Sunil Venkataram
Back to garden
•
🌰 seedling
AI Productivity - The Micro-Macro Disconnect
type/noteaieconomicslaborproductivity

AI Productivity — The Micro-Macro Disconnect

The pattern

Robert Solow observed in 1987 that “you can see the computer age everywhere but in the productivity statistics.” Nearly four decades later, the same dynamic plays out with AI. By early 2026, over a dozen randomized controlled trials document productivity gains from generative AI:

  • Customer support agents resolve 14-15% more issues per hour with AI assistance (Brynjolfsson, Li, Raymond 2025)
  • Software developers complete pull requests 26% faster across three Microsoft/Accenture RCTs (Cui et al. 2025)
  • Professional writers finish tasks 0.8 standard deviations faster at 0.4 SD higher quality (Noy & Zhang 2023, Science)
  • BCG consultants produce 12% more tasks, 25% faster, at 40% higher quality — but only within AI’s capability frontier (Dell’Acqua et al. 2023)

Meanwhile, the Penn Wharton Budget Model estimates AI contributed approximately 0.01 percentage points to total factor productivity growth in 2025 — essentially zero.

Four explanations for the gap

1. Endogenous adoption. In micro studies, participants are told to use AI and often receive training. In the real world, workers self-select into adoption. A 2025 BCG survey found only 36% of workers feel properly trained to use AI. Untrained users apply AI to unproductive tasks, choose suboptimal models, or fail to integrate outputs into workflows.

2. Who adopts differs from who benefits. Micro studies repeatedly show an equalizing effect — less experienced workers gain the most. But the Anthropic Economic Index and BCG surveys show real-world adoption concentrates among middle-to-upper wage white-collar workers. Managers use AI at nearly twice the rate of frontline workers. The gap between experimental beneficiaries and actual adopters means the largest potential gains go unrealized.

3. Bottleneck tasks. Jobs consist of many tasks. Even if AI accelerates some tasks dramatically, the ones it does not touch become binding constraints. A developer who writes code twice as fast still waits for code reviews, attends meetings, coordinates across teams, and navigates organizational approvals. Until organizations restructure workflows around AI-augmented speed, task-level gains get absorbed by job-level bottlenecks.

4. The productivity J-curve. When firms adopt transformative general-purpose technologies, measured productivity often falls initially because resources flow into investment, reorganization, and learning that standard statistics do not capture as output. See Productivity J-Curve - Why Transformative Technologies Suppress Measured Output Before Harvest for the detailed mechanism.

The update: signs of convergence

By March 2026, the gap may be closing. Jason Furman, previously skeptical, noted that revised BLS data shows nonfarm labor productivity 2.2% above the CBO’s pre-pandemic forecast, with annual growth at 2.8% — nearly double the prior decade’s average. Erik Brynjolfsson argues the harvest phase has begun. European data from Aldasoro et al. (2026) shows a 4% average productivity increase in AI-adopting EU firms.

The strongest counter-data point: Humlum and Vestergaard (2026) find precise null effects on earnings, hours, and wages in Denmark two years after ChatGPT, despite 93% adoption in surveyed firms. What moves is the structure of work — 8% of users report taking on entirely new tasks, and occupation-switchers see earnings grow 12 percentage points faster.

Why this matters for investment and strategy

The micro-macro disconnect determines the timeline for AI returns. If the gap closes quickly (Brynjolfsson’s view), AI-exposed sectors should see measurable output gains within 1-2 years and the equity premium for AI-forward companies is already justified. If the gap persists (the Denmark data), the J-curve is longer than expected and the investment thesis depends on patience and organizational transformation, not just technology deployment.

Backlinks (4)
Connected Notes