cs.AIMar 21, 2026
SaveWhere can AI be used? Insights from a deep ontology of work activities
Organizations: Center for Collective Intelligence, Massachusetts Institute of Technology, 100 Main Street, Cambridge, 02142, Massachusetts, USA. · Sloan School of Management, Massachusetts Institute of Technology, 100 Main Street, Cambridge, 02142, Massachusetts, USA. · Mens, Manus, and Machina, Singapore-MIT Alliance for Research and Technology, 1 CREATE Way, Singapore, 138602, Singapore.
Abstract
Artificial intelligence (AI) is poised to profoundly reshape how work is executed and organized, but we do not yet have deep frameworks for understanding where AI can be used. Here we provide a comprehensive ontology of work activities that can help systematically analyze and predict uses of AI. To do this, we disaggregate and then substantially reorganize the approximately 20K activities in the US Department of Labor's widely used O*NET occupational database. Next, we use this framework to classify descriptions of 13,275 AI software applications and a worldwide tally of 20.8 million robotic systems. Finally, we use the data about both these kinds of AI to generate graphical displays of how the estimated units and market values of all worldwide AI systems used today are distributed across the work activities that these systems help perform. We find a highly uneven distribution of AI market value across activities, with the top 1.6% of activities accounting for over 60% of AI market value. Most of the market value is used in information-based activities (72%), especially creating information (36%), and only 12% is used in physical activities. Interactive activities include both information-based and physical activities and account for 48% of AI market value, much of which (26%) involves transferring information. These results can be viewed as rough predictions of the AI applicability for all the different work activities down to very low levels of detail. Thus, we believe this systematic framework can help predict at a detailed level where today's AI systems can and cannot be used and how future AI capabilities may change this.
Figures & tables
Figure 1: Excerpt of the ontology structure. (a) The top 4 levels of the ontology, (b) a sample of the next 5 levels beneath “Create information,” including generic activities, such as “Decide” and “Select,” and leading all the way down to atomic activities, such as “Select method.” At the last level are O*NET tasks, such as “Select banding methods to permit identification of chromosome pairs.” Collections are groupings of specializations that often differ on a dimension such as “Decide how?” or “Decide what?”
Figure 2: Demonstration of inheritance of the property AI Applicability . This stylized example illustrates how different values of the property of “AI applicability” can be inherited down different branches of the ontology from very abstract types of activity like “Create information” to very specific activities like “Write grant application for research projects.”
Figure 3: Where are AI software applications used today? Sunburst diagram illustrating the concentrations of AI software applications across activity types. Color intensity represents the proportion of AI applications associated with each activity (including its specializations in the outer rings) relative to the total number of AI applications in the dataset. We view this as a rough indicator of AI applicability . (Percentages in a ring sum to more than 100% when activities are classified in multiple places, but these double counts are removed at subsequent higher levels.) Note : For this and all other sunburst diagrams in the paper, readers are encouraged to zoom in to examine fine-grained specializations.
Figure 4: The 20 activities with the most AI software applications. These 20 activities alone account for more than 35% of all AI software applications.
Figure 5: Distribution of AI applications across activities within the “Create information” branch of the ontology . Consistent with the term “generative AI,” most current AI applications (58%) involve various forms of creating information, including images, videos, and other kinds of content.
Figure 6: Distribution of AI applications across activities within the “Create Information – other” branch of the ontology . The diagram includes detailed examples of activities both with and without AI applications, and as discussed in Section 3.1 , this provides a kind of checklist for different types of opportunities. (Readers are encouraged to zoom in to see detailed activity names in the diagram.)
Figure 7: Sunburst diagrams illustrating the changes in usage intensity of AI software applications from 2016 to 2025 . Each diagram represents the distribution of AI applications across ontology activities for a given year. Color intensity encodes the percentage of applications relative to the 2025 distribution. Color shading is thus fixed and directly comparable across years, enabling longitudinal comparison of category growth and concentration. This suggests substantial growth in the number of applications over time but relatively little broadening in the types of activities for which applications are available. (See larger diagrams for each year in Appendix 13.6 .)
