tasks. By this demanding standard, the 2024 adoption landscape across G7 economies reveals significant variation.

The United States leads at 6.1%, reflecting its mature technology ecosystem, deep venture capital markets, and large technology workforce. Canada follows at 5.7%, benefiting from proximity to the US market and strong AI research institutions including the Mila – Quebec AI Institute and the Vector Institute. The United Kingdom registers 5.4%, supported by London’s position as a global financial and technology hub.

Germany’s 4.8% reflects strong industrial adoption, particularly in manufacturing and automotive sectors that have begun integrating AI into production processes and supply chain management. France at 3.1% shows a growing but more cautious adoption pattern. Italy at 2.2% and Japan at 1.9% occupy the lower end of the range, though for different reasons — Italy faces structural challenges in digital infrastructure and firm size (many SMEs), while Japan’s lower adoption rate reflects a more conservative corporate culture toward new technology deployment despite the country’s strong technical capabilities.

These adoption differences compound over time. Under the OECD’s fast adoption scenario — modelled after mobile phone diffusion curves — the United States could see AI adoption grow by 57 percentage points over the next decade, potentially reaching approximately 63% of firms using AI at high intensity. Italy, starting from a lower base, might add 43 percentage points but would reach only approximately 45% adoption over the same period.

Modelling Future AI Adoption With S-Shaped Diffusion Curves

Predicting how quickly firms will adopt AI requires historical analogies, since AI’s full diffusion trajectory has yet to unfold. The OECD calibrates its adoption projections using S-shaped logistic curves — the mathematical pattern that describes how most technologies spread through an economy: slowly at first, accelerating through a middle phase of rapid adoption, then plateauing as the technology reaches saturation.

Three historical technology benchmarks anchor the OECD’s scenarios. The pessimistic scenario mirrors the slow adoption of electricity in the early twentieth century, where only 23% of US firms had adopted the technology 10 years after a user-friendly commercial breakthrough. The medium scenario follows the trajectory of personal computers and the internet, which reached approximately 40% adoption within a decade of becoming commercially accessible. The optimistic scenario maps to mobile phone diffusion, the fastest major technology adoption in modern history, reaching 60% adoption within 10 years.

The choice of ceiling matters enormously for long-term projections. The OECD sets a long-run adoption ceiling at 80%, reflecting the assumption that some firms will never adopt AI at high intensity due to industry characteristics, firm size, or strategic choices. This ceiling is deliberately below 100% — not every business in every sector will find transformative AI use cases, and adoption requires ongoing investment in skills, data infrastructure, and organisational change.

One factor favouring faster adoption compared to historical technologies is the sharp decline in AI model costs. Quality-adjusted costs have fallen approximately 80% over the past two years according to research by Andre et al. (2025). This cost trajectory more closely resembles the rapid cheapening of internet access in the late 1990s than the slower cost decline of earlier technologies. Combined with the increasing ease of deploying AI through cloud services and pre-trained models, the cost barrier to adoption is falling faster than for any previous general-purpose technology.

Projected Macroeconomic AI Productivity Impacts by Country

The core output of the OECD’s analysis is a set of projected annual labour productivity growth contributions from AI across G7 economies over a ten-year horizon. These projections combine the micro-level evidence, sectoral exposure maps, and adoption trajectories into country-specific estimates under three scenarios.

Under the pessimistic scenario — assuming slow (electricity-like) adoption and baseline AI capabilities — the projected annual productivity contribution ranges from approximately 0.2 percentage points for Japan and Italy to approximately 0.4 percentage points for the United States and United Kingdom. Even this conservative estimate represents a meaningful addition to labour productivity growth, which has averaged well below 2% annually in most G7 economies since the 2008 financial crisis.

The central scenario — combining medium (computer/internet-like) adoption speed with expanded AI capabilities — projects annual gains of approximately 0.5 to 1.0 percentage points. Countries with both high sectoral exposure and faster adoption, principally the United States and United Kingdom, cluster toward the upper end of this range. Germany and Canada fall in the middle, while France, Italy, and Japan occupy the lower end but still show substantial absolute gains.

The optimistic scenario — featuring fast (mobile phone-like) adoption and expanded AI capabilities — projects annual productivity contributions of 0.8 to 1.3 percentage points. At the upper end, this would approach the productivity contribution of the entire ICT boom in the United States during the mid-1990s, which added approximately 1.0 to 1.5 percentage points to annual labour productivity growth. This comparison provides both context and a reality check: achieving ICT-boom-level productivity gains from AI would represent a historically significant technology-driven acceleration.

Cross-country differences in projected gains are driven by two primary factors. First, economic structure matters — countries with larger knowledge-intensive service sectors have more of their economy exposed to AI augmentation. Second, adoption speed matters — countries with higher current adoption rates and more supportive institutional environments tend to project faster future adoption. The interaction of these two factors creates a landscape where the United States and United Kingdom consistently project the largest gains, while Italy and Japan project the smallest.

AI Cost Decline and Accelerating Technology Diffusion

One of the most striking findings supporting the case for rapid AI adoption is the pace of cost decline in AI model deployment. Research by Andre et al. (2025), cited extensively in the OECD report, documents that quality-adjusted AI model costs fell by approximately 80% over just two years. This rate of cost reduction is exceptional even by the standards of digital technology, where Moore’s Law drove decades of declining computing costs.

