AI Productivity Measurement in National Accounts

AI’s contribution to productivity is measured in national accounts primarily through its impact on multifactor productivity (MFP) and the quality-adjusted output of the information and communication technology (ICT) sector. In practice, statisticians do not directly measure “AI” as a separate input; instead, they capture it indirectly through the residual in growth accounting, the prices and volumes of AI-related goods and services, and the capital services from AI hardware and software.

Background: Why National Accounts Matter

National accounts are the framework countries use to measure economic activity—GDP, investment, and productivity. Productivity is typically expressed as output per hour worked (labor productivity) or as multifactor productivity (MFP), which is the portion of output growth not explained by increases in labor and capital. When AI improves processes, reduces errors, or automates tasks, its effect should appear in these aggregates. However, because AI is a general-purpose technology that diffuses unevenly across industries, its measurement is complicated by the need to separate price changes from quality changes, and by the fact that much AI activity is embedded in existing capital or in intangible assets.

How AI Enters the Production Function

In the standard growth accounting framework, output (Y) is a function of capital (K), labor (L), and MFP (A): Y = A × f(K, L). AI affects productivity through three channels:

Capital deepening: Investment in AI hardware (servers, GPUs) and software adds to the capital stock, raising output per worker even if MFP is unchanged.
Multifactor productivity: AI can improve the efficiency with which capital and labor are used—for example, by optimizing supply chains or reducing energy consumption—showing up as a rise in MFP.
Labor quality and composition: AI may complement high-skill workers, shifting the labor mix toward more productive tasks, which is captured in labor composition adjustments.

Statisticians measure the volume of AI capital by deflating nominal investment with a price index that adjusts for quality improvements. For software, this often involves hedonic pricing—modeling how the price of a software product changes with its features. For AI services (e.g., cloud-based machine learning APIs), the challenge is that they are often bundled with other services, making it hard to isolate their contribution.

Measuring AI Output: The ICT Sector and Beyond

One direct place AI shows up is in the output of the ICT sector itself. When a company sells an AI-based service—say, a speech recognition API—the revenue from that sale is counted as output. But the volume of that output must be adjusted for price changes. If the API improves in accuracy over time, the price per unit of accuracy may fall, meaning that the volume of output is higher than the nominal revenue suggests. National statistical offices use quality-adjusted price indices to convert nominal output into real (volume) terms. For AI, this is particularly difficult because the “unit” of output is not well defined—is it per API call, per hour of use, or per task completed?

Beyond the ICT sector, AI is used as an input in other industries. For example, a bank that uses AI for fraud detection may not report “AI” as a separate line item; instead, the cost of the AI service is part of its intermediate consumption. The productivity gain from AI would appear as a reduction in the bank’s costs per transaction, which would show up in the bank’s value added (output minus intermediate consumption). However, if the bank passes the savings on to customers in the form of lower fees, the measured value added may not increase, even though real output (number of transactions) rose.

Common Problems and Misconceptions

A common misconception is that AI’s contribution can be observed directly in GDP growth statistics. In reality, the effect is usually buried in the MFP residual, which is also affected by measurement error, cyclical factors, and other technological changes. Another problem is the “productivity paradox”—the observation that rapid advances in AI have not always coincided with strong measured productivity growth. This can occur because:

AI investment is often intangible (e.g., data, algorithms, organizational change) and is not fully captured in national accounts.
Price indices for AI services may overstate inflation if they do not fully adjust for quality improvements, leading to understated real output.
There are time lags between AI adoption and measurable productivity gains, as firms need to reorganize processes.

Statisticians also face the challenge of double-counting: if AI software is embedded in a physical product (like a smart sensor), it may be counted as part of the hardware’s value, not as separate software. This is why national accounts rely on detailed product classifications that attempt to separate goods and services, though the lines are blurring.

