Working with AI: Measuring the Applicability of Generative AI to Occupations

Summary

Tomlinson, Jaffe, Wang, Counts, and Suri (Microsoft Research, 2025) construct an empirical AI applicability score per occupation based on 200,000 anonymised Bing Copilot conversations, mapped to the O*NET Intermediate Work Activity (IWA) taxonomy. The study is the Microsoft-side complement to Anthropic's Economic Index paper (sources/2025-handa-which-economic-tasks-ai).

Method. An LLM-based pipeline classifies each conversation into all applicable O*NET IWAs along two axes simultaneously:

User goal — what the human is trying to accomplish.
AI action — what the AI does in service of that goal.

This two-sided classification is the paper's distinctive methodological move. Example: a user trying to learn how to print a document has a user-goal IWA of operate office equipment; the AI-action IWA is train others to use equipment. The 40% disjointness between user-goal and AI-action IWA sets (96% have more unique-to-one-side IWAs than common) shows that AI often performs a different activity than the one the user is trying to do.

Success metrics. Three orthogonal measures per conversation: completion (LLM judgement validated by thumbs up/down), user thumbs feedback, and scope (6-point ordinal scale for how broad an IWA the AI can assist with / perform). AI applicability score aggregates coverage × completion × scope, split further by user-goal side and AI-action side.

Findings.

Most frequent work activities are information work — creation, processing, and communication of information. GWAs most over-represented vs. total US workforce: Getting Information, Updating and Using Knowledge, Communicate with External People, Interpreting Information for Others.
Highest-scoring occupations: interpreters/translators (0.492), historians (0.462), writers/authors (0.454), sales reps of services (0.449), CNC tool programmers (0.419), customer service reps (0.408). Lowest: pile-driver operators, dredge operators, foundry mold makers, roofers (all ~0).
Major-group ranking: Computer & Mathematical (0.29), Sales & Related (0.29), Office & Admin Support (0.26), Community & Social Service (0.25); physical trades and healthcare support at the bottom.
Image generation and data analysis have notably lower completion × scope than text-based information tasks.
Distinctive patterns: over-assisted activities (purchase goods, execute financial transactions, perform athletic activities) vs. over-performed activities (train others, coach others, teach subjects). Useful for predicting which occupations may shift focus (AI takes over teaching/training subtasks) vs. which will be primarily assisted.

Contribution relative to Handa et al. Handa et al. map each conversation to one O*NET task (occupation-specific). Tomlinson et al. map each conversation to all relevant IWAs (cross-occupation). This enables occupation-applicability scoring even when no user explicitly comes from that occupation — the AI's demonstrated capability transfers across occupations that share IWAs.

Connections

Microsoft-platform complement to Anthropic-platform sources/2025-handa-which-economic-tasks-ai; together they anchor concepts/ai-adoption at the bottom-up usage level.
Enterprise-leader-side complement to sources/2025-korst-wharton-gen-ai-enterprise-adoption.
Info-work dominance + image/data-analysis weakness contextualises the methodology of sources/2024-xu-the-agent-company-benchmark and the productivity surprise in sources/2025-becker-metr-ai-developer-productivity.
User-goal / AI-action split is relevant to concepts/agentic-bpm agent-role framing and concepts/agent-process-observability trajectory analysis.
New entities: entities/kiran-tomlinson, entities/sonia-jaffe.