The cost decline operates through multiple channels. Foundation model providers have dramatically reduced API pricing as competition intensifies among providers including OpenAI, Anthropic, Google, Meta, and emerging players like Mistral AI. Simultaneously, open-weight models have become competitive with proprietary offerings, providing firms with lower-cost alternatives that can be deployed on their own infrastructure. Hardware improvements, particularly in GPU efficiency and the emergence of specialised AI accelerator chips, further reduce the cost of running AI workloads at scale.

For firms, lower AI costs translate directly into higher potential returns on AI investment, lowering the threshold at which adoption becomes economically rational. The OECD emphasises that this cost trajectory supports faster diffusion than assumed in the pessimistic scenario and may even exceed the medium scenario’s assumptions. If costs continue declining at current rates, the economic case for AI adoption becomes compelling even for smaller firms and less technology-intensive sectors that have historically been slow to adopt new digital tools.

However, the report cautions that model costs are only one component of the total cost of AI adoption. Integration costs — including data preparation, workflow redesign, employee training, and ongoing maintenance — often exceed the cost of AI models themselves. These complementary investments create a significant gap between having access to affordable AI tools and actually deploying them productively in core business functions. The firms that capture the largest AI productivity gains will be those that invest systematically in these complementary capabilities alongside the technology itself, a pattern consistent with historical experience from previous technology adoptions documented by the National Bureau of Economic Research.

Policy Recommendations for Maximising AI Productivity Gains

The OECD’s analysis carries clear implications for government policy across G7 economies. The researchers identify five priority areas where policy intervention can meaningfully accelerate AI-driven productivity gains while managing associated risks.

First, the report emphasises the urgent need for harmonised, comparable official statistics on AI adoption. Current cross-country comparisons rely on different survey methodologies, definitions of AI use, and sampling frameworks, making it difficult to benchmark progress or evaluate policy effectiveness. The OECD recommends developing a standardised statistical framework for measuring high-intensity AI adoption in core business functions, similar to existing frameworks for ICT usage surveys.

Second, investment in digital infrastructure — including broadband connectivity, data centre capacity, and access to affordable cloud computing — remains essential. Countries with weaker digital infrastructure face structural barriers to AI adoption regardless of their firms’ willingness to invest in new technology. This is particularly relevant for Italy, where digital infrastructure gaps contribute to lower adoption rates despite the country’s strong industrial capabilities.

Third, workforce development requires both expanded STEM education and comprehensive reskilling programmes. The productivity gains from AI depend not just on having the technology available but on having workers who can use it effectively. This means investing in AI literacy across all skill levels, from executive decision-makers who need to understand AI’s strategic implications to frontline workers who need practical skills in AI-augmented workflows.

Fourth, supporting SME adoption through targeted policy instruments — including grants, vouchers, and technical assistance programmes — can help bridge the gap between large firms (which adopt AI faster) and smaller firms (which often lack the resources and expertise for effective integration). Given that SMEs account for the majority of employment in most G7 economies, their adoption rates significantly influence aggregate productivity outcomes.

Fifth, data governance and competition policy must balance innovation incentives with protections for privacy, safety, and market fairness. The OECD recommends ensuring data portability and competitive access to training data, preventing excessive concentration of AI capabilities, and maintaining regulatory environments that encourage responsible innovation. The EU AI Act provides one model for risk-based regulation, though the OECD emphasises that regulatory approaches must be flexible enough to adapt as AI capabilities evolve.

Limitations and Future Research Priorities for AI Economics

The OECD report is transparent about the substantial uncertainties embedded in its projections. Several limitations deserve attention from policymakers and researchers interpreting the results.

The micro-level evidence, while growing rapidly, remains concentrated in specific tasks and settings. Most published experiments examine relatively simple, well-defined tasks — writing customer service responses, generating code snippets, summarising text — rather than complex, multi-step business processes. The extent to which task-level gains translate into firm-level and sector-level productivity improvements remains an open empirical question.

Adoption measurement, despite the OECD’s harmonisation efforts, still relies on survey responses that may not accurately capture the depth and quality of AI integration. A firm that reports using AI in core business functions may be using it superficially for a single process, while another firm in the same survey category may have deeply integrated AI across multiple workflows. This measurement imprecision propagates through the modelling framework into the macro-level projections.

The report also acknowledges that AI-driven productivity gains may not be evenly distributed across workers, firms, or regions. Concentration of gains among high-skill workers and large firms could exacerbate inequality, and sector-specific productivity improvements may create Baumol-type cost disease effects — where sectors experiencing slower AI-driven productivity growth see relative cost increases that partially offset aggregate gains.

Future research priorities identified by the authors include developing better firm-level panel data on AI adoption and productivity outcomes, conducting more field experiments in realistic business settings, and building dynamic general equilibrium models that can capture the feedback effects between AI adoption, labour market adjustment, and capital reallocation. The paper also calls for more research on the role of complementary investments — skills, data governance, organisational redesign — in determining whether AI’s theoretical productivity potential is actually realised in practice. For organisations committed to making complex research accessible, these findings highlight the growing importance of bridging the gap between academic evidence and practical implementation.