AI Adoption & Diffusion Across the Economy

Hub page for evidence on how generative AI is being adopted and what it is being used for, drawing on three complementary measurement strategies.

Three complementary measurement lenses

Lens	Source	Population	Unit	Key signal
Enterprise leader survey	sources/2025-korst-wharton-gen-ai-enterprise-adoption	~800 US decision-makers at 1000+-employee firms	Self-reported usage & ROI	82% weekly, 46% daily; 72% measuring ROI; 75% positive returns
Anthropic platform usage	sources/2025-handa-which-economic-tasks-ai	4M+ Claude.ai conversations	O*NET task per conversation	~36% of occupations use AI for ≥25% of tasks; 57% augmentation / 43% automation
Microsoft Copilot usage	sources/2025-tomlinson-working-with-ai	200K Bing Copilot conversations	O*NET IWA (user-goal + AI-action split)	Cross-occupation applicability score; info work dominates

A fourth, methodologically distinct, anchor measures actual productivity rather than adoption: sources/2025-becker-metr-ai-developer-productivity (METR RCT) finds frontier AI tools slow experienced OSS developers by 19%, despite forecasts of ~25–40% speedup.

Converging patterns

Information work dominates. Writing, learning, communicating, teaching/explaining, coding, and data analysis account for the bulk of usage across platforms.
Software engineering is the single biggest vertical in both Claude and Bing Copilot usage; confirms predictions from exposure-based forecasts (Webb 2019; Eloundou et al. 2023) with direct measurement.
Cognitive skills over physical or managerial. Reading Comprehension, Writing, Critical Thinking dominate; Installation, Equipment Maintenance, Negotiation barely appear.
Mid-to-high wage, bachelor's-degree peak. Usage is inverted-U across wages and preparation requirements, not monotonic with skill.
Mix of augmentation and automation. Neither pure collaboration nor pure replacement; 57/43 split in Handa et al. Tomlinson et al. extend this with a per-occupation user-goal/AI-action split, identifying which occupations shift focus vs. get assisted.

Tensions and contradictions

Perceived vs. actual productivity gap. Enterprise leaders report 75% positive ROI (sources/2025-korst-wharton-gen-ai-enterprise-adoption); experienced developers in RCT are 19% slower with AI while predicting 20% speedup (sources/2025-becker-metr-ai-developer-productivity). The gap is setting-specific — synthetic tasks and less-experienced workers still show speedups — but underscores that self-reported ROI may be inflated.
Augmentation-vs-automation is not a binary. Tomlinson et al. show the AI's action often differs from the user's goal (40% disjoint IWA sets); pure-automation framings understate the co-production pattern.
Skill atrophy as adoption grows. 43% of enterprise leaders warn of skill declines (sources/2025-korst-wharton-gen-ai-enterprise-adoption) — independently causally demonstrated in the novice-learning setting of concepts/ai-skill-formation.

Three-wave adoption trajectory (Wharton/GBK)

2023 — Exploration. 37% weekly use; "fascinated but cautious"; simple office tasks.
2024 — Experimentation. 72% weekly; +130% spend; pilots across functions.
2025 — Accountable Acceleration. 82% weekly; 46% daily; 72% measuring ROI; guardrails tightening; people (not tools) set the pace.

Sectoral ranking (converging across sources)

Leaders: Tech/Telecom, Banking/Finance, Professional Services, Computer & Mathematical, Sales & Related, Office & Admin Support. Laggards: Manufacturing, Retail, Healthcare Support, Farming, Construction, Building/Grounds Maintenance.

concepts/ai-skill-formation · concepts/ai-agent-benchmarks · concepts/agentic-bpm