The macro-economic evidence on AI presents a paradox that is both statistically robust and analytically contested. At the aggregate level, Goldman Sachs finds no meaningful relationship between AI adoption and economy-wide productivity. PwC reports 56% of CEOs getting nothing measurable from their AI investments. Yet at the firm level, European data shows a 4% average labor productivity gain from AI adoption, and the Stanford Enterprise AI Playbook documents a 71% productivity improvement in the best-designed escalation workflows. The reconciliation is not reassuring: the gains are highly concentrated, captured by roughly 20% of deploying firms, and they have not diffused to the aggregate in any measurable way five years into the investment cycle.
The Jevons paradox finding complicates the story further. Rather than efficiency gains reducing AI consumption, improved models triggered a 44% surge in agent usage — driven disproportionately by smaller, more agile firms. This is consistent with a general-purpose technology reading, where lower effective cost per task expands the frontier of tasks worth attempting. The GPT framework suggests the aggregate productivity lag is historically normal: electricity took 40 years, IT took 25. The J-curve hypothesis is plausible. It is also, as the evidence base acknowledges, conveniently unfalsifiable in the near term.
The labor displacement data has shifted materially. Goldman now estimates 16,000 net US jobs per month eliminated, with the burden falling disproportionately on young, female, and entry-level workers. The Brookings “vanishing ladder” framing — where AI erodes the gateway occupations through which lower-wage workers build expertise and advance — captures something the aggregate employment statistics conceal. The employment headcount may hold while the mobility architecture underneath it collapses.
Europe’s structural failure to convert AI investment into captured value compounds this: brilliant science, deep research capability, and genuine industrial AI strengths in robotics and energy, but a venture funding gap that widens brutally at growth stage. We do not yet know whether these distributional and structural asymmetries will self-correct or compound.