St. Louis Fed: AI spending surges but productivity gains stay invisible

A new Federal Reserve Bank of St. Louis study analyzing roughly 490,000 corporate earnings calls confirms that artificial intelligence has not delivered a measurable increase in aggregate productivity. Despite the technology dominating boardroom discussions since late 2022, official economic data continues to show no meaningful bump once capital investment is accounted for.

By the end of 2025, AI accounted for roughly 15% of all productivity commentary on earnings calls, up from near zero before ChatGPT's launch. However, about 95% of those sentences describe expected future gains rather than realized improvements, a ratio that has held steady since 2023. Executives remain overwhelmingly bullish, with 95% of AI-related commentary predicting rising productivity compared to 75% for non-AI topics.

For investors, the critical question is whether this optimism translates into tangible capital allocation. The St. Louis Fed researchers found that it does, noting a strengthening correlation between positive AI commentary and increased R&D and capital expenditures. “we trust what people, do not what they say,” said Aakash Kalyani, one of the paper's authors.

A separate San Francisco Fed study corroborates this, showing AI-positive firms posted substantially higher investment and R&D growth by 2025, though this spending remains heavily concentrated among the largest technology companies building infrastructure. Previous St. Louis Fed research estimated generative AI drove only a 1.1% productivity increase by late 2024. That modest figure compares to overall productivity growth of 2.3% in 2024 and 1.6% in 2023.

The abundance paradox

Beyond the standard lag in technology adoption, the paper's authors highlight a structural risk unique to AI. The technology may be generating real efficiency gains that are inherently invisible to macroeconomic statistics because it simultaneously destroys the value of the output it creates. “Some things are going to become more abundant,” said Serdar Ozkan, another author of the paper. “That means they’re also going to become probably less valuable.”

If AI makes a task radically cheaper to produce, the resulting abundance drives down the price of that output. The mathematical gain in production efficiency gets erased by falling prices, meaning a real economic improvement vanishes from aggregate statistics without ever registering as a loss.

A decades-long wait

The current delay mirrors a historical pattern for general-purpose technologies. The researchers compare AI to electrification and early computers, which took decades to reorganize workflows and retrain workers before appearing in productivity data. “The aggregate gains will be in the future, whereas what you see right now is a lot of investment and a lot of excitement and optimism for the future,” Kalyani said.

Historically, technology diffusion unfolds over 20 to 30 years, and compressing AI's timeline to just three to five years would break sharply with precedent. Other Fed research aligns with this slow-burn thesis. A Kansas City Fed analysis found recent productivity gains are not yet broad-based, while Fed Chair Kevin Warsh recently told Congress that while AI has made workers “a bit more productive,” “the long term can be quite far out.”

Furthermore, AI cannot eliminate all operational bottlenecks. While the technology accelerates research and drafting, physical and scheduling constraints still move at human speed, limiting aggregate output. Kalyani noted that the specific applications which will ultimately drive measurable productivity are discovered through a decentralized, unpredictable trial-and-error process.