Firm Productivity

Bottom Line

Adopting AI and digital tools generally raises firm productivity, innovation, and resilience. Gains are uneven, depend on complements (infrastructure, skills, data, governance), and often follow a productivity J-curve (near-term profit pressure and integration costs) Bilgin and Ottaviano; Ropele and Tagliabracci; Highfill and Samuels; Kim et al. (2026); Kim and Baek; Oliaro et al. (2026)

What This Means in Practice

• Plan complements from day one: redesign workflows, build data pipelines, train workers, set governance, and budget for a J-curve (near-term dip before gains) Espinal Maya (2026); Kwon et al.; Pitic et al..
• Fix infrastructure and integration before scaling: connectivity, cloud, model serving (how models run in production), and ERP/data integration often cap speed and gains, especially for SMEs (small and medium-sized enterprises) Bilgin and Ottaviano; Srikanta Prasad S and Arora (2026); OECD (2026).
• Start with human-in-the-loop augmentation (people review/approve AI output) before automating judgment-heavy steps; use governance to manage fatigue and quality risks Pelechanoa et al. (2026); Lee et al. (2026).
• Budget hidden costs: evaluation, monitoring, error correction, and security can offset task-level gains. Build them into ROI models and vendor selection Oliaro et al. (2026); van der Maden et al. (2026).
• Use policy to spread gains: invest in digital infrastructure, simplify pro-innovation tax credits and pilots, and fund SME capability building to close adoption gaps Bilgin and Ottaviano; Chen et al. (2026); Rao and Chen (2026); OECD (2026).

What the Research Finds

Causal effects: AI and digital adoption raise firm productivity, innovation, and resilience

• Connectivity upgrades raised AI adoption, labor productivity, and exports, with larger effects for SMEs and software-intensive firms (natural experiment using staggered rollout) Bilgin and Ottaviano.
• In Italy, firm AI adoption increased labor productivity and profitability and shifted jobs toward higher-skill roles without reducing total employment (natural experiment) Ropele and Tagliabracci.
• Training startups to embed AI in production increased completed tasks, customer acquisition, and revenues without proportional input growth (randomized controlled trial, RCT) Kim et al. (2026).
• In US industries, AI adoption raised productivity and reduced inputs, with some shifts toward younger, less-educated workers (natural experiment) Highfill and Samuels.
• AI adopters produced more and higher-quality patents and raised productivity, lifting total factor productivity (TFP) in post-adoption years (natural experiment) Acharya et al..

Heterogeneity and complements: who gains, by how much, and why

• Bigger gains were associated with standardized processes, capable management, and relevant skills; diffusion alone delivered modest averages in Central and Eastern Europe (observational) Pitic et al..
• Faster decision cycles and flexibility explained much of the AI–performance association; the association varied in more turbulent, complex markets (observational) Kwon et al..
• Stronger management practices were associated with higher returns to technology investment in manufacturing (observational) Espinal Maya (2026).
• Cross-functional AI integration was associated with better firm performance; narrow, task-only use was not and sometimes coincided with lower employment in specific functions (observational) Bonney et al..

Adjustment costs, the productivity J-curve, and execution risks

• Short-run profit dip: in KOSDAQ-listed firms, AI adoption cut operating margins, left return on assets flat, and raised market value mainly in ICT (information and communications technology) firms, consistent with transition costs (natural experiment) Kim and Baek.
• Lab-to-production gap: agents that generate system code often failed to beat production throughput baselines (descriptive) Oliaro et al. (2026).
• Workflow design: copy-paste use lowered workers' self-efficacy and ownership; draft-first collaboration preserved them (RCT) Lee et al. (2026).
• Serving architecture: re-architecting model serving (running models in production) cut latency by over 50% and costs by 30–40% (descriptive) Srikanta Prasad S and Arora (2026).

Distributional shifts tied to productivity improvements

• AI adoption raised the firm-level skill premium in Chinese listed firms via substitution away from low-skilled labor and via productivity and capital-deepening channels (natural experiment) Liang et al. (2026).
• In Italy, adopters reallocated employment toward white-collar roles as productivity and profitability increased (natural experiment) Ropele and Tagliabracci.
• AI exposure in job postings was associated with lower demand for routine cognitive skills and higher demand for nonroutine analytical skills, especially in small, low entry-threshold firms (observational) Zhang and Zhang.

Policy and governance levers that raise firm-level productivity

• R&D tax incentives for intelligent manufacturing were associated with higher TFP and better sustainable development performance by easing financing constraints, especially in large, capital-intensive, non-state firms (natural experiment) Chen et al. (2026).
• National digital economy pilot zones increased firms' "new quality productive forces" by improving factor allocation, deepening digital adoption, and boosting green innovation, especially in non-state, high-tech, eastern firms (natural experiment) Rao and Chen (2026).
• Privacy regulation coincided with stronger positive associations between big-data adoption and firm value in China, suggesting rules can improve data quality and trust (natural experiment) Yan and Cai (2026).
• Infrastructure-first strategies for SMEs are associated with higher AI adoption and productivity; G7 SME gaps persist without targeted support (descriptive) OECD (2026).

What We Still Don't Know

• Long-run firm-level ROI that fully includes governance, monitoring, cybersecurity, and integration costs is under-measured; most evidence is short- to medium-term or task-level Oliaro et al. (2026); van der Maden et al. (2026).
• How agentic AI (systems that can act autonomously) changes organizational boundaries, spans of control, and profit dynamics at scale remains largely theoretical; few studies track multi-year outcomes after deep workflow redesign Pelechanoa et al. (2026); Srikanta Prasad S and Arora (2026).
• External validity beyond China, Europe, and a few OECD economies is thin; Sub-Saharan African and LATAM firm-level causal evidence is limited, with newer studies still mainly correlational Adediran (2026); OECD (2026).
• The timing and size of the productivity J-curve by industry, and which managerial or policy practices compress it, are not well established on comparable datasets Kim and Baek; Highfill and Samuels.
• Standardized measurement to separate human, AI, and interaction contributions to firm productivity is emerging but not yet mature, complicating benchmarking and accountability Nashed (2026).