Operations and workflow data for AI training
Operations and workflow data records how work actually gets done: task histories with decisions and outcomes, field-service jobs, quality inspections, and recordings of people doing real work. AI labs license it to train agents and robots on real-world processes that public data barely covers.
Listings
| Data | Typical sources | Modality | Availability |
|---|---|---|---|
| Task histories with context, actions and outcomes | ServiceNow, Jira Service Management, Salesforce | Structured data and text | On request |
| Field-service work orders with technician notes and photos | ServiceTitan, Salesforce Field Service, IFS | Structured data, text and images | On request |
| Quality inspection records with photos and defect logs | SafetyCulture, Procore, ETQ Reliance | Images and structured data | On request |
| Screen recordings of real software workflows | Task-mining and session-recording tools a company already uses | Video and event logs | On request |
| Recordings of hands-on work | New recordings arranged for each project | Video | On request |
What’s included
- Work items with context, actions, decisions and outcomes
- Field-service jobs with technician notes, parts and photos
- Inspection checklists, defect photos and corrective actions
- Screen recordings and event logs of real software tasks
- New video recordings of hands-on work, coordinated with partner businesses
Public datasets and licensed data
Public agent benchmarks are built on crowdworker demonstrations or simulated companies. They’re useful tests, but small: hundreds to a few thousand tasks. Licensed workflow data adds years of real task histories from operating businesses, with outcomes attached.
| Public dataset | Released | What it contains | Limits for enterprise use | License |
|---|---|---|---|---|
| Mind2Web | Ohio State University, 2023 | 2,350 tasks on 137 real websites across 31 domains, with step-by-step demonstrations by crowdworkers. | Consumer websites and crowdworker demonstrations, not company workflows. | CC BY 4.0 (research use) |
| WebArena | Carnegie Mellon University, 2023 | 812 multi-step tasks on self-hosted shopping, forum, GitLab and content-management sites. | A benchmark environment with simulated sites and data. | Apache 2.0 |
| OSWorld | HKU, Salesforce Research, CMU and Waterloo, 2024 | 369 computer tasks on Ubuntu desktop and web applications, plus 43 Windows tasks. | A test set of hundreds of tasks, not a record of real work. | Apache 2.0 |
| TheAgentCompany | Carnegie Mellon University, 2024 | 175 tasks inside a simulated software company with its own code host, files, task tracker and chat. | The company and its data are simulated. | MIT |
Preparation and rights
Preparation requirements are agreed with each supplier before any work begins. For this kind of data they typically include:
- Pseudonymizing customer and employee identities and redacting free-text notes
- Getting consent from people who appear in recordings, under the company’s policies and local law
- Blurring or removing faces, addresses and documents visible in photos and video
Every dataset has a documented owner and confirmed licensing rights, and is delivered only after an executed agreement and the supplier’s authorization. See data governance on sourcex.si.
Use cases
More on sourcex.si
Questions
What is workflow data for AI agents?
Records of how real work moves from request to completion: who did what, in which system, with what information, and how it turned out. Agents learn multi-step processes from it, and outcomes give a way to score them.
Can SourceX arrange new recordings?
New recordings of hands-on work can be coordinated with partner businesses. Describe the tasks and settings you need in your request.
Need operations and workflow data?
Describe what you need: domain, volume, history, format and licensing terms. SourceX looks for companies that hold it and can license it.