Canaries in the Gold Mine
A · Strong evidence NBER working paper by Tania Babina (Maryland, NBER) giving a 15-year firm-level answer to the AI productivity paradox with an instrumental-variables check and discussant comments; strong evidence, not yet a reference point for the field.
Summary based on the full text·Original: about 150 minutes to read
In short
NBER Working Paper 35684 (August 2026) by Tania Babina, Alex X. He and Renhao Jiang of the University of Maryland. Measuring AI investment through firms' employment of AI-skilled workers, they find that AI investment was unrelated to productivity growth over 2010-2018 but strongly positive over 2018-2024. They trace the gain to organisation capital: durable, firm-specific knowledge that AI helps codify from the firm's own operations. The slow build-up of that knowledge explains why the productivity effect took about a decade to appear.
Key numbers
- 7.5 ppextra revenue TFP growth over 2018-2024 per one-SD increase in AI worker share growth
- 2.1%share of US public firms' US workforce that is AI-skilled in 2024 (over 300,000 workers)
- 58% vs 31%AI jobs vs non-AI jobs that build organisation capital
- 1,321firms in the final regression sample
As reported by the source.
English·Web page·opens nber.org
What it reports
- By 2024 US public firms employ more than 300,000 AI-skilled workers, 2.1% of their US workforce, with AI shares rising in every industry. Firms outside the public-firm sample invest at roughly half the rate.
- Over 2018-2024, a one-standard-deviation increase in AI worker share growth goes with 6.2 percentage points more sales-per-worker growth and 7.5 points more revenue TFP growth over six years, about one point a year.
- Over 2010-2018 there is no such relationship.
- Recent gains come from more sales with flat headcount: a one-SD increase in AI investment predicts 6.7 points greater sales growth and a near-zero employment effect.
- AI-investing firms expand intangibles (SG&A, R&D) rather than plant and equipment, and shift toward advanced-degree and senior workers while non-college and junior shares fall.
- 58% of AI jobs build organisation capital vs 31% of non-AI jobs; the AI share rose from about one-half in 2010 to almost two-thirds in 2024. 66% of generative and agentic AI jobs build it vs 57% of earlier ML, NLP and computer-vision jobs.
- AI jobs build organisation capital more often in non-tech firms than in tech firms.
- Productivity growth comes from organisation-capital-building AI jobs; other AI jobs show no effect. Gains are larger for firms starting with little organisation capital.
How they got it
AI skills were identified from about 367 million Lightcast US job postings (2010-2025) and used to classify AI jobs in Revelio Labs resumes covering about two-thirds of the US workforce, with a false-positive rate of about 2%. These were matched to Compustat firms and analysed with long-difference regressions. An instrument based on pre-2010 hiring ties to universities strong in AI research supports a causal reading. Job titles and descriptions were scored to measure organisation-capital-building work.
Limits
- Only US public firms; firms relying solely on off-the-shelf AI tools are not captured.
- Revenue TFP can mix price and quantity effects, though the authors find no price rises in AI-intensive industries.
- Resume data over-represent white-collar workers and are updated with a lag.
Why it matters for companies
The paper ties AI productivity gains to building internal capability, not to buying tools, and to a lag of years. Its case examples show the pattern: models trained on records a firm already produces, such as Morgan Stanley's advisor meeting notes, P&G's plant telemetry or Expedia's incident histories. For a traditional company the practical reading is that the value sits in turning operating knowledge into reusable systems, and that firms with little codified process knowledge have the most to gain.
Summary based on the full text. Read the original for the full argument and data.Read the original on nber.org
- Tags
- Productivity