Working with AI: Measuring the Applicability of Generative AI to Occupations
Labor performed with AIpaperactive
Working with AI: Measuring the Applicability of Generative AI to Occupations: a paper with results files (metrics aggregated at the work-activity and occupation level). Operator: Microsoft Research (Kiran Tomlinson, Sonia Jaffe, Will Wang, Scott Counts, Siddharth Suri). Its unit: an anonymized Bing Copilot conversation, classified by an LLM pipeline to the O*NET intermediate work activities (IWAs) it serves as a user goal and as an AI action; metrics aggregated to IWAs and to detailed 2018 SOC occupations. Whose labor: the O*NET work activities that Bing Copilot users seek help with (user goals) and that the AI performs (AI actions), mapped to the occupations whose tasks contain them.
The goal of these metrics is to measure how AI may be useful in different occupations.
In its own words: https://raw.githubusercontent.com/microsoft/working-with-ai/main/README.md, read 2026-10-29
Who supplies, who directs, who gains
Supplies the laborBing Copilot users bring the goals; the AI performs actions, recorded separately as user goals and AI actions
Directs itthe user's goal, as classified; the paper states its methodology allows it to predict which occupations are more likely to delegate tasks to AI and which are more likely to use AI to assist existing workflows
Receives the gains, bears the costsnot recorded
Each as the resource itself records it. Where it records nothing, the census says so.
What it records
| Intersections | Labor performed with AI: assistance, delegation, task fidelity, correction burdens and user labor drag |
| Kind of evidence | interaction dataexposure estimatedata access |
| Unit of record | an anonymized Bing Copilot conversation, classified by an LLM pipeline to the O*NET intermediate work activities (IWAs) it serves as a user goal and as an AI action; metrics aggregated to IWAs and to detailed 2018 SOC occupations op |
| Whose labor | the O*NET work activities that Bing Copilot users seek help with (user goals) and that the AI performs (AI actions), mapped to the occupations whose tasks contain them |
| Method | an LLM-based pipeline classifying 200k anonymized Bing Copilot conversations to O*NET work activities, with completion, impact-scope and thumbs-feedback measures combined into an occupation-level AI applicability score op |
| Coverage | geography: United States (Bing Copilot conversations); period: 1 January 2024 to 30 September 2024 op |
| What it finds | the most common and successful AI-assisted work activities involve information work; widespread AI applicability cuts across sectors, as most occupations have information work components; occupation-level AI applicability increased over the 9-month data period, driven primarily by people using AI for more varied tasks |
| Its stated limits | the metrics are not to be read as measuring the ability of AI to replace jobs; success measures are proxies; consumer usage does not represent the full distribution of AI use in the workplace; LLM-based classifiers are imperfect; United States only, January to September 2024 |
op the resource's own statement · 3p a third-party account · obs our observation · rd our reading. Hover a tag for its quote; each opens below with its source.
What this evidence can settle
Can answer
- which work activities a consumer LLM (Bing Copilot) is used for, as user goal and as AI action
- which of those uses are relatively more or less successful, by its proxies
- which occupations may find AI more applicable, by an applicability score
Cannot answer
- whether AI can replace jobs
- employment or wage change
- AI use through workplace-integrated or specialised tools
- use outside the United States or after September 2024
- who receives gains from the assisted work
In the worked questions
Where it meets Labor Vector
rd The archive's reading: the separation of user goal from AI action, and the completion classifier ('the LLM completed the user's goal'), are the resource's own instruments for whether the system's labor served the user's commission; they meet DSL there, measured by classifier and thumbs feedback. Task-origin is recorded as the user goal.
Entity
| Canonical name | Working with AI: Measuring the Applicability of Generative AI to Occupations |
| Also known as | Working with AI, AI applicability score, "Working with AI" results files, microsoft/working-with-ai |
| Address | https://github.com/microsoft/working-with-ai |
| Operator | Microsoft Research (Kiran Tomlinson, Sonia Jaffe, Will Wang, Scott Counts, Siddharth Suri) op |
| Founded | 2025-07-10 (arXiv v1) |
| Kind | paper with results files (metrics aggregated at the work-activity and occupation level) |
| Lifecycle | active — a fixed 2024 data window; the paper revised to arXiv v7 (23 October 2026) for NeurIPS 2026; results files at v1.1; the operator's publication page at Microsoft Research still dates the work July 2025 and shows the abstract of the July 2025 version (applicability scores by occupation group), and links a Microsoft Research note, 'Applicability vs. job displacement: further notes on our recent research on AI and occupations' |
| Count | 200k conversations; 785 SOC codes covering 149.8 million workers (2023 OEWS) (as of results v1.1 / arXiv v7 (2026-10-23)) op |
| Data access | results files on GitHub (IWA-level and SOC-level metrics, applicability scores, weights) |
| Licence | CC BY 4.0 op |
| Versions and dates | arxiv v1: 2025-07-10; msr publication page: July 2025 (the page's date and abstract, read 2026-10-29); arxiv v7: 2026-10-23; results: v1.1; venue: NeurIPS 2026; read: 2026-10-29 |
| Observation | 2026-10-29: read |
Relations
Neighbours
Captures: the composition layer's account
- «Microsoft Working with AI» on Google AI Overview, 2026-10-29
the composition layer's account of the entity at its common name, signed out, seated in the Capture Registry with its dimensional PER; at this name the composition gives Microsoft's own AI products and workforce, with the entity's own page as its first card
The /non row
Evidenceevery field's quote, its source and the hash of the saved copy25
- self description op
The goal of these metrics is to measure how AI may be useful in different occupations.
