The Task Graph connects descriptions of work to reference data and AI research. You get a clearer picture of what AI means for your role or organisation, with the evidence and limits behind the findings.
Start with the role and the tasks people actually do. Existing role descriptions and task lists are useful starting points.
02
Review the matches
The Task Graph connects the work to reference tasks and research. Review the proposed matches so the analysis reflects the work you recognise.
03
Understand the implications
The report brings together findings about tasks, roles and the wider work. Use it to decide where to investigate, build skills or rethink responsibilities.
For your organisation
A view across the work.
Bring organisational context, roles and the task detail you have. The intended analysis connects that work to reference evidence and brings the implications together across roles and units. Proposed tasks and assumptions need review.
Describe your work in your own words. We suggest tasks and connect them to reference roles and AI research where available. Check the connections, including partial matches, then create your private report. See what AI could change, where your judgement matters and what to learn or try next.
Your invitation includes one finalised analysis, without payment, in exchange for short feedback. No code? Join the waitlist on the personal report page.
Understanding the analysis
Different evidence. Different questions.
Research-based potential, recorded AI use and your own information stay separate. A reference finding is a starting point for judgement; it does not predict what will happen in your workplace.
Technical potential
Which tasks met a study’s threshold for AI-assisted work? This is a research assessment, not measured customer savings.
Recorded AI use
How was AI used in a particular sample? This describes those interactions, not adoption across all workers.
Your context
Which findings fit your tasks, tools and responsibilities? Your information and review determine the local interpretation.
A public task finding
Analyze employment-related data and prepare required reports.
Human Resources Specialists
MIT-WAL 2026 E1 · Agentic exposure
A frontier agent with browsing, tool use and standard workspace access could halve task time at equivalent quality, assuming the worker's access and permissions. This is a model assessment, not a successful execution test. This is a research finding, not measured savings in your organisation.
Separate evidence: Anthropic’s published task-use sample also records AI use for this task. It does not establish adoption in your organisation.
Illustrative implication to review
If AI helps prepare the report, review how checking source data, handling exceptions and handing over the output would change.
Who checks the data and remains responsible for the report?
Missing evidence remains visible. The analysis does not measure employees’ skills, predict job losses or generate an optimal organisation chart. Customers decide and test changes in their own setting.
Public reference data and published research provide distinct evidence. The Task Graph selects relevant findings and adds connections, plain-language labels and interpretation. It does not own the underlying public sources or imply their providers’ endorsement.
Agentic task exposure · MIT-WAL 2026
The public Taskgraph exposure score uses MIT Work Analytics Lab’s 2026 Gemini agentic task classifications. The rubric asks whether an agent could use at least 50% less task time at the same quality. Agents can browse and use standard workspace tools when the worker’s access and permissions are arranged. These are model assessments, not measured workplace results or execution tests.
The score summarises technical exposure across the assessed tasks on a 0–100 scale. Higher values indicate greater research-based potential for AI-assisted work. Coverage shows how much of the reference role has a matched assessment; missing assessments are excluded, not treated as E0.
This is not the percentage of a job automated, time saved or likelihood of job loss. Recorded AI use and other supporting research do not change the score.
E1 contributes 1 to the index, E2 and E3 contribute 0.5, and E0 contributes 0. Core tasks weigh twice other tasks. These are index contributions, not estimated time-saving percentages.
OpenAI’s 2023 GPTs are GPTs study remains historical comparison. Both model and rubric changed; the difference is not a measured trend in AI progress. Older classes never fill missing MIT-WAL classifications.
Read the MIT-WAL E0, E1, E2 and E3 task classes
E1 · Agent with standard tools
The threshold could be met with browsing and standard workspace tools.
E2 · Enterprise integration
Deeper enterprise integration or custom software is needed to meet the threshold.
E3 · Visual capabilities
Visual understanding or generation is the critical extra capability for meeting the threshold.
E0 · Threshold not met
The threshold was not met. Smaller gains and other uses remain possible.
Anthropic’s Economic Index release of 2026-06-26 describes Claude interactions linked to tasks. Different chat and connected-software samples stay separate, with their observation dates.
Earlier task-use findings from Anthropic’s labour-market research describe whether AI use was observed for a task in that study’s sample. They remain separate from the June monthly interaction shares. Each finding describes its own sample; neither establishes adoption in your workplace.
“Asked AI to do it” describes requests to produce a result, sometimes with feedback. “Worked with AI” describes shaping, learning or checking work together. The shares describe interactions with a clear work pattern, not successful automation or the share of people using AI.
Different observation periods contain different samples. A comparison does not establish a trend or statistically significant change. Missing observations do not mean no AI use.
OpenAI Signals v2.0 describes work-related ChatGPT messages on individual plans in the US. Its broader work-activity findings provide context for related tasks; they do not directly measure an individual task or occupation.
Message shares describe relative attention within that sample. They do not establish workplace adoption, time savings or growth in the absolute amount of work. Activity observations and task-use samples do not change the exposure score.
OpenAI Signals data is used under CC BY 4.0. The Task Graph adds connections and interpretation.
O*NET supplies reference occupations, tasks, skill definitions and occupation ratings. ESCO supplies European occupation and skills context where available. These describe typical work; they do not establish a customer’s responsibilities, employee proficiency or future skill demand.
O*NET® 30.0 Database, U.S. Department of Labor, Employment and Training Administration, used under CC BY 4.0. O*NET is a trademark of USDOL/ETA. Task connections and change interpretations are added by The Task Graph and do not imply USDOL/ETA endorsement.
This service uses the ESCO classification of the European Commission.
EconEvals modelled time-saving estimates and occupation-level benchmarks, including Microsoft’s Working with AI research, answer separate questions. They remain supporting context and do not change the task exposure score. Modelled estimates are not realised customer results.
AI-assisted explanations interpret the available evidence. They do not create the research values. Reports show the basis of findings, relevant dates, coverage and gaps. Review the reference connections and local assumptions, then choose a bounded experiment with a responsible owner and a way to assess the result.
Source credits and licences
Agentic task exposure · primary source
MIT Work Analytics Laboratory / Nicolò Piergiovanni Bagnoli, Gemini 3.1 Task AI Exposure Scores (2026). Dataset published under the MIT licence. Gemini is the assessment model; this is a separately published dataset, not an official update by the original OpenAI paper’s authors.
Earlier task exposure study · comparison
Tyna Eloundou, Sam Manning, Pamela Mishkin and Daniel Rock, GPTs are GPTs (2023). Task dataset: OpenAI, copyright 2024, under the MIT licence.
Anthropic Economic Index
Maxim Massenkoff, Eva Lyubich, Szymon Sacher, Zoe Hitzig, Shaoyi Zhang, Ryan Heller and Peter McCrory, Anthropic, Anthropic Economic Index report: Cadences (26 June 2026). Original report. Data released under CC-BY ( publisher’s licence notice).
Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Drew Johnston, Alex Martin Richmond, Christopher Ong, Carl Yan Shan and Kevin Wadman, OpenAI. Data under CC BY 4.0; source definitions and suggested citation.
