If you’re training an agent to do work inside software, or buying data for a team that is, you’ve already hit the problem this guide is about. Language models learned from the written internet. Agents have to learn from work, and almost none of the work people do in a CRM, a ticketing system, a codebase or a spreadsheet was ever posted online. Read on to find out what agent training data is, why labs can’t scrape it, what labs are buying in 2026, how to judge a dataset before you license it, and the three ways to get one.
What agent training data is
An agent is a model that takes actions in software to finish a task. To learn that, it needs three kinds of data, and most serious programs use all three.
| Kind of data | What it teaches the model | Where it comes from |
|---|---|---|
| Workflow records | How a real task moves through real systems from start to outcome (a ticket opened, escalated and resolved; a deal logged, updated and closed) | Operating companies: ticketing, CRM, project, finance and engineering systems, with the history that links them |
| Action trajectories | The step-by-step actions a person takes to do a task on a computer (clicks, keystrokes, commands, what changed on screen) | Recordings of people using real software and websites, usually collected to order |
| Environments | Practice and scoring: a replica of the software with a goal and a verifier that says whether the agent succeeded | Built on top of the first two, from real workflows and real codebases |
Workflow records give the model the shape of the job. Action trajectories give it the mechanics. Environments give it somewhere to practice. Our guide to what RL environments are covers the third row in detail, and our glossary entry on workflow trajectories covers the unit of work the first two rows are built around.
Why the open web has none of it
Pretraining worked because people published text. Nobody publishes their support queue, their project history or their month-end close. That work lives inside company systems, and the parts that did leak onto the web (help-center articles, product documentation, tutorials) describe how software is supposed to be used, with none of the mess of how it’s used in practice.
That mess is the point. An agent that only saw the documentation knows what a CRM is. An agent that saw ten thousand real deals move through one, with the stalls, the reassignments and the fields nobody fills in, knows how to use one. The labs training agents have run out of the first kind of data and are paying for the second.
What labs are buying in 2026
Frontier labs like OpenAI, Anthropic and Google DeepMind, hyperscalers like Microsoft and Amazon, and well-funded newer labs like Reflection AI are all training agents that use software, and the data they want follows the three rows above.
- Real work product and the systems around it. Labs have hired contractors to hand over real work files and code repositories, and buy the operating records of companies (tickets, CRM data, project histories, code with its commit history) to learn what finished work looks like.
- Recordings of people using computers. Screen recordings with the actions attached, web navigation traces and software usage logs, collected at scale to teach agents to operate a browser or an operating system.
- Environments built from both. Replicas of enterprise software and real codebases where the agent practices and gets scored. Our guide to RL environment companies covers who sells them and what they cost.
Enterprise software companies like Salesforce and ServiceNow are training agents on the same kinds of records, from inside the products where the work happens. And a public price exists for one company’s operating history. Google bid 10 million dollars for Spirit Airlines’ internal archive in 2026, and our guide to who buys company data for AI covers what that sale says about the market.
What makes a dataset useful for agents
Four things separate agent data that trains a model from an export that fills a drive.
- Outcomes are visible. The record shows whether the ticket was resolved, the deal closed, the bug fixed. A model can’t learn to finish a task from data that stops before the finish.
- The systems connect. One ticket on its own is thin. The same ticket linked to the customer record, the engineer’s commit and the follow-up conversation is what a job looks like, and that cross-system context is the most expensive thing to recreate.
- Names and personal details are removed to a documented standard, so the records can be used without exposing customers or employees. Our guide to enterprise operational data covers what companies hold and how that scoping works.
- The rights are documented per asset. Workflow data is full of people and third parties, so an agent trained on records a company wasn’t allowed to share carries the same risk as a model trained on scraped books. The questions buyers ask about rights-cleared training data apply here in full.
Code gets its own checklist. Our guide to what code AI labs pay for covers why the commit history, tickets and reviews matter more than the lines.
Where to get agent training data
Three routes, and most labs use more than one.
- Collect it yourself. Hire people to do tasks on instrumented machines, or run your own agents and keep the trajectories. You control the format and the volume. You also pay for every hour, and the tasks are only as real as the briefs you write. Our guide to AI data collection companies covers the vendors that do this to order.
- Build from your own product. Enterprise software companies can train on the work that happens inside their tools, which is why Salesforce and ServiceNow are building agents from the inside. If you don’t own the software, this route is closed.
- License it from the companies that produced it. Operating companies hold years of real workflows with real outcomes and are starting to license them. This is the only route to the cross-system, outcome-complete records in the checklist above, and the one that gets you data no other lab can scrape.
Where Troveo fits
Troveo licenses business data from operating companies: support tickets, CRM records, project histories, transaction logs, workflows and code, with names and personal details replaced with pseudonyms and the rights documented per asset. We run the process from the first assessment of what a company holds through agreement, verification, snapshot and data extraction, and the same real workflows are what RL environments are built from. More than 7,000 rights holders have licensed through Troveo, 95 percent exclusively, and we’ve paid out more than 20 million dollars. If you’re training agents and need the records of real work, tell us what your agents need to learn from and we’ll show you what we hold.