For three years, executives have told investors on earnings calls that artificial intelligence is about to make their companies dramatically more productive. New research from the Federal Reserve Bank of St. Louis says that transformation remains almost entirely a forecast, not a result, and the gap between the pitch and the ledger is widening even as companies lean on AI productivity assumptions to justify hiring freezes and layoffs.
What the earnings-call data actually shows
Economists Serdar Ozkan, Aakash Kalyani and Nicholas Sullivan built their analysis on 490,000 earnings call transcripts from 5,198 U.S. publicly traded companies, covering every quarter from 2000 through 2025. Rather than rely on keyword counts, the team used sentence-level extraction followed by classification through an open-source language model, Qwen3-3B, to identify roughly 910,955 sentences about productivity and tag each one for whether it mentioned AI, whether it described productivity as rising, falling or unchanged, and whether the claim was about the past, the present or the future.
The share of productivity sentences that mention AI was close to zero before ChatGPT’s late 2022 launch. It plateaued through 2024, then accelerated sharply in 2025 to reach roughly 15% of all productivity commentary on earnings calls by year end. The language shifted too: 2023 calls leaned on the phrase “generative AI,” while 2025 calls increasingly used plainer terms like “using AI” and “AI tools,” a sign the technology has moved from novelty to operating assumption in how executives describe their businesses.
The gap between the pitch and the ledger
The more striking finding is about timing, not sentiment. Ninety five percent of AI related productivity sentences describe gains that have not happened yet, compared with 75% of productivity sentences that do not mention AI. AI commentary is also more uniformly upbeat: 95% of it frames productivity as increasing, versus 75% for non AI commentary. Put together, executives are not just optimistic about AI, they are talking almost exclusively about a future that has not arrived.
Meanwhile, the economy wide number that would confirm any of it, utilization adjusted total factor productivity, grew just 0.07% over the four quarters ending in the first quarter of 2026, according to the same research. That is close to flat. “Executives predominantly describe AI related productivity effects in positive terms, suggesting a strong sense of optimism,” the authors write, before adding the caveat that matters most for anyone making workforce decisions off that optimism: “Firms may be investing, experimenting and reorganizing around AI today, while the measurable productivity effects remain mostly ahead.”
Utilization adjusted TFP is a deliberately strict yardstick. Rather than simply dividing output by hours worked, the measure, built on the Basu, Fernald and Kimball framework used by the San Francisco Fed, corrects for labor quality, capital mix and how intensively firms are actually running their workers and equipment in a given quarter. The point of the adjustment is to strip out cyclical noise, such as a company temporarily throttling back during a slow quarter, so that what remains is closer to a true read on technological progress. A near flat reading on that stricter measure is a harder result to explain away than a soft quarter in a simpler productivity statistic would be.
What this means for the HR leader
The disconnect matters because workforce decisions, not just investor messaging, are increasingly built on the assumption that AI has already delivered efficiency gains. Hiring freezes, restructuring plans and headcount reduction targets are frequently justified internally by reference to “AI productivity,” even though the Fed’s data says that phrase, on average, describes a hope rather than a measurement. A separate Culture Amp study found HR leaders’ own confidence in AI is cracking just as adoption peaks inside their organizations, which lines up with a workforce absorbing AI driven restructuring while the productivity case for it is still being written.
For HR and people analytics teams, the practical implication is a documentation gap. If a business case for eliminating roles or freezing requisitions leans on AI productivity, that claim should be traceable to an actual before and after measurement inside the company, not to a general sense that AI is helping. Few organizations currently have the instrumentation to make that comparison cleanly, since most AI productivity claims live in usage metrics, such as logins, prompts and feature adoption, rather than output metrics, such as units produced, cases closed or revenue per employee hour.
Questions worth asking before the next round of workforce cuts
Leaders weighing AI justified headcount decisions should be able to answer a short list of questions with data, not narrative. What was the team’s output before AI tools were introduced, and what is it now. Is any measured gain attributable to AI specifically, or to other changes made in the same period. Is the productivity claim based on a small pilot group or the full deployed population. And is the timeline for realizing the gain measured in months, which the Fed data would support, or is it open ended, which is closer to what earnings calls are actually describing.
The bigger signal
None of this means AI will never deliver the productivity gains executives keep promising. The share of AI related commentary keeps climbing, and the technology’s footprint in corporate language is real and growing. But the Fed’s own read of six years of earnings call data is that the industry is still in the investment and reorganization phase, not the payoff phase. Companies making people decisions as though the payoff has already landed are, by this measure, ahead of their own evidence.
For HR leaders, that is worth raising before the next restructuring memo goes out: ask for the productivity number, not the promise.