Figure 8: Growth in AI software applications from 2016 to 2025. Significant growth in the number of applications began in 2022 with the emergence of generative AI, but it appeared to slow in 2024-25. The breadth of activities for which AI applications were available also increased over time, but at a much slower rate than the number of applications.
Figure 9: Where are robotic systems used today? Sunburst diagram showing the distribution of robotic systems deployed in 2024. Percentages and color intensity indicate the proportion of robotic systems associated with each activity relative to the total number of robotic systems deployed in that year (Percentages in a ring sum to more than 100% when activities are classified in multiple places, but these double counts are removed at subsequent higher levels.).
Figure 10: Top 20 robotic activities in 2024 (excluding “Clean floor” robots, which account for 76.7% of total installations).
Figure 11: Sunburst diagrams illustrating the changes in distribution of robotic systems across activities from 2015 to 2024 . Each diagram shows how installations of robotic systems are distributed across ontology activities for a given year. Color intensity encodes the proportion of robotic systems deployed in an activity, relative to the total robotic systems deployed that year. In general, this figure suggests that there may be less broadening over time of robot uses across different types of activities than of AI software applications. (See larger diagrams for each year in Appendix 14.3 .)
Figure 12: Overall view of how AI is used today. Both AI software applications and robotic systems are included. Numbers and color intensity show the percent of market value associated with each activity in the ontology. (Percentages in a ring sum to more than 100% when activities are classified in multiple places, but these double counts are removed at higher levels.)
Figure 13: “Sell online” and “Sell in retail store” are specializations of the generic sales process “Sell product.” Sub-activities that are different from those in the generic process are shaded. Adapted from Malone et al. (1999) .
Figure 14: Flow chart demonstrating the ontology construction process, including steps: (1) extracting verb-object pairs from O*NET items (to achieve atomicity ), (2) assigning appropriate definitions from WordNet for each verb in context (to achieve specificity ), (3) organizing activities into hierarchical structure using WordNet’s “hypernym” (generalization) trees, and (4) significant human editing to improve the structure (to achieve deep similarity ).
Figure 15: Overview of the O*NET hierarchy structure with an example of a “woodworking” occupation and its associated GWAs, IWAs, DWAs, and tasks. The top-level activities in (a) can be found at https://www.onetonline.org/find/descriptor/browse/4.A.
| Type of activity | Activity level | Examples | Number of activities |
|---|---|---|---|
| Generic activities | 1 | Act | 1 |
| 2 | Act on information, Act on physical objects, Act with other actors and activities | 3 | |
| 3 | Create information, Modify information, Transfer information, Store information, Create physical objects, Modify physical objects, … | 10 | |
| 4 | Decide, Calculate, Plan, Fabricate, Buy, Improve, … | 69 | |
| Cumulative 1 - n-3 | Select, Count, Budget, Cast, Rent, Optimize, … | 1,113 | |
| Select (Information), Select (Actor), Budget (Activity), Cast (Physical Object), Optimize (Information), … | 1,551 |
Table 1: Counts and examples across different levels of the ontology. Note that the Total row is not the direct sum of the rows above, but rather counts only once the activities that exist at multiple levels.
Figure 16: Example of how O*NET tasks break down into atomic tasks. Demonstration of how this breakdown helps with the analysis of AI applicability, because each atomic task requires different abilities.
Figure 17: Demonstration of the importance of specificity , achieved using word sense disambiguation, in the case of “Make”, with distinct senses meaning creation and causation; “Produce” and “Create” also are synonyms of the Creation meaning.
