Summary: "What is the impact of AI on productivity?" by Alex Imas
Source: aleximas.substack.com (living document, current through Jan 2026)
This is a comprehensive literature review of the research on AI's productivity impact, organized into micro (controlled studies) and macro (aggregate economic data) evidence.
The Core Finding: A Micro-Macro Disconnect
Micro studies show real productivity gains. Macro statistics don't show them yet. The author compares this to Solow's 1987 observation: "You can see the computer age everywhere but in the productivity statistics."
Micro Evidence (Task-Level Studies)
Results lean heavily positive, but with important nuances:
Strongest gains:
- Customer support (Brynjolfsson et al., QJE 2025): 14-15% more issues resolved/hour; 30-35% gains for less experienced agents
- Software development (Cui et al., 2025): 26% more pull requests across 5,000 devs at Microsoft/Accenture — but build success rate fell 5.5 points ("guess-and-check" behavior)
- GitHub Copilot (Peng et al., 2023): 55.8% faster task completion; (Yeverechyahu et al., 2024): 37-55% more commits
- Writing tasks (Noy & Zhang, Science 2023): 0.8 SD faster, 0.4 SD higher quality. Lower performers benefited most.
- Ad copy (Ju & Aral, 2025): Human-AI teams 73% more productive per worker (but human-human teams still better for images)
- Mammography (MASAI trial, 2026): 44% workload reduction, 29% more cancers detected, 12% fewer interval cancers — the first completed RCT on AI in medical screening
- E-commerce (Fang et al., 2025): AI chatbots increased sales 16%
- BCG consultants (Dell'Acqua et al., 2023): 12.2% more tasks, 25.1% faster, 40% higher quality — but 19 points worse on tasks outside AI's frontier (the "jagged frontier")
Negative/cautionary results:
- METR study (Becker et al., 2025): Experienced open-source devs were 19% slower with AI — yet believed it helped by 20%. Major perception-reality gap.
- Learning penalty (Shen & Tamkin, 2025): Engineers using AI scored 17 points lower on subsequent quizzes. AI accelerates tasks but undermines skill acquisition.
- Kenyan entrepreneurs (Otis et al., 2023): No average effect on revenue. High performers improved ~15%, low performers got 8-10% worse — widening the gap.
- Book publishing (Reimers & Waldfogel, 2026): Titles tripled but average quality declined. Top-tier quality improved; flood of low-quality new entrants.
Who benefits most? Micro studies generally show an equalizing effect — less skilled/experienced workers gain the most. But there are notable exceptions (Google found senior devs benefited more).
Macro Evidence (Aggregate Data)
Results are mostly null or ambiguous so far:
- Denmark (Humlum & Vestergaard, 2025): Essentially zero effects on earnings or hours, even for heavy AI users. Key insight: workers may be taking productivity gains as on-the-job leisure.
- International executive survey (Yotzov et al., 2026): 70% of firms use AI, but 80%+ report no impact on employment or productivity over 3 years. Execs predict only 1.4% productivity boost over next 3 years.
- Penn Wharton (2025): AI contribution to TFP growth ≈ 0.01 percentage points — essentially zero.
- Yale Budget Lab: "The picture… largely reflects stability, not major disruption."
- US survey (Bick et al., 2024): 40% of Americans use generative AI; self-reported savings average ~6 extra minutes/day.
Emerging signals:
- Brynjolfsson (FT, 2026): US productivity grew ~2.7% in 2025, nearly double the prior decade's average. Payroll revised down 403K while GDP stayed strong — the decoupling signature of productivity growth.
- "Canaries in the Coal Mine" (Brynjolfsson et al., 2025): Entry-level workers (22-25) in AI-exposed jobs show 15-16% employment declines. Senior workers stable. Driven by slower hiring, not firings.
- Résumé data (Hosseini Maasoum & Lichtinger, 2025): Junior employment drops sharply at GenAI-adopting firms; senior employment unchanged. AI as "seniority-biased technological change."
- European firms (Aldasoro et al., 2026): AI adoption → 4% labor productivity increase on average. Training investment amplifies gains (5.9 pp per additional % spent).
Who uses AI in practice? Unlike micro studies, real-world adoption is concentrated among middle-to-upper wage white-collar workers. Managers use AI at 2× the rate of frontline workers (BCG survey). This suggests AI may widen rather than narrow disparities in practice.
Why the Disconnect? (Author's Framework)
- Adoption friction — Only 36% of workers feel properly trained. People don't know the best models or productive use cases.
- Selection bias — Higher-skilled workers self-select into AI adoption, so the equalizing micro-level effect gets reversed at the macro level.
- Bottleneck tasks — A dev who codes 2× faster still waits for code reviews, meetings, and organizational processes. Non-AI tasks constrain overall throughput.
- Productivity J-curve — Firms are investing in reorganization, learning, and integration that suppresses measured output now but will pay off later (just as IT did in the 1990s).
Author's Prediction
AI will show up in aggregate productivity numbers quite soon, especially as agents automate end-to-end workflows and AI accelerates scientific research. The dynamics mirror early IT adoption — the gains are real but take time to propagate through organizations and measurement systems.
