Firms that install AI products patent more and patent better — their inventions are more cited, more original and span new technological domains — and these innovation gains translate into higher R&D, productivity and market value, implying roughly a 1.5% boost to aggregate TFP.

Main Finding

Firms that install AI products subsequently increase both the quantity and quality of innovation. Using a novel firm-level AI adoption timing measure and a stacked difference‑in‑differences design, the paper shows adopters raise patenting, produce patents that are more cited, more original/generic, and span farther technological classes, and ultimately exhibit higher productivity and market value. The authors interpret AI as an innovation‑enabling intangible investment that raises aggregate value‑added TFP by about 1.51% in a representative post‑adoption year.

Key Points

• AI adoption — measured from the timing of firm AI product installations — leads to higher patent counts for adopters relative to matched nonadopters.
• Post‑adoption patents look qualitatively different:
  • More citations and more patent claims (higher impact/complexity).
  • Greater originality and generality.
  • A higher share of exploitative patents (building on firm’s prior tech) while also exploring new technological domains and more distant technology classes.
• Effects indicate both refinement (exploitation) and exploration of knowledge enabled by AI.
• Heterogeneity: effects are stronger for firms with a more focused business scope; they are not concentrated among large firms, financially unconstrained firms, or firms already in high‑tech industries.
• Mechanisms: evidence points to AI improving R&D efficiency, shifting inventor composition, and enhancing knowledge recombination.
• Complementary firm outcomes: adopters increase R&D spending (but not capital expenditures), show no significant changes in operating costs, and display higher measured productivity and market valuation.
• Aggregate implication: estimated AI adoption raises aggregate value‑added TFP by ~1.51% in a representative post‑adoption year.

Data & Methods

• AI adoption measure: novel firm‑level indicator constructed from the timing of AI product installations at firms (installation dates used to define treatment timing).
• Identification: stacked difference‑in‑differences design exploiting staggered adoption timing (event‑study / dynamic treatment effects framework) to compare adopters to nonadopters while accounting for differential timing.
• Outcomes analyzed:
  • Extensive patent outcomes: patent counts, citations, number of claims.
  • Patent composition measures: exploitative vs exploratory share, originality, generality, technological class distance.
  • Firm financial and input responses: R&D spending, capital expenditures, operating costs.
  • Productivity and market value measures; authors compute aggregate value‑added TFP effects from firm‑level estimates.
• Heterogeneity and mechanism tests: variation by firm focus, size, financial constraints, industry tech intensity; analyses of inventor composition and recombination patterns to probe channels.
• Robustness: dynamic/event‑study checks to address pre‑trends and timing heterogeneity (stacked DID used to mitigate biases from staggered adoption).

Implications for AI Economics

• AI as an intangible, innovation‑enabling investment: Evidence supports modeling AI not just as a productivity shortcut but as a capital‑like intangible that changes firms’ innovation processes and expands their technological reach.
• Dual role in exploration and exploitation: AI appears to simultaneously increase refinement of existing capabilities and enable entry into more distant technological areas, with implications for endogenous growth and technological diffusion models.
• Policy and firm strategy:
  • Policies that lower adoption frictions for AI (e.g., training, matching, standards) could have larger welfare effects through increased innovation, not only through operational efficiency.
  • Because effects are not restricted to large/financially unconstrained/high‑tech firms, broad adoption support may foster widespread innovation gains.
• Measurement and macro aggregation: firm‑level adoption timing and stacked DID approaches are useful for assessing dynamic adoption effects; the ~1.51% aggregate TFP lift highlights potentially sizable macroeconomic impacts from diffusion of AI as an intangible.
• Open research directions: long‑run persistence of innovation gains, optimal complementary investments (human capital, organizational change), cross‑firm spillovers and sectoral reallocation effects, and welfare implications of AI‑driven changes in the technological frontier.