| GWA Branch | GWA Title | DWA ID | DWA Title (Inspection Activity) |
| Branch 1: Information Input Inspecting Equipment, Structures, or Material | |||
| 4.A.1.b.2 | Inspecting Equipment, Structures, or Material | 4.A.1.b.2.I07.D18 | Inspect equipment to ensure proper functioning |
| 4.A.1.b.2 | Inspecting Equipment, Structures, or Material | 4.A.1.b.2.I10.D03 | Inspect shipments to ensure correct order fulfillment |
| Branch 2: Information Input Monitor Processes, Materials, or Surroundings | |||
| 4.A.1.a.2 | Monitor Processes, Materials, or Surroundings | 4.A.1.a.2.I02.D09 | Inspect operational processes |
| 4.A.1.a.2 | Monitor Processes, Materials, or Surroundings | 4.A.1.a.2.I11.D02 | Inspect products or operations to ensure standards are met |
Table 2: ONET scatters functionally similar inspection activities across three separate GWA branches. All six DWAs share the core cognitive operation of inspection (perceive, compare, judge), yet ONET’s hierarchy places them under different top-level categories: two under “Inspecting Equipment,” three under “Monitor Processes,” and one under “Handling and Moving Objects.” In our ontology, all six activities inherit from a common “Inspect” ancestor, enabling AI applicability to propagate across the entire inspection family.
Figure 18: User interface for ontology platform, including (a) an outline view of the specialization hierarchy, starting with the most generic activity type called “Act,” (b) an editable view of the chosen activity (“Select”), including properties such as Title, Description, Generalizations, and Specializations, and (c) a search bar to find activities by keywords or semantically similar phrases.
Figure 19: Human-AI inter-annotator agreement for three GPT-5.1 (high-reasoning-effort) classification pipeline variants: SPPO = Single Prompt Partial Ontology, MPPO = Multi-Prompt Partial Ontology, SPFO = Single Prompt Full Ontology. Left: Mean Wu-Palmer Similarity. Right: Weighted Cohen’s . Plots show 95% confidence intervals computed via nonparametric bootstrap with 1,000 resamples.
Figure 20: A diverging stacked bar chart showing human evaluation of model classifications relative to their own. The skew toward the right (red) indicates that in over 15% of cases, human annotators rated the AI’s classification and its reasoning as superior to their own initial judgment. The skew toward the left (blue) indicates that in over 18% of cases, human annotators rated the AI’s classification and its reasoning as inferior to their own initial judgment. Overall, human annotators judged the AI classifications to be as good as their own in over 60% of cases, with roughly comparable proportions rated as better than (over 15%) or worse than (over 18%). With the final model we used (SPFO), human annotators rated the AI classifications as equal to or better than their own in 81.4% of cases.
Appendix figures & tables41 assets
Supplementary material from the paper’s appendix.
Appendix
| Study sponsor | Anthropic | Microsoft | OpenAI |
|---|---|---|---|
| LLM employed | Claude | Copilot | ChatGPT |
| Approximate number of conversations with LLMs in dataset | 1 million | 200,000 | 1 million |
| Method to map user conversations | Claude used to position each conversation within three-level hierarchical groupings of O*NET tasks developed by the Anthropic research team | GPT-4o used to map each conversation to best-fitting O*NET Intermediate Work Activity (IWA) | GPT-5 mini used to map each conversation to the best-fitting O*NET Intermediate Work Activity (IWA) |
| Level of aggregation at which task classification is reported | O*NET tasks and associated occupations/skills | IWAs, which are linked to relevant Generalized Work Activities (GWAs); IWA data then used to assess potential future impact of LLMs on occupations | IWA data aggregated and reported at GWA level; IWA data also mapped to 24 categories created by OpenAI to classify features and capabilities of current and future GPTs |
Table 3: Summary of approaches used by Anthropic, Microsoft, and OpenAI in studies of users’ conversations with their LLMs. 0 0 footnotetext: Source: Anthropic ( Handa et al. (2025) ; Appel et al. (2025) ), Microsoft ( Tomlinson et al. (2025) ), and OpenAI ( Chatterji et al. (2025) )
| Activity | Generalizations |
|---|---|
| Agree (Agree.v.01) | Collaborate (Collaborate.v.01) |
| Exchange information (Change.v.06) | |
| Brush (Brush.v.04) | Reorient physical objects |
| Touch (Touch.v.01) | |
| Casket (Casket.v.01) | Cover (Cover.v.01) |
| Enclose (Enclose.v.03) |
Table 4: Activities and their parent generalizations. .
