An RL (reinforcement learning) environment is a simulated workplace where an AI agent practises multi-step tasks and an automatic grader scores each attempt. It has three parts: working copies of the software a worker uses, realistic data inside that software, and tasks paired with checks that decide whether the agent succeeded. Labs now treat environments as a core training input: Anthropic's leaders discussed spending more than $1 billion on them over a year, The Information reported in September 2025. Real company records make environments realistic, so builders license them. One data vendor's chief executive said in August 2026 that it had committed more than $20 million in 11 days to license operational data from companies for its environments.
- Environments need data that looks like a real company: messy, linked and with years of history.
- Enterprise workflows such as CRM work, reports and spreadsheets are the fastest-growing area, according to Epoch AI's interviews.
- Records that show a task and its correct outcome, such as a ticket and its resolution, are the most useful.
What an RL environment is
In reinforcement learning, a model is given a task and a set of actions it can take. It attempts the task, a grader scores the attempt, and the scores are used to update the model. The environment is everything around the task: the files, the applications and their state. Epoch AI gives examples such as a code repository with a failing test to fix, or a clone of a booking site where the agent must find the cheapest listing.
For agents meant to do office work, the environment is a replica of a company: its ticketing system, CRM, inbox and shared drives, filled with records that look real. Builders also call them RL gyms. The RL environment companies tracker lists who builds them for coding, computer use, finance and enterprise apps.
| Part | What it is | Where company data fits |
|---|---|---|
| Replica software | Working copies of the apps a worker uses: CRM, ticketing, email, spreadsheets, code repositories | Your configuration, fields and workflows show builders what a real setup looks like |
| Data inside the replica | The records the agent reads, edits and acts on | De-identified tickets, messages, documents and records make the workplace believable |
| Tasks and graders | A goal, such as "resolve this ticket", and an automatic check of the result | Resolved tickets, closed jobs and approved reports show what a correct outcome looks like |
Source: Epoch AI: An FAQ on reinforcement learning environments (Jan 2026)
Why labs are spending on environments now
Labeled datasets powered the chatbot era; environments are becoming the equivalent for agents. Leading labs are demanding more environments and also buying them from third-party vendors, TechCrunch reported in September 2025, and a new group of well-funded startups has formed to build them.
For a January 2026 report, Epoch AI interviewed 18 people across environment startups and labs. They said math and coding tasks came first, and that enterprise workflows such as navigating a CRM, filing reports or editing spreadsheets are now a major growth area. A top quality concern was reward hacking: models finding ways to game a grader. Realistic data with genuine edge cases makes that harder, which is why builders look for real company records.
Buyers building simulated workplaces look for varied, deep records, including disagreements and judgment calls, MIT Technology Review reported in October 2026.
Source: TechCrunch: Silicon Valley bets big on environments to train AI agents (Sept 2025)
Which company records make good environment data
The most useful records show a task, the steps taken and the outcome, with links between systems. A ticket tied to the email thread and the fix that closed it is worth more than any one of them alone. The workflow data guide shows how to describe one workflow end to end.
| Record | Why environment builders want it | Preparation |
|---|---|---|
| Support tickets with resolutions | Each one is a task with a known correct outcome | Medium: customer details in messages |
| CRM pipelines and activity history | CRM work is among the most-cited enterprise workflows | Medium: contact details |
| Spreadsheets, models and reports with their inputs | Report filing and spreadsheet work are growth areas | Low to medium |
| Slack, Teams and email threads | Realistic coordination, interruptions and context | High: staff and customer personal information |
| Code repositories with issues and commits | Bug-fix tasks graded by tests | Medium: secrets and licenses in history |
| Project and issue trackers | Multi-step tasks with dependencies and owners | Medium |
| Work orders and field job records | Scheduling, dispatch and diagnosis tasks | Medium: customer addresses |
What makes records valuable for an environment
- Linkage: consistent IDs across systems, so a customer, ticket and invoice can be followed end to end.
- Recorded outcomes: whether the ticket was solved, the deal closed or the report approved. Graders need ground truth.
- Real mess: exceptions, escalations, rework and disagreements. Clean, synthetic-looking data is what builders already have.
