AI productivity payoff may arrive later than investors expect

Artificial intelligence may take years longer to lift broad productivity than many investors expect, according to Goldman Sachs economist Elsie Peng. In a research note, Peng compared the current AI buildout with the computer revolution and found that technology spending alone did not quickly show up in economy-wide data.

Goldman Sachs still expects AI to raise productivity growth meaningfully over the next decade, according to the bank’s view cited by Peng. Her warning is about timing: the personal computer took far longer to affect measured output than its early backers expected.

The PC comparison

Peng wrote that the PC was commercialized in 1981, while investment in information and communications technology rose across industries in the early 1980s. Her industry analysis found a J-curve pattern: productivity was a small drag for the first four years, did not produce statistically significant gains until year eight and reached a peak effect of about 0.6 percentage points in year 12.

The broader productivity boom associated with personal computers did not appear in macroeconomic data until 15 years after commercialization, according to Peng. If ChatGPT’s 2022 launch is treated as a similar starting point, her framework suggests a measurable productivity payoff around 2030 and a peak around 2034.

Peng identified three reasons for the earlier delay. Goldman Sachs said key components, including semiconductors and telecom equipment, stayed costly through the 1980s, while applications such as the internet needed broad adoption before they created much value. The largest constraint, according to Peng, was the work of changing organizations so they could use the technology effectively.

Goldman Sachs estimated that each dollar spent on ICT hardware required at least $1.70 in related intangible investment, including software, data systems and organizational redesign. Peng said that spending did not accelerate until the mid-1990s, about a decade after PCs began reaching desks.

AI spending is uneven

Goldman Sachs data cited by Peng shows AI hardware investment rising faster than ICT investment did at a similar stage. Spending on work-process changes appears to be moving more slowly than it did during the 1990s computer cycle, though Goldman said some activity may be missing from official statistics.

An Atlanta Fed survey cited by Goldman implied about $280 billion in AI-related intangible spending in 2026. Even with that measurement gap, Peng said the reorganization side is trailing hardware investment by more than it did in the PC era.

Worker resistance may add another delay. An April survey of 2,400 knowledge workers by Writer and Workplace Intelligence found that 29% of employees said they had actively undermined their company’s AI strategy. Among Gen Z workers, the figure was 44%, up from 41% a year earlier, according to the survey.

A separate WalkMe survey of executives and employees in 14 countries found that more than 54% of workers had avoided company AI tools in the previous 30 days and done tasks manually instead. The survey commissioners attributed much of the resistance to fear of becoming obsolete.

Harvard Business School researchers have described a related behavior as “symbolic adoption,” in which employees appear to comply with AI rollouts while quietly limiting their use. In the Writer survey, 30% of self-described AI saboteurs said they did not want AI to take their job, and 26% said it reduced their sense of value or creativity at work. The same survey found that 69% of executives said their companies were already carrying out AI-related layoffs.

Labor data and market risk

Stanford’s Erik Brynjolfsson and ADP Research have tracked 4.6 million workers across more than 730 occupations through the Canaries Dashboard. Their data shows employment for workers ages 22 to 25 in AI-exposed occupations shrinking more than 4% annually, a pattern they said is not visible in broad labor-market figures.

Apollo Global Management chief economist Torsten Slok has warned that a slower AI payoff could create wider market risks because AI has supported both the economy and equities. On Apollo’s Daily Spark blog, Slok wrote that delayed returns could threaten a recession and an S&P 500 correction.

Peng’s conclusion is that AI may still be productive, but broad gains depend on companies changing how work is done. Her comparison with the PC era suggests that the bottleneck is less about buying hardware than about the slower task of rebuilding organizations around new tools.

This story draws on original reporting from Fortune.