| Variant | Ontology context | Model calls | Summary |
|---|---|---|---|
| SPPO (Single Prompt Partial Ontology) | Retrieved shortlist (top- ) | 1 | Retrieve candidate nodes using application text; one prompt extracts the verb-object activity and selects the best node from the shortlist. |
| MPPO (Multi-Prompt Partial Ontology) | Retrieved shortlist (top- ) | 2 | First prompt extracts a clean verb-object activity; retrieval uses this activity; second prompt selects the best node from the shortlist. |
| SPFO (Selected) | Full ontology (cached) | 1 | Cache the full ontology and use one prompt to extract the verb-object activity and select the best node from the full hierarchy (no retrieval). |
Table 5: Summary of automated classification pipeline variants evaluated in the human validation study. All variants used the same base model (GPT-5.1, high-reasoning-effort).
| Mean Wu-Palmer Similarity | Weighted Cohen’s | |||||
| Interns | Experts | All | Interns | Experts | All | |
| Human-SPFO | 0.591 [0.518, 0.666] | 0.631 [0.561, 0.703] | 0.619 [0.559, 0.680] | 0.347 [0.233, 0.456] | 0.411 [0.304, 0.509] | 0.392 [0.296, 0.476] |
| Human-SPPO | 0.583 [0.504, 0.660] | 0.585 [0.511, 0.655] | 0.584 [0.519, 0.645] | 0.343 [0.225, 0.451] | 0.349 [0.242, 0.444] | 0.347 [0.254, 0.427] |
| Human-MPPO | 0.554 [0.473, 0.635] | 0.619 [0.543, 0.697] | 0.600 [0.528, 0.670] | 0.314 [0.195, 0.421] | 0.412 [0.304, 0.514] | 0.383 [0.282, 0.470] |
| Human-Human | 0.525 [0.465, 0.586] | 0.584 [0.525, 0.649] | 0.557 [0.514, 0.609] | 0.258 [0.161, 0.349] | 0.348 [0.260, 0.433] | 0.307 [0.234, 0.372] |
| AI-AI | 0.812 [0.726, 0.895] | 0.703 [0.566, 0.822] | ||||
Table 6: Inter-annotator agreement (IAA) across human annotation groups, using Mean Wu-Palmer Similarity and Weighted Cohen’s (mean with 95% confidence intervals).
Figure 21: Distribution of human-AI annotation similarity (mean Wu-Palmer similarity) across rating categories. When human evaluators rated the AI’s classification as equal to their own, the corresponding human-AI annotation pairs showed higher similarity. In contrast, when the AI’s classification was rated as better or worse than the human’s, annotation similarity was lower.
| Name | &facts |
|---|---|
| Link | https://www.andfacts.com/ |
| Type | AI |
| Launch Date | 10-Jun-23 |
| Task | Market analysis |
| Tagline | Analyze market signals for real-time consumer insights. |
| Description | &facts is a market intelligence platform for consumer brands that analyzes real-time market and product signals to identify trends, opportunities, and risks. It supports product development, marketing optimization, competitor analysis, and international expansion by providing AI-driven insights into consumer preferences and behavior without the need for extensive manual research or consultants. |
Table 7: Sample AI application from the TAAFT dataset.