- method op
Drawing instead on real-world AI usage, we analyze a dataset of 200k anonymized conversations with Microsoft Bing Copilot to measure AI applicability to occupations. We use an LLM-based pipeline to classify the O*NET work activities assisted or performed by AI in each conversation.
- unit op
In particular, they include our metrics aggregated at the O*NET intermediate work activity (IWA) and occupation (detailed 2018 SOC code) level.
- direction op
`iwa_metrics.csv`: IWA-level metrics, reported separately for user goals and AI actions.
- completion op
where our completion classfiier indicates that the LLM completed the user’s goal.
- directs op
Our methodology also allows us to predict which occupations are more likely to delegate tasks to AI and which are more likely to use AI to assist existing workflows.
- operator op
This research was conducted by members of [Microsoft Research](https://www.microsoft.com/en-us/research/).
- msr page date op
Kiran Tomlinson , Sonia Jaffe , Will Wang , Scott Counts , Siddharth Suri July 2025 Download BibTex
- msr page abstract op
We find the highest AI applicability scores for knowledge work occupation groups such as computer and mathematical, and office and administrative support, as well as occupations such as sales whose work activities involve providing and communicating information.
- msr page note op
Blog & Podcasts Applicability vs. job displacement: further notes on our recent research on AI and occupations
- coverage op
The conversations contributing to these results are from Bing Copilot in the United States from January 1, 2024 to September 30, 2024, so may not represent AI usage in other countries.
- count op
This leaves us with 785 SOC codes covering 149.8 million workers according to the 2023 BLS Occupational Employment and Wage Statistics (OEWS) data (total US employment in 2023 OEWS data is 151.9 million).
- finding information work op
We find that the most common and successful AI-assisted work activities involve information work--the creation, processing, and communication of information.
- finding sectors op
At the occupation level, we find widespread AI applicability cutting across sectors, as most occupations have information work components.
- finding trend op
Finally, we see an increase in occupation-level AI applicability over our 9-month data period, driven primarily by people using AI for more varied tasks.
- limit replacement op
Our metrics should not be misconstrued or misrepresented as measuring the ability of AI to replace jobs. There is more to an occupation than the sum of its O*NET tasks, and our measures of success (completion, scope, and thumbs feedback) are proxies that do not capture full workplace capabilities.
- limit scope op
Rather, our metrics indicate (1) what a consumer LLM is used for, (2) in a relative sense, which of these uses are more or less successful, and (3) which occupations may find AI more useful or applicable as a result.
- limit classifier op
LLM-based classifiers are imperfect, so all metrics should be interpreted with caution.
- limit consumer op
As these results are based on consumer AI usage, they do not represent the full distribution of AI use in the workplace.
- license op
This repository is licensed under CC BY 4.0.
- version op
(current) v1.1: referenced by the NeurIPS '26 version and by arXiv paper revisions v6 and v7.
- arxiv v7 op
[Submitted on 10 Jul 2025 ( v1 ), last revised 23 Oct 2026 (this version, v7)]
- venue op
50 pages; NeurIPS 2026
- ethics op
Data analysis was conducted with approval from the Microsoft Institutional Review Board.
- cited by aei 3p
See also Tomlinson et al (2025) for a related AI applicability score.
Negative tagswhat it is not4
- NOT a measure of the ability of AI to replace jobs
- NOT an employment outcome measure
- NOT a measure of all workplace AI use: one consumer product, Bing Copilot, in the United States
- NOT independent of the provider: a platform-specific observation of Microsoft's own product, published by Microsoft Research
Sourcesevery source read, with its date and hash4
Not yet checkedwhat remains to read4
- the paper's full text (classifier prompts, interrater reliability, validation)
- the Microsoft Research note 'Applicability vs. job displacement: further notes on our recent research on AI and occupations', linked from the publication page and not fetched
- the results CSV files themselves
- the Copilot-Thumbs feedback data
Cite this record
Lee Sharks, ed. "Working with AI: Measuring the Applicability of Generative AI to Occupations." Labor Vector meta-aggregator, laborvector.org, compiled 2026-10-29. https://www.laborvector.org/aggregators/microsoft-working-with-ai
Record https://www.laborvector.org/aggregators/microsoft-working-with-ai · entity https://www.laborvector.org/aggregators/microsoft-working-with-ai#entity · the census as data
@misc{lv-microsoftworkingwithai2026,
editor = {Sharks, Lee},
title = {Working with AI: Measuring the Applicability of Generative AI to Occupations},
howpublished = {Labor Vector meta-aggregator, laborvector.org},
year = {2026},
note = {Compiled 2026-10-29; every field sourced and dated},
url = {https://www.laborvector.org/aggregators/microsoft-working-with-ai}
}