Research opportunity · Aggregate null or distributional shift? Employment
How much of a firm's AI productivity gain reaches the wages of workers who stay?
Earlier version (v1)
This is RO-LD02-05 as it was published at v1, kept unchanged. The current version is v3; its changes are listed there.
Question
When a firm starts using generative AI, how much does its value added per worker change over the following two to three years, and what share of that change shows up in the wages of incumbent workers who stay, compared with profits and with the wages of new hires?
Causal mechanism not isolatedOne PhD chapter or paperLow to medium costPartly addressed
Why it matters
Task experiments show large productivity gains, and incumbent earnings look flat, but no study found connects the two inside the same firms. Whether workers share in the gains, or whether gains go to profits and lower hiring, is the distributional core of this debate and a central question for wage bargaining. A pass-through estimate, even an imprecise one, would replace speculation with a number that can be compared with known rent-sharing rates.
The debate lists this among what we still do not know: Will productivity gains pass through into wages, profits, prices or output?
Evidence that motivates it
- The current R05 evidence base does not establish a causal pass-through from AI-induced productivity gains into worker wages.
The current evidence does not establish causal pass-through from AI productivity gains to wages; it also does not license saying workers receive none of the gains, so a direct estimate is needed either way. - Instrumented AI adoption among more than 12,000 firms raises labour productivity by about 4% with no adverse short-run firm-level employment effect.
Instrumented AI adoption in over 12,000 firms raises labour productivity by about 4% without short-run employment loss, but the claim is not genAI-specific and does not license wage pass-through, which this design targets. - In 11 AI-exposed Danish occupations, Humlum and Vestergaard estimate precise null effects on earnings and recorded hours two years after ChatGPT, ruling out effects larger than 2%.
Incumbent workers in 11 exposed Danish occupations show precise null earnings and hours effects two years after ChatGPT; the claim does not license a no-hiring-effect reading, which is why the design separates stayers from new hires.
A possible design
Firm-level rent-sharing design in linked registers. Take firms whose answers to the national ICT-usage-in-enterprises survey show they began using generative AI (or AI for text generation) in a known year, matched to non-adopters on sector, size and pre-period trends. Estimate event-study changes in value added per worker, operating profit per worker, the wage bill, and the individual wage growth of workers employed before adoption who stay, separately from the starting wages of new hires. The pass-through elasticity is the change in stayers' wages divided by the change in value added per worker. Compare it with the same register's pre-AI rent-sharing estimates and across firms with and without collective-agreement coverage.
Data needed
Linked employer–employee registers with firm accounts (Denmark, Norway, Sweden, Finland, the Netherlands or France); the ICT-usage-in-enterprises survey linked by firm identifier (the harmonised EU survey asked about AI that generates written or spoken language in its 2023 and 2024 waves, covers enterprises with at least 10 employees, and national versions and earlier waves must be checked); collective-agreement coverage where recorded.
Feasibility
One PhD chapter or paper. Low to medium; register access fees and approval time. Adoption is self-reported and not randomly assigned, so the design identifies pass-through only under the matching and parallel-trend assumptions, and should say so. Firm value-added gains of a few percent mean wage effects will be small; power calculations should come before the analysis.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026; not yet checked in depth).
Danish worker-level work estimates 3–7% pass-through of time savings to earnings, but mainly from self-reports, with the authors noting they lack power in the administrative data. A European firm study finds higher average wages at AI adopters, for AI and big-data analytics in general, not generative AI, and without a worker-level analysis. The authors of a US state-by-industry study calculate that labour captures about 29 cents of each dollar of AI-driven output growth, using exposure, not adoption. No study was found that estimates generative-AI pass-through from firm value added to stayers' wages in linked administrative data with observed adoption.
How strong this evidence is
- Who searched and checked: an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: 1 of the 3 nearest works below was judged from an abstract, summary or search listing only; it is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web search on the date shown; Google Scholar, SSRN and paywalled journal pages were not systematically searched, and one item was read through the authors' article rather than the full paper. Absence of a hit is not proof that the work does not exist.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Humlum & Vestergaard, "Large Language Models, Small Labor Market Effects," Becker Friedman Institute Working Paper No. 2025-56 (May 2025)
Two survey rounds of about 25,000 Danish workers each in 11 exposed occupations, linked to registers: no significant effect on earnings or recorded hours; self-reported time savings of about 2.8% of hours, with an estimated 3–7% passed to earnings based on self-reports; the authors say the administrative data lack the power to detect comparable pass-through. Worker-side, not firm value added; no profit split. - Aldasoro, Gambacorta, Pal, Revoltella, Weiss & Wolski, "AI adoption, productivity and employment: evidence from European firms," BIS Working Paper No. 1325 (23 January 2026)
Matched survey and accounts data (EIBIS-ORBIS) for over 12,000 EU and US firms, EU adoption instrumented with matched US peers' adoption: about 4% higher labour productivity, no short-run employment loss, higher average wages at adopters. The adoption question covers big-data analytics and AI in general, not generative AI; firm-level averages only, with no estimate of pass-through to incumbents' wages. - Makridis & Johnston, "More AI-Exposed Industries and States Are Benefiting, But Results Are Heterogeneous," ProMarket (21 May 2026), the authors' summary of Johnston & Makridis, "The Labor Market Effects of Generative AI: A Difference-in-Differences Analysis of AI Exposure," SSRN Working Paper No. 5375017 Judged from an abstract, summary or search listing only
The authors' own article about the paper, not the paper itself: US QCEW state-by-industry data 2017–2024 with occupation-based exposure; a one standard deviation higher exposure goes with about 10% higher output, 3.9% higher employment and 4.8% higher wage bill by 2024, from which they calculate that labour captures about 29 cents per dollar of AI-driven output growth; effects begin in 2021, before ChatGPT. Exposure-based, not adoption; no firm-level pass-through.
Search queries used
generative AI adoption firm productivity gains wages profits rent sharing pass-through firm-level evidence 2026instrumented AI adoption 12,000 firms labour productivity 4% employment no adverse effectAI adoption wages incumbent workers firm-level linked employer-employee wage effect generative AI value added split labour sharegenerative AI adoption firm value added wages rent sharing linked employer-employee 2026 working paper pass-through incumbents
Provenance
RO-LD02-05 v1. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-02 debate on 2026-10-12. Checked by an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.