| Name | Launch Date | Tagline | Description | Classification |
|---|---|---|---|---|
| &AI | 1-Jan-24 | AI-powered patent due diligence in seconds | &AI streamlines patent due diligence by rapidly generating detailed claim charts, performing prior art searches, and analyzing claim robustness. The platform supports exporting results to Word or Excel, semantic search across proprietary and USPTO documents, and interactive querying through a patent-aware chatbot. Its goal is to reduce the time and cost associated with patent analysis while improving coverage and reliability. | Research (Information) |
| &facts | 10-Jun-23 | Analyze market signals for real-time consumer insights. | &facts is a market intelligence platform for consumer brands that analyzes real-time market and product signals to identify trends, opportunities, and risks. It supports product development, marketing optimization, competitor analysis, and international expansion by providing AI-driven insights into consumer preferences and behavior without the need for extensive manual research or consultants. | Analyze Market |
| 008 Agent | 22-Nov-23 | AI-powered open-source softphone revolutionizing VoIP. | 008 Agent is an open-source, AI-powered softphone designed for intelligent VoIP communication. It captures and processes call events, transcribes conversations, performs sentiment analysis, and generates summaries. The system integrates with CRM tools, supports programmable conversational agents, and provides transparent, customizable workflows through a community-driven development model. | Communicate (transfer information end-to-end) |
| 04-x | 1-Aug-23 | Privacy-first chat with powerful AI models | 04-x is a privacy-focused chat platform for interacting with large language models. It supports multiple commercial and open-source models, encrypts all user data end-to-end, and allows tasks such as conversational interaction, summarization, and text generation. The system emphasizes data ownership, security, and extensibility through continuous model integration. | Converse |
| 0gpt | 9-Sep-24 | Unmask AI-generated text with precision. | 0gpt is an AI content detection tool that distinguishes between human-written and AI-generated text using linguistic and statistical analysis. It evaluates sentence structure, vocabulary diversity, and contextual coherence, highlights suspected AI-generated segments, supports multiple languages, and ensures privacy by not storing analyzed content. The tool is free and designed for broad accessibility. | Classify Content |
Table 8: Examples of AI applications and their classification results.
Figure 22: Sunburst diagram of the complete ontology with aggregated counts of AI software applications. Color intensities and percentage values represent the proportion of AI applications classified directly in each activity and in all of its specializations , relative to the total number of AI applications in the dataset. Readers are encouraged to zoom in to examine fine-grained specializations.
Figure 23: Sunburst diagram of the complete ontology with direct counts of AI software applications. Color intensities and percentage values represent the proportion of AI applications directly classified directly into each activity, relative to the total number of AI applications in the dataset. Readers are encouraged to zoom in to examine fine-grained specializations.
| Atomic Activity | O*NET Tasks |
|---|---|
| Create content (3.53%) | (O*NET) 16193 — Collaborate with Web, multimedia, or art design staffs to create multimedia Web sites or other internet content that conforms to brand and company visual format. |
| (O*NET) 5219 — Plan and develop instructional methods and content for educational, vocational, or student activity programs. | |
| Develop application (1.79%) | (O*NET) 16184 — Develop transactional Web applications using Web programming software and programming languages such as HTML and XML. |
| (O*NET) 16786 — Develop new software applications or customize existing applications to meet specific scientific project needs. | |
| (O*NET) 17706 — Develop or maintain applications that process biologically based data into searchable databases for analysis, calculation, or presentation. | |
| (O*NET) 18093 — Develop, implement, or evaluate health information technology applications, tools, or processes to assist with data management. |
Table 9: Selected atomic activities and associated O*NET tasks in the “Create information – other” branch.
Figure 24: Distribution of AI applications in 2016.
Figure 25: Distribution of AI applications in 2017.
Figure 26: Distribution of AI applications in 2018.
Figure 27: Distribution of AI applications in 2019.
Figure 28: Distribution of AI applications in 2020.
Figure 29: Distribution of AI applications in 2021.
Figure 30: Distribution of AI applications in 2022.
Figure 31: Distribution of AI applications in 2023.
Figure 32: Distribution of AI applications in 2024.
Figure 33: Distribution of AI applications in 2025.
| Handling Operation | Welding | Dispensing | Cutting/ Processing | Assembling | Others |
|---|---|---|---|---|---|
| Metal casting | Arc welding | Painting, enamelling | Laser cutting | Assembling | Cleanroom for FPD |
| Plastic molding | Spot welding | Sealing material | Water jet cutting | Assembling unspecified | Cleanroom for semiconductors |
| Stamping, forging, bending | Laser welding | Other dispensing, spraying | Grinding | Disassembling | Cleanroom for others |
| Handling operation at machine tools | Welding other | Dispensing unspecified | Processing other | Service applications | |
| Machine tending | Soldering | Processing unspecified | Medical applications | ||
| Measurement, inspection, testing | Welding unspecified | Others |
Table 10: IFR taxonomy of industrial robots. Column headings are classes and items within each column are subclasses.
