I spend most of my working hours inside a database of 247,000 earnings call transcripts. It has changed how I think about market narratives, because I no longer have to guess how a theme is spreading — I can watch it.
Here is the experiment. Take one phrase, "agentic AI", and run the same full-text query over every earnings call in the archive, quarter by quarter, from the first time anyone said it until today. No sampling, no "top 20 tech names", no vibes. Every call, every company, every sector.
The first mention was a chip company nobody was watching
The earliest hit in the archive is from May 2023: Ambarella, a mid-cap semiconductor company, months before the phrase showed up anywhere else. In all of 2023 the phrase appears in exactly 5 earnings calls. HubSpot and Zoom picked it up that November. NVIDIA — the company most people would guess said it first — doesn't use it on a call until January 2024.
That alone is worth pausing on. If you only read the megacap calls, you would have dated the narrative a year late and attributed it to the wrong companies.
Then the curve did what narrative curves do
Companies mentioning "agentic AI" on at least one call, per quarter:
| Quarter | Companies | Calls |
|---|---|---|
| H2 2024 | 45 | 59 |
| Q1 2025 | 147 | 190 |
| Q2 2025 | 174 | 219 |
| Q3 2025 | 287 | 333 |
| Q4 2025 | 384 | 446 |
| Q1 2026 | 494 | 583 |
| Q2 2026 | 463 | 555 |
Across the whole run, that's roughly 2,400 earnings calls quoting the theme. Each row is one API request:
curl 'https://earningscalls.dev/api/v1/search/by_ticker?q="agentic+AI"&date_from=2026-04-01&date_to=2026-06-30' \
-H "X-API-Key: $KEY"
Two things in that table would be invisible from any single call.
First, the diffusion pattern. In Q2 2026 the sector split looks like this: Information Technology 210 companies, Financials 74, Industrials 54, Consumer Discretionary 43, Communication Services 40, Health Care 28 — and a tail that includes Real Estate, Consumer Staples, Energy, Utilities and Materials. That is all eleven GICS sectors. The theme left tech about four quarters ago, and it left in a specific order: software first, then the banks, then everyone who sells to consumers.
The banks are my favorite part, because they aren't bragging — they're worrying. Regional banks like KeyCorp and U.S. Bancorp got asked, in the same reporting week, what happens to deposit pricing when a customer's AI agent can sweep cash to the highest-yielding venue automatically. When two banks field the identical question days apart, that's not two coincidences. That's the sell side converging on a consensus risk, and you can watch it happen in the transcripts. Meanwhile Walmart's analysts ask about agentic AI the way they asked about cloud migration in 2017. I pulled ten of the best verbatim quotes in an earlier piece if you want the raw material.
Second, the plateau. Q2 2026 is the first quarter where the count went down — 494 companies to 463. One quarter isn't a trend, and Q2 reporting always skews the mix. But if you hold a position that leans on "AI narrative still accelerating", this is exactly the number you want on a chart, refreshed every quarter. Mention counts flatten before headlines do. Management teams decide what goes into prepared remarks based on what they believe will move the stock; when they collectively stop adding a phrase, that decision is the signal. It's the first of the five text signals quant teams pull from this corpus, and it's the easiest one to reproduce.
Why scale is the whole point
None of this requires clever NLP. It's a phrase count. What makes it interesting is coverage: the archive spans 12,692 companies, so the query catches the mid-cap chip vendor in 2023, the education publisher in 2026, and the regional bank treating the theme as a funding risk — companies that would never make anyone's hand-picked watchlist for an AI narrative study.
Reading earnings calls tells you what one management team wants you to believe. Querying all of them tells you what the market is telling itself. Those are different datasets, and the second one only exists at scale.
The archive behind these numbers is queryable at earningscalls.dev — 247,000+ transcripts, full-text search, one GET request per quarter of narrative history.