Field experiments show AI tools deliver sizable task-level productivity gains, but national productivity improvements in the United States will depend on wider adoption, complementary software and data investments, and workforce reskilling.

Main Finding

AI adoption in the United States has a positive but uneven effect on labour productivity. Strong, replicable task‑level gains (faster task completion and higher quality in language, coding, and standardized customer service tasks) are documented in experiments, but economy‑wide productivity effects remain limited so far because diffusion is incomplete and gains depend on complementary investments, worker skills, firm readiness, and governance.

Key Points

• Task‑level evidence: controlled experiments report large productivity gains
  • Professional writing: ~40% faster completion and ~18% quality improvement (Noy & Zhang, 2023).
  • Software development: ~55.8% faster completion with AI pair programming (Peng et al., 2023).
  • Customer support: notable productivity improvements, especially for less experienced workers (Brynjolfsson, Li & Raymond, 2023).
• Adoption rates and diffusion
  • U.S. business AI use: ~17–20% (Dec 2025–May 2026); expected use in next six months ~20–23% (U.S. Census Bureau, 2026).
  • Adoption concentrated in larger, data‑intensive firms and sectors (software, finance, information, professional services), raising concerns about widening firm productivity gaps.
• Conceptual nuances
  • Productivity gains are task‑specific: AI helps well‑defined, reviewable tasks (language, coding, search, routine decisions) but underperforms on deep contextual judgment, sensitive human interaction, and high‑risk decisions.
  • AI behaves like a general‑purpose technology that requires complementary investments (skills, data, software, management) and time before broad macro gains appear.
• Quantitative potential (estimates and caveats)
  • McKinsey estimate: generative AI could raise labour productivity growth by ~0.1–0.6 percentage points annually through 2040 under varying adoption scenarios.
  • IMF/OECD: large shares of jobs are exposed to AI (~60% exposure in advanced economies), but exposure ≠ replacement; many roles can complement AI.
• Risks and failure modes
  • “Jagged frontier”: AI can harm performance when applied outside its strengths (Dell'Acqua et al., 2023).
  • Small firms face higher barriers (data, technical staff, cybersecurity, funds), slowing inclusive diffusion.
  • Productivity gains at task level may not translate to aggregate gains unless organizations redesign workflows and redeploy saved time into valuable output.

Data & Methods

• Approach: qualitative, evidence‑based literature synthesis rather than new econometric estimation; focused on sources from 2020 onward due to the technology’s recency.
• Main data and evidence sources:
  • Official statistics: U.S. Census Bureau (business AI use), U.S. Bureau of Labor Statistics (productivity series).
  • International reports: OECD, IMF.
  • Research syntheses and indices: Stanford AI Index, McKinsey Global Institute, NBER.
  • Academic experimental and working‑paper studies: Noy & Zhang (2023), Peng et al. (2023), Brynjolfsson et al. (2023), Dell'Acqua et al. (2023), and related literature.
• Dependent variable: labour productivity (output per worker or per hour).
• Independent variable (conceptual): AI adoption (use of generative AI, ML, predictive analytics, AI‑assisted software, automation).
• Moderators considered: worker skills, firm readiness (data quality, software investment, workflow redesign, management), governance (privacy, bias control, human review).
• Limitations highlighted:
  • Short time series and evolving adoption make robust macro causal inference premature.
  • Many firms remain in pilot/early stages; national statistics may undercapture informal or embedded AI use.
  • Task‑level experimental gains may not scale without organizational change.

Implications for AI Economics

• Measurement and research priorities
  • Need for firm‑level panel data on AI adoption, expenditures, task allocation, and complementary investments to quantify micro→macro transmission mechanisms.
  • Improve statistical capture of AI as part of software capital and services in national accounts.
  • Distinguish “exposure” from effective adoption and complementarity in empirical work.
• Policy and managerial implications
  • Invest in reskilling and AI literacy to realize complementarities and reduce uneven gains across the workforce.
  • Support SMEs with access to technical assistance, affordable data infrastructure, and cybersecurity to prevent widening firm‑level productivity gaps.
  • Encourage governance frameworks (quality assurance, human‑in‑the‑loop review, bias mitigation) to avoid costly errors that erode productivity gains.
  • Promote organizational change: workflow redesign, performance metrics, and managerial capacity are essential for converting task gains into aggregate output.
• Macro‑economic outlook
  • AI has credible potential to contribute meaningfully to future labour productivity growth, but magnitude and distribution depend on diffusion speed, complementary capital/skills, and institutional responses.
  • Policymakers should prepare for transitional distributional effects (firm heterogeneity, possible wage polarization) while enabling broad, responsible adoption.
• Researchable questions for AI economics
  • What fraction of task‑level time saved by AI is redeployed into higher‑value activities versus leisure or retained as cost savings?
  • How do firm size, industry, and data governance mediate the productivity returns to AI?
  • What policy mixes (training subsidies, adoption grants, data infrastructure) most cost‑effectively accelerate inclusive productivity growth from AI?

Summary: Martinez (2026) synthesizes recent experimental and official evidence to conclude that AI adoption is a likely net positive for U.S. labour productivity but that real economy‑wide gains will require deeper diffusion, complementary investments, managerial redesign, and governance to scale task‑level improvements into sustained aggregate productivity growth.