How Statisticians Actually Estimate AI’s Contribution

In practice, national statistical offices (NSOs) do not have a dedicated “AI” line item. Instead, they use a combination of methods:

Investment surveys: Firms report capital expenditure on computer software, databases, and R&D. AI-related spending is often classified under these categories, so NSOs can approximate AI investment by analyzing the composition of these expenditures.
Hedonic price indices: For products like AI software, NSOs may use hedonic regression to control for quality changes. For example, if a new version of a translation tool is 10% more accurate, the price index is adjusted so that the real output reflects the quality improvement.
Growth accounting decomposition: Once capital and labor inputs are measured, the residual (MFP) is computed. Economists then use supplementary data—such as the share of firms using AI, or patent counts—to attribute part of the residual to AI. This is done in research, not in official statistics.
Supply-use tables: These tables trace the flow of AI-related goods and services through the economy, helping to ensure that output is counted once and that intermediate consumption is correctly assigned.

For example, the U.S. Bureau of Economic Analysis (BEA) publishes statistics on “digital economy” that include software, cloud services, and data. While AI is not separately identified, the BEA has experimented with satellite accounts that attempt to measure the economic impact of AI by combining investment data with productivity estimates. Similarly, Eurostat has explored ways to capture AI in its statistics on the information society.

Comparisons and Alternatives

An alternative to the national accounts approach is to use micro-level data, such as firm-level surveys or case studies, to estimate AI’s productivity impact. These studies often find large effects in specific settings (e.g., a 20% reduction in manufacturing defects), but they are not representative of the whole economy. Another approach is to use “task-based” models, where AI is treated as a technology that automates certain tasks. These models can be used to simulate the effect of AI on labor demand and productivity, but they rely on assumptions about the cost and capability of AI that are hard to verify.

Compared to these alternatives, national accounts have the advantage of being consistent, comprehensive, and internationally comparable. However, they are also backward-looking and slow to capture new technologies. The delay between AI’s emergence and its appearance in official productivity statistics can be several years, partly because investment surveys are conducted annually and partly because it takes time for the effects to show up in output.

Costs and Data Limitations

Measuring AI in national accounts is expensive. It requires detailed surveys, sophisticated price index construction, and ongoing research to keep up with technological change. Many NSOs have limited budgets and must prioritize. As a result, some countries do not separately identify AI at all, instead lumping it into broader categories like “computer software.” This means that international comparisons of AI’s contribution to productivity should be interpreted with caution, as the underlying data may not be consistent.

Frequently Asked Questions

Does GDP include AI’s contribution?

Yes, but indirectly. If AI is sold as a service, the revenue from that sale is part of GDP. If AI is used internally to improve efficiency, the resulting increase in output (or decrease in costs) is captured in the value added of the firm, provided it is measured accurately. However, because AI is not a separate category, its contribution is not visible in the headline GDP numbers.

Why doesn’t AI show up in productivity statistics yet?

There are several reasons: measurement lags, the intangible nature of AI investment, and the time needed for firms to reorganize around AI. Also, if AI is used to reduce prices rather than increase output, the measured volume of output may not rise, even though consumers benefit.

How do statisticians adjust for AI quality improvements?

They use hedonic price indices, which model the price of a product as a function of its characteristics. For AI services, this might include accuracy, speed, or the number of tasks it can perform. By controlling for these quality changes, the price index can separate pure price changes from quality improvements, allowing real output to reflect the improved service.

Is AI counted as capital or labor?

AI is generally counted as capital, specifically as part of software and equipment investment. However, some AI-related spending, such as on data collection and cleaning, may be classified as intermediate consumption, which is not counted as investment. This can lead to an understatement of AI’s contribution to the capital stock.

Can the productivity paradox be explained by measurement issues?

Partially. If the price index for AI services overstates inflation, then real output is understated, and productivity growth appears lower than it actually is. However, measurement issues are not the whole story; there are also genuine time lags and adjustment costs that slow productivity gains.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.

More cards in this drawer

The whole Numbers and Powers drawer →