Why AI’s Productivity Gains Are Invisible in India’s Official Numbers

Artificial intelligence is now embedded in everyday work across the world. Lawyers draft contracts with it, programmers debug code with it, consultants prepare presentations with it, students learn with it, and bureaucracies quietly experiment with it. India is no exception. From startups in Bengaluru to Government departments piloting AI-enabled systems, adoption is rising steadily.And yet, when economists and policymakers look at the headline numbers, the story appears strangely flat. Productivity growth remains modest. GDP has not surged. Wages have not jumped. The long-promised “AI dividend” seems delayed, if not absent. This has led to a familiar conclusion: perhaps artificial intelligence is being oversold. A more uncomfortable possibility deserves attention. The problem may not be AI. The problem may be how we measure productivity itself. Modern productivity statistics were designed for a very different economic world — one in which intelligence was scarce, slow, and inseparable from human labour. AI quietly violates all three assumptions. As a result, some of its most important effects remain largely invisible to our economic instruments.

Consider what the data show

In advanced economies such as the United States, labour productivity growth has hovered around 1–1.5 per cent annually over the past decade, with no sustained acceleration even after the release of powerful generative AI tools in late 2022. Yet adoption has been extraordinarily rapid. A 2024 working paper from the National Bureau of Economic Research found that nearly 28 per cent of U.S. workers were already using generative AI at work within 18 months —one of the fastest technology diffusion curves ever recorded.

High adoption. Modest productivity growth. That is the puzzle

At the level of individual tasks, however, the effects of AI are anything but modest. Field experiments conducted by researchers associated with MIT examined customer-support workers using generative AI tools. Average productivity rose by about 14 per cent. For newer and less experienced workers, gains reached 30–35 per cent. Employee attrition fell. Response quality improved.

These are large effects by historical standards. Yet overall firm output barely moved. Companies used the gains to reduce training time, standardise responses, and shorten resolution cycles rather than to increase the number of calls handled.

Output stayed the same. The system changed

A similar pattern emerged in a 2023 Harvard — Boston Consulting Group experiment involving management consultants. With access to advanced AI tools, consultants completed tasks 25–40 per cent faster and produced higher-quality work. But total billable output did not rise proportionally. Firms absorbed the gains as time savings, error reduction, and smoother coordination.

From the perspective of GDP, little happened. From the perspective of organisations, work was reorganised. India is already seeing this pattern. Public disclosures and executive commentary from large Indian IT services firms suggest that generative AI is being used extensively to reduce turnaround times in coding, testing, and customer support. In 2024, senior executives across major firms indicated that AI tools were cutting task completion time by roughly 20–30 per cent in selected service lines, even as overall revenue growth remained subdued. The gains were absorbed as efficiency, quality control, and margin protection rather than expanded output. From a national accounts perspective, little changed. From an organisational one, workflows were being quietly rewritten.

This distinction matters because artificial intelligence primarily compresses time, not labour. It reduces the latency involved in thinking, drafting, searching, summarising, and deciding.

Productivity statistics, however, are designed to capture changes in output per hour worked—not changes in how quickly uncertainty is resolved.

In physics, choosing the wrong variable to measure does not produce a slightly wrong answer; it produces a misleading one. The same principle now applies to economic measurement.

Another blind spot lies in how averages dominate productivity analysis. AI’s most consistent impact is not on the average level of performance, but on its distribution. Study after study shows that AI narrows performance gaps. Lower-performing workers improve the most. Error rates fall. Variability shrinks. These changes matter enormously for organisations, which care about reliability, predictability, and worst-case outcomes. But productivity statistics largely ignore them. They track averages, not distributions.

There is also a historical lesson worth recalling. During the early decades of electrification in the late 19th and early 20th centuries, productivity growth remained stubbornly low. Factories initially treated electric motors as drop-in replacements for steam engines, preserving old layouts and workflows. Only when production was reorganised around electricity did productivity surge — decades later.

In 1987, economist Robert Solow famously observed that computers were visible everywhere except in the productivity statistics. The computer-driven productivity boom arrived only in the late 1990s, once organisations, management practices, and skills caught up. Artificial intelligence appears to be following a similar trajectory. It is reorganising work faster than it is increasing measurable output.

This has direct implications for India

India’s economic strategy increasingly rests on digital public infrastructure, services exports, and the productivity of its vast working-age population. AI is already being deployed in software development, logistics, education technology, and parts of Government. But much of its early impact will appear as faster workflows, fewer errors, and improved coordination rather than higher output per worker.

If policymakers rely solely on headline productivity numbers, they risk missing early signals of structural change. Worse, they may conclude prematurely that AI is economically insignificant — leading to underinvestment in skills, power infrastructure, data systems, and institutional reform. There is a second risk as well. If AI-driven gains are absorbed as invisible efficiency — shorter decision cycles, tighter feedback loops, fewer mistakes —the benefits may remain concentrated within firms rather than translating into wages or employment growth.

That is a distributional challenge, not a technological one, and it requires policy attention. The solution is not to abandon productivity statistics.

It is to recognise their limits. Artificial intelligence does not behave like earlier labour-saving machines. It behaves more like cognitive infrastructure — closer to electricity or the internet than to a mechanical tool.

When intelligence becomes abundant and cheap, measuring its impact through labour-based accounting alone becomes increasingly misleading.

The AI era is unlikely to announce itself through a sudden explosion in GDP. It will announce itself through something quieter and harder to capture: faster decisions, fewer errors, tighter coordination, and organisations that tolerate far less inefficiency than before.

That creates a political and policy risk. When official statistics fail to register real change, Governments are tempted to conclude that nothing important is happening. Investment is delayed. Reforms are postponed. Skepticism hardens into complacency. The danger is not that AI will fail to transform the economy. The danger is that it already is — while our measurement systems insist that it hasn’t.

Author is a theoretical physicist at the University of North Carolina at Chapel Hill, US, and the author of the forthcoming book The Last Equation Before Silence; views are personal