I counted AI vocabulary on 5,305 earnings calls held by current S&P 500 companies since January 2024, using speaker-by-speaker transcripts. The short version: executives have mostly stopped saying "generative AI" and started saying "agentic".
| Calls held | Said "agentic" | Said "generative AI" | Mentioned AI at all |
|---|---|---|---|
| Jan to Mar 2024 | 0.2% | 16.8% | 43.1% |
| Jan to Mar 2025 | 6.3% | 16.0% | 52.2% |
| Oct to Dec 2025 | 14.5% | 12.8% | 65.5% |
| Jul to Sep 2026 | 20.3% | 5.6% | 69.5% |
| Oct 1 to 23, 2026 | 28.2% | 3.8% | 71.8% |
Each full quarter has 461 to 487 calls; October is only the first 78 calls of the season. These are words, not deployments: the count shows what companies choose to talk about on their calls, which tracks attention, not how many of them have agents in production. The full quarterly series, with charts, is in the study on Equibles.
Five things I didn't expect
1. It's a management word, not an analyst word. On calls held from July to September 2026, company speakers said "agentic" 371 times, analysts 46 times, and 52 mentions had no identified speaker. Even if all 52 were analysts, at least 79% came from the company side. On 83 of the 97 calls that used it, an executive said it first; an analyst did on 7 and an unidentified speaker on 7. That matters because an analyst asking about a word is a different signal: it means investors expect an answer.
2. DeepSeek was the opposite. It came up on 53 calls (11.2%) in January to March 2025, the highest share any AI developer reached until OpenAI hit 11.6% a year later. An analyst raised it first on 31 of those 53 calls. By July to September 2026 it was on none.
3. Anthropic has nearly caught OpenAI. In early 2024, OpenAI or ChatGPT was named on 17 calls and Anthropic or Claude on 5. Last quarter it was 40 to 31 (8.4% vs 6.5% of calls), and on the October calls so far it's 5 each.
4. It spread well beyond software. Technology companies used "agentic" on 53% of their calls last quarter. Consumer staples went from 1 call a year earlier to 8 of 31 (Walmart, Target, Costco, Dollar General and others), and real estate from 1 to 6 of 27, mostly data center owners. Energy, utilities and materials: zero, even though most energy calls mention AI.
5. "Agentic commerce" is a payments story. Visa, Mastercard, American Express, Global Payments and Coinbase use the phrase. When an analyst asked Costco about it, the CFO said he wasn't sure he'd call it that.
The phrase it replaced is fading fast: "generative AI" was said 459 times in January to March 2024 and 40 times in July to September 2026. Of the 79 companies that said it on a call in mid-2024, 76 held a call in the same months of 2026, and 40 of those said "agentic" while 17 still said "generative AI".
How I counted, and the traps
- Case-insensitive whole words, one call per company per fiscal quarter, longest transcript kept, calls under 1,500 words dropped.
- Speech recognition spells the new word several ways. "Agentic" includes "agentec", "agentech", "a gentic", and "a genetic" only when the next word is AI, commerce, workflows or similar. Without that condition you also count "a genetic disease".
- "Claude" counts only with an AI term within about 90 characters. FedEx's interim CFO is named Claude, and so was a Norfolk Southern board member in 2024.
- "Gemini" doesn't count on Alphabet's own calls, "Copilot" not on Microsoft's, "Llama" not on Meta's.
- Speaker roles come from the transcripts; 52 of the 469 "agentic" mentions last quarter had no identified speaker.
- Two transcription sources cover different years. In the quarters where both were common, they gave nearly the same shares (11.7% vs 12.0% for "agentic" in July to September 2025), so I don't think the trend is an artifact.
Run it on one company
The single-company version is a phrase count across that company's calls. I walked through it for Microsoft and "agents" in an earlier post. With the Equibles MCP server connected, a prompt like "Using Equibles, how often has Visa said 'agentic' on its last eight earnings calls?" does it in one tool call.
The full write-up has the sector breakdown, the 13 companies that say it most (Adobe leads with 27 mentions on one call), quotes linked to each call, and the methodology: Agentic AI Has Overtaken Generative AI on Earnings Calls: Evidence From 5,305 S&P 500 Calls.
Top comments (8)
That bound is the right way to close it, and it changes the job.
If the claim survives its own worst case, the unknowns stop being a reason to wait. State the bucket, state the bound, then make the claim that is left: 371 of 469 with all 52 unidentified mentions given to the other side, and 83 to 14 on who spoke first. A reader can check that arithmetic. They cannot check a number you quietly resolved by assumption.