- Years of history across many people, so tasks vary rather than repeat.
- Coherent de-identification: names replaced consistently, not deleted, so threads still make sense.
How companies get paid
Most deals license a de-identified copy of records for a fee, the same way as other AI data licensing. One startup that turns company data into agent environments advertises payouts of up to $300,000 for particularly valuable datasets and has said it paid out $1 million to companies, Founderland reported. Public reports put closed startups' workspace archives at roughly $10,000 to $100,000, and the deals tracker lists larger company-record sales.
Some buyers also pay for people. Environments need tasks and graders written by people who know the work, so a deal can include paid time for current or former staff to write tasks and check that graders reward the right outcome.
What to check before licensing records for an environment
- Scope: say whether the buyer may train on the records, use them to build environments, use them to evaluate models, or all three. Each is a different use; see licensing data for AI evaluation.
- Personal information: de-identify staff and customer details before delivery. See how to de-identify company data.
- Third-party software: you can license your records and configuration, not the vendor's software or documentation.
- Customer contracts: confidentiality clauses can cover records that mention clients.
- Exclusivity: environment builders sometimes ask for it, and it should cost them more. See exclusive vs non-exclusive licenses.
Who buys company data, and who helps you sell it
Buyers, intermediaries and what deals have paid:
- Who buys company data for AI training?
- Brokers, marketplaces and advisors compared
- AI data buyers compared
- Can your company license its operational data?
- What AI buyers pay: 2026 study
- AI data licensing deals tracker
- How much do AI companies pay for data?
- Selling company data to hedge funds
- Where to sell data to hedge funds, compared
- RL environment companies and deals tracker
- SimpleClosure review and alternatives
Next step
Pick the record type with outcomes attached, usually tickets or work orders, and describe it at company level: the system, years covered and volume. No export is needed. The data inventory template gives you the columns.
Origin Data Partners is independent and is not affiliated with any AI data buyer. We help owners compare buyers and find the best fit for their records. See what your data is worth.
Common questions
What is an RL environment in AI?
A simulated setting where an AI agent attempts tasks and an automatic grader scores each attempt, so the model learns which strategies work. For office agents it is usually a replica of a company's software filled with realistic records, plus tasks and checks.
Why do AI labs need company data for RL environments?
Agents trained on tidy synthetic data fail on real work. Real records add the exceptions, history and linked systems that make tasks realistic and graders hard to game, which is why builders license operational data from companies.
How much do companies get paid for data used in RL environments?
There is no public rate card. One environment startup advertises payouts of up to $300,000 for valuable datasets, closed startups have sold workspace archives for roughly $10,000 to $100,000, and one vendor said it committed more than $20 million in 11 days to license company data.
What company data is most useful for RL environments?
Records that pair a task with its outcome and link across systems: support tickets with resolutions, CRM history, spreadsheets and reports with their inputs, issue trackers, work orders and code with its commits.
What are RL environments for LLMs?
For a language model, an environment is usually a container holding a task, the tools the model may use and a grader: a code repository with a failing test, a replica of a CRM with a ticket to resolve, or a set of documents to analyse. Environments for robots are physics simulations, a separate field.
What is an RL gym?
Another name for an RL environment: a simulated setting where an AI agent practises tasks and is scored automatically. Several builders sell "RL gyms" to AI labs.
Which companies build RL environments?
Human-data companies that added environments and a growing group of specialist startups. The RL environment companies tracker lists them with funding, domains and whether they pay for company data.
Is licensing data for an RL environment different from licensing it for training?
Often the same deal covers both, but they are different uses. The license should state whether the buyer may train on the records, build environments from them, evaluate models with them, or all three, and for how long.
Sources and further reading
Origin Data Partners is independent and not affiliated with any buyer named on this page.
- Epoch AI: An FAQ on reinforcement learning environments (Jan 2026)
- TechCrunch: Silicon Valley bets big on environments to train AI agents (Sept 2025)
- LinkedIn: a data vendor CEO on $20M committed in 11 days to license operational data for RL environments (Aug 2026)
- Founderland: a startup that turns company data into agent training environments (Aug 2026)
- MIT Technology Review: AI labs buying data from dead startups (Oct 2026)