Source: Fortune, covering St. Louis Fed research — 2026-07-31
Summary
Researchers at the Federal Reserve Bank of St. Louis used an LLM to analyze roughly 490,000 corporate earnings-call transcripts from 5,198 publicly traded U.S. companies, spanning 2000 to 2025, flagging every sentence about productivity and classifying it as past, present, or future-tense. They found that about 95% of AI-related productivity mentions describe gains executives expect in the future rather than gains already realized, and that share has held steady since 2023. Separately, AI's share of all productivity-related sentences went from near zero before ChatGPT, climbed through 2023, plateaued in 2024, then jumped to about 15% by the end of 2025.
Key Takeaways
- The study covers a large dataset: ~490,000 earnings-call transcripts across 5,198 U.S. public companies from 2000 through 2025, using an LLM to classify productivity sentences as realized, present, or future-tense.
- About 95% of AI-related productivity commentary is forward-looking (executives talking about expected future gains), not describing results already achieved — and that ratio hasn't shifted meaningfully since 2023 despite growing AI adoption.
- AI's share of all productivity-related sentences rose from near zero pre-ChatGPT, climbed through 2023, plateaued through 2024, and accelerated again to roughly 15% of all such sentences by the end of 2025.
- Researchers frame the pattern as consistent with prior general-purpose technologies like electrification, where the measurable productivity payoff has historically lagged years behind the hype and adoption curve.
Reel Script
Hook (~15-20s, 35-45 words) The Fed just read almost half a million earnings calls to answer one question: is AI actually making companies more productive yet, or are executives just talking about it? The answer is not what the hype cycle wants you to hear.
Core Concept (~45-90s, 105-200 words) St. Louis Fed researchers took roughly 490,000 earnings-call transcripts from over five thousand public companies, going back to the year 2000, and ran them through an LLM. The model's job: find every sentence about productivity, then tag it as describing a gain already achieved, a gain happening now, or a gain expected in the future. That's a clever use of AI to study AI — instead of guessing at corporate sentiment, they mechanically classified tense across hundreds of thousands of sentences. The headline finding is that the overwhelming majority of AI-productivity talk is still future tense — executives promising gains, not reporting them. And that's been true consistently since 2023, meaning as more companies started talking about AI, the ratio of promise to proof didn't change. The researchers connect this to a well-documented historical pattern: general-purpose technologies like electrification took years, sometimes decades, between widespread adoption and a measurable productivity payoff showing up in the data.
Hands-On (~45-150s, 105-350 words) Picture this as a timeline you could sketch on a whiteboard. Before ChatGPT launched, AI's share of all productivity-related sentences in earnings calls was basically at zero — nobody was connecting the two. Once ChatGPT came out, that share started climbing through 2023 as executives began weaving AI into how they talked about efficiency and output. Then in 2024, that climb flattened out into a plateau — companies had said their piece and weren't escalating the AI-productivity talk further. But in 2025, the line kicks up again, accelerating to reach about 15% of all productivity-related sentences on earnings calls by year's end. Now overlay the second number on that same chart: throughout this entire period, roughly 95% of the AI-specific productivity sentences were future-tense — "we expect," "we anticipate," "this will enable" — rather than "we achieved" or "we're seeing now." That 95% figure barely moved even as the overall volume of AI talk surged in 2025. So you have rising volume of AI-productivity chatter, but a flat, dominant future-tense bias underneath it the whole time. That gap between talk and realized results is the actual story here, not the raw growth in mentions.
Takeaway (~20-30s, 45-70 words) More AI mentions on earnings calls doesn't mean more AI results — it means more AI promises, and that ratio has barely budged in three years. History says the payoff eventually shows up, but it's slower than the marketing suggests. Next earnings season, listen for "expect" versus "achieved" before you believe the productivity story.