Keep the wording one-sided, though. That bound is a floor, not an estimate. "At least 79% of mentions came from company speakers" is what the data defends; "about 79%" is a midpoint the unknown bucket never measured, and it reads as precision you did not buy.
Which leaves the title doing more work than the count can carry. "Agentic is now the more common word in earnings calls" survives everything above. "Agentic AI has overtaken generative AI" only survives if the piece says plainly that the measure is vocabulary and the signal is attention, not deployment. That is still the finding, and the bound you just ran is what lets you say it without hedging.
Fair on the wording, and my last reply has exactly that slip: "about 79%" inside a worst case. The bound is a floor, so it should read at least 79% of mentions from company speakers (371 of 469), analysts at most 21%, and 83 to 14 on who spoke first.
On the title, agreed that the measure is vocabulary. The post opened with "I counted AI vocabulary" but never said what that leaves out, so I've added a line after the table: these are words, not deployments, and the count tracks what companies choose to talk about on their calls, not how many have agents in production.
You are right on depth, and the correction makes my point sharper rather than weaker: 371 was the speaker share and I read it as the whole count. That is the same mistake I came to check for, a true number with the wrong population attached, so thank you.
With the full split the executive-first pattern loses some of its force. If 46 of the 469 mentions come from analysts, the vocabulary is not purely a top-down deck word; some of it is analysts asking about it, which is a different signal from management positioning. I would still keep the two separate in the write-up: mentions by speakers, mentions by analysts, and the 52 unknown. The unknown bucket is the one worth stating out loud, because it caps how strong any conclusion can be.
Also worth saying: you corrected the number down to a bigger gap, which is the honest direction. Most corrections I get go the other way.
Agreed on keeping the three apart, and the unknown bucket gives a usable bound. Even if all 52 unidentified mentions came from analysts, company speakers would still account for 371 of 469, about 79%, and analysts at most 98, about 21%.
The who-said-it-first count holds up the same way: a company speaker raised the word first on 83 of the 97 calls, an analyst on 7, and an unidentified speaker on the other 7, so the worst case is 83 against 14.
Your reading of the analyst share still stands. Analysts using the word means it has become a question investors expect answered, which is a different signal from a word management puts in the deck, and it's a fair ask to show that split, unknowns included, in the write-up itself.
Ran the checks. This one holds, and I want to say that before the two notes.
What the number counts: a call counts as "said agentic" at one mention, so 28.2% is breadth across companies, not depth inside them. You report depth too, and it survives: 371 mentions of "agentic" against 40 of "generative AI" in Jul-Sep. The overtake is real on both measures.
Where the title still outruns the count: "agentic AI has overtaken generative AI" names two technologies. You counted two vocabulary items in prepared remarks and Q&A. Your point 1 is the most interesting datum in the post and it complicates the verb. On 83 of 97 calls an executive said it first, before the analyst asked. A word that arrives from the top, in a deck, is a management word, so the curve tells you about budget before it tells you about shipped agents. Still a good measurement. It measures something narrower than the sentence does.
Credit on the last row: the overtake already shows in the completed Jul-Sep quarter (20.3% vs 5.6%), so the headline is not riding the 78-call October slice. That would have been the easy way to make this number.
Thanks for running the checks. Fair point on the verb: it's two ways of describing AI, not two technologies, and a word count measures attention rather than spending. The executive-first pattern is exactly why I read it as a budget and positioning signal, not as evidence of agents in production.
One small correction on depth: 371 is the company speakers' share. The full count last quarter was 469 mentions of "agentic" against 40 of "generative AI", with 46 from analysts and 52 from speakers the transcripts couldn't identify. Same conclusion, slightly bigger gap.
Then this is closed on my side, and one loose end stays in yours.
You added the sentence the title needed. The split is still the part only you can add: company speakers, analysts, unidentified, in the post itself. If the analyst share stays out, the vocabulary wording is what the piece can carry. If it goes in, so does the reading that analysts asking about a word is a different signal from management putting it in a deck.
Nothing else here needs checking. Good piece, and your correction ran the hard direction twice.
Done, it's in point 1 now: 371 company, 46 analysts, 52 unidentified, so at least 79% from the company side, and who spoke first (83, 7 and 7), with a line on why the analyst share is its own signal. Thanks for taking it the whole way.