I wanted to count how often executives outside tech say AI is changing their headcount, from the speaker-labelled transcripts of S&P 500 earnings calls (4,335 calls by 403 companies, January 2024 to October 2026, technology sector and a few AI-product companies excluded).
What I tried and what broke:
- Keyword co-occurrence alone over-counts. A sentence with an AI term (“AI”, “GenAI”, “agentic”, “LLM”, “machine learning”), a staff word and a change word in the same or a nearby sentence is a decent candidate filter, but many of the hits are executives saying the opposite: “this reduction was not driven by AI”, “we are not using AI to cut jobs”. Calls with a denial went from none in 2024 to 13 in 2026, against 17 calls with an actual claim, so this year denials are almost as common as claims.
- Analyst questions look like claims. “How should we think about AI and headcount?” matches every pattern, and analysts asked it on 18 calls in 2026, up from 1 in 2024. Speaker roles fix most of it; a few turns have no role or merge the question into the answer, and those needed reading in context.
- “Hiring less” is the main case, not layoffs. Of the 41 calls with a claim, 22 said AI lets the company grow without adding people and 5 that it replaces people who leave; 14 described cut roles. A classifier trained on layoff language misses most of them.
- A second, wider pass mattered. AI sentences mentioning staff or hiring without a change word found further calls the first filter missed.
In the end I read every candidate in its turn and labelled it by hand (claim, broader mention, denial, analyst question), which is fine at this size but won’t scale to every call in the archive.
Question for people who have done stance detection on financial text: has anyone had luck with an NLI-style model (premise: the turn, hypothesis: “the company is reducing headcount because of AI”) for separating claims from denials and questions, or is a small fine-tuned classifier on a few hundred hand labels the better route?
I work at Equibles, which published this. The study with the counts and the method: Executives Outside Tech: AI Means Hiring Less, Not Layoffs - Equibles
Data: earnings call transcripts, processed by Equibles.