| Agriculture | Cleaning | Inspection | Construction/ Demolition |
|---|---|---|---|
| Cultivation | Clean floor | Inspect building | Construction demolition |
| Milking | Disinfect | Inspect other | |
| Farming other | Cleaning other | ||
| Transportation/ Logistics | Search/Rescue/ Security | Hospitality | Other |
| Indoors without public | Search, rescue services | Prepare food | Other |
| Indoors with public | Security | Mobile guidance, telepresence |
Table 11: IFR taxonomy of service professional robots.
| Medical robots |
|---|
| Diagnostics and medical analysis |
| Perform surgery |
| Rehabilitation and non-invasive therapy |
| Other medical robots |
Table 12: IFR taxonomy of medical robots.
| Domestic tasks | Social interaction, education | Care at home | Consumer other |
|---|---|---|---|
| Cleaning floors indoor | Social interaction, companions | Mobility assistants, assist eating | Other consumer robots |
| Window cleaning, gardening and other domestic tasks | Education | Manipulation aids and other care robots | |
| Cleaning outdoors |
Table 13: IFR taxonomy of service consumer robots.
| IFR Subclass | Ontology Work Activity |
|---|---|
| Industrial robots | |
| Metal casting | Pour Metal |
| Handling for plastic molding | Mold Material |
| Handling for stamping/forging | Shape Material |
| Handling operations at machine tools | Operate Tool |
| Machine tending for other processes | Tend Machine |
Table 14: Mapping of IFR subclasses to ontology work activities.
Figure 34: Sunburst diagram showing the distribution of robotic systems deployed in 2024 excluding the node “Clean floor”.
Figure 35: Distribution of robot units in 2015.
Figure 36: Distribution of robot units in 2016.
Figure 37: Distribution of robot units in 2017.
Figure 38: Distribution of robot units in 2018.
Figure 39: Distribution of robot units in 2019.
Figure 40: Distribution of robot units in 2020.
Figure 41: Distribution of robot units in 2021.
Figure 42: Distribution of robot units in 2022.
Figure 43: Distribution of robot units in 2023.
Figure 44: Distribution of robot units in 2024.
| Industry segment | 2024 market share (%) | 2024 market value (USD B) |
| Industrial | ||
| Automotive | 7% | 3.2 |
| Electrical/Electronic | 6% | 2.8 |
| Logistics | 5% | 2.3 |
| Metal | 2% | 0.9 |
| Chemical | 2% | 0.9 |
Table 15: Market share and corresponding market value by robotic industry (2024).
| Subclass | Units | Price range (USD) | Initial estimate of avg price (USD) | Relative avg price (ratio) | Adjusted avg price (USD) | Market value (USD) |
|---|---|---|---|---|---|---|
| Diagnostics/ Lab analysis | 3,293 | 100k–150k | 125k | 1.3 | 145k | 0.5 bn |
| Surgical | 6,612 | 600k–2,500k | 1550k | 16.4 | 1,800k | 11.9 bn |
| Rehab/ non-invasive therapy | 5,759 | 69k–120k | 94.5k | 1.0 | 110k | 0.6 bn |
| Other | 995 | 100k–150k | 125k | 1.3 | 145k | 0.2 bn |
| Total (USD) | 13.2bn |
Table 16: Estimated Market Value of Medical Robotics Subclasses (2024) . The Medical Robotics industry represents 29% of the global robotics market value (13.2bn USD). Average prices for each subclass are initially estimated as the midpoint of the reported price ranges and then converted to relative prices. The relative prices are multiplied by the price adjustment factor ( ) to obtain adjusted price estimates. Then, the adjusted price estimates are multiplied by the number of units to calculate the market value of each subclass.
Figure 45: The global market value distribution of AI software applications in 2024.
Figure 46: Sunburst diagram showing the global market value of robotic systems (industrial and service robots) for the year 2024.
Figure 47: Sunburst diagram showing the percentage contributions to market value of AI software applications and robotic systems for each work activity in 2024. Red shading indicates activities performed by AI software applications. Blue shading indicates activities performed by robotic systems. And purple shading indicates activities performed by a combination of both.
More in this topic
AI and Labor
AI and labor studies how artificial intelligence changes jobs, wages, skills, hiring, and the organization and quality of work.
See all 94 papers