Where AI Value Lands

AI–Productivity Link at the Firm Level

6 min read

AI–Productivity Link at the Firm Level

AI adopters are more productive than non-adopters. But most of the premium is not caused by AI — it reflects pre-existing competitiveness, digital readiness, and workforce capabilities. Once these are controlled for, the AI-specific productivity premium shrinks substantially.

This is the empirical and methodological caveat at the heart of the AI-productivity literature, as synthesized by oecd-sme-ai-adoption-2025 (Calvino & Fontanelli 2023; Calvino, Costa & Haerle, forthcoming; Acemoglu et al. 2022; McElheran et al. 2024).

The headline empirical findings

  • Productivity-quantile gap. Share of AI adopters in the top decile of the productivity distribution vs. the bottom decile: +40% in France (2018), +120% in Germany (2018), +240% in Italy (2020). Adopters are concentrated at the top.
  • Within-cohort premium. Comparing firms of the same size, age, and sector, AI users have labour-productivity premia of >4% in most G7 countries, sometimes >15%. Documented in Canada, France, Germany, Italy, UK, US.
  • Macro projection. oecd-g7-productivity-gain-projection — annual labour-productivity gains from AI of 0.2 to 1.3 percentage points across G7 economies over the next decade, depending on the adoption scenario. For reference: the US ICT boom of the mid-1990s delivered 0.5–1.5 pp.

The methodological caveat

The headline gap is not all causal. Three confounders consistently shrink the estimated AI premium when added:

  1. Pre-existing competitiveness. Firms that are already more digital and more competitive are also more likely to adopt AI. The productivity gap partly reflects selection, not treatment.
  2. Digital capabilities. Once cloud computing, other digital technologies, and ICT-specialist presence are controlled for, the AI premium shrinks substantially (Acemoglu et al. 2022; Calvino & Fontanelli 2023).
  3. Worker skills. Firms with more ICT-trained workers benefit more from AI — the premium is mediated by workforce capability, not by AI per se.

The honest interpretation: AI is a multiplier of pre-existing digital and human capital, not an independent productivity-boost driver. This is one of the most important findings in the source for our paper.

The J-shaped productivity dynamic

McElheran et al. (2024) document a J-shaped firm-level productivity trajectory after AI adoption in US manufacturing: productivity often declines before rising, consistent with Brynjolfsson, Rock & Syverson’s (2021) “complementary-investments-take-time” model of general-purpose technologies.

This matters because:

  • Macro projections that compound today’s adopter-premium forward will overstate the gain. Some current adopters are still in the J’s downward leg; some non-adopters will accelerate once they cross the J’s pivot.
  • Returns to AI are gated by organizational redesign. Agrawal, Gans & Goldfarb (2023) argue the AI surplus is unlocked by reorganization, not by the technology in isolation. Firms that adopt AI without restructuring workflows capture little.

Why this is load-bearing for the U-curve

  • It re-routes surplus. If AI’s productivity gains are mediated by complementary digital capital (cloud, data, skills), then the surplus does not stop at AI vendors — it accrues to whoever built and owns the complementary stack. For most SMEs, that complementary stack is rented from hyperscalers and SaaS vendors. The application-layer firm captures the operational productivity gain; the substrate vendor captures the recurring rent. Cross-check H2_u-curve-of-value — the bottom of the U is reinforced by exactly this mediation.
  • It explains the scaling-gap. If AI productivity gains depend on complementary investments and organizational redesign, the 74% of firms struggling to scale AI are not necessarily failing at AI — they are failing at the complements. The lever is enablers, not models.
  • It complicates the SME productivity story. SMEs have less of the complementary stack (cloud, ICT specialists, training budgets). Their AI premium is therefore likely to be smaller than the headline cross-firm regression suggests — both because they have less complementary capital and because they tend to use AI for peripheral tasks (only 29% in core activities, per oecd-sme-ai-adoption-2025).

Generative AI as a special case

The OECD source argues generative AI exhibits the canonical general-purpose-technology markers:

  • Pervasiveness — usable across sectors and functions.
  • Continuous improvement — model capability rising on a steep slope.
  • Innovation-spawning — accelerating R&D, lowering entrepreneurship entry barriers.

In SME surveys (OECD 2025), generative AI shows up as the most-used AI type, and 39% of SMEs that adopted generative AI and had a recent skill gap reported it helped compensate. This is a different kind of productivity gain: not “AI does the task faster” but “AI substitutes for missing skills” — a substitutability story that could be either bullish (broadening the productive workforce) or bearish (compressing skilled-labour wage premia). See capital-labor-divergence.

Tensions

  • Causal identification is hard. All the cited studies are observational. Selection bias is partially controlled but not fully eliminated.
  • The premium varies by AI type. Firms developing AI internally (Calvino & Fontanelli 2024 for France) capture larger premia than firms buying externally — a finding that pushes value upstream toward Makers / Shapers and away from Takers, conditional on firms having the capital to develop in-house. Most SMEs cannot.
  • Macro and micro estimates disagree in scale. The +15% within-cohort labour-productivity premium does not aggregate cleanly to the 0.2–1.3 pp annual macro projection — too many firms have not adopted, and the J-curve drag is unmodelled.
  • task-based-framework is the bear cap. Acemoglu’s task-based estimate suggests cumulative US TFP gains from AI of <0.71% over a decade — incompatible with the upper end of the OECD projection. The two views are unreconciled.

Graph View

Backlinks