RL Environment Companies for Knowledge Work and Enterprise Workflows
Patronus AI TeamNovember 202610 min read
Patronus AI publishes this list and is one of the companies on it. Entries are listed alphabetically, not ranked, and each entry links its sources.
The main companies building RL environments for knowledge work and enterprise workflows in 2026 are Fleet, Halluminate, Mercor, Patronus AI, Plato, Scale AI, Surge AI and Turing. Some simulate whole businesses or jobs, some clone the enterprise apps people work in, and some sell ready-made task sets written by professionals. To choose, match the environment to the job your agent must do, ask who wrote and graded the tasks, and check how long and how many apps a task can span.
Knowledge work agents are judged on the result, such as a memo, a decision or a reconciled account. They learn in RL environments that hold the documents, tools and messages a professional works with, a task that ends in a deliverable and a grader that scores it. Grading is the hard part, since much of the work calls for judgment and the sources often disagree. This guide examines eight companies serving knowledge work buyers in 2026, listed alphabetically. It is written by Patronus AI, a behavioral research lab whose open benchmark FinanceBench has 350+ citations and 82,000+ Hugging Face downloads of its open sample as of November 2026.
The eight sit at different layers. Fleet and Plato simulate whole businesses or jobs, and Halluminate goes deep in finance and consulting. Mercor, Surge AI and Turing are expert data companies that sell knowledge work environments and task sets next to their human data work, and Scale AI builds environments from a real company’s operating history. Patronus AI is a research lab that builds environments with practitioners from the job.
Last updated
Key takeaways
- Start from the job, not the app. Knowledge work spans documents, spreadsheets, email and chat, so a single app rarely holds a whole task. Plato builds worlds around the tasks of a job, and Halluminate’s diligence tasks combine a desktop, office apps, a data room and tools in one trajectory.
- Ask who wrote and graded the tasks. Judgment work needs people who do the job. Halluminate’s problems are written by practicing deal professionals, and Mercor grades with expert-authored rubrics.
- Check how the grader handles judgment. Where a number can be checked, code should check it, and where judgment matters, someone has to write the criteria. Halluminate mixes rubric grading with deterministic checks, and Scale AI pairs tasks with process and outcome verifiers.
- Long-horizon means state that changes. Professional tasks run for many steps, and the documents and records change along the way. Halluminate’s diligence trajectories reach hundreds of steps, and Scale AI’s environments carry a company’s operating history.
- Open sets help you start. Public tasks let a team test its harness before it buys. Mercor’s APEX-Agents tasks are open source, and Plato’s Computer Work Val is a validation set of 300 worlds.
Comparison table
All eight companies at a glance, listed alphabetically. Security shows what each company lists on its own site. Sort by any column marked with arrows.
| Company | Best for | Type | What it sells | How tasks are graded | Security |
|---|---|---|---|---|---|
| Fleet | Simulated businesses with people supervising | Specialist | Training gyms that simulate how businesses run | Expert grading criteria and reviews | Not listed on its site |
| Halluminate | Finance and consulting work | Specialist | Benchmarks and RL environments for financial services and consulting | Deal professionals, rubrics and deterministic checks | Not listed on its site |
| Mercor | Professional services work, open and licensed | Expert network | Ready-made datasets in enterprise app clones, custom projects, the open APEX-Agents set | Expert-authored rubrics with an LM judge | SOC 2 Type II |
| Patronus AITalk to us → | Decisions from sources that disagree, from a research lab | Research lab | Environments built with practitioners, ready-made and custom, and parts on their own | In finance, code where a number can be checked and practitioner criteria where judgment matters | SOC 2, HIPAA, TISAX |
| Plato | Simulations of the software a job runs on | Simulator | Worlds that simulate the tools and systems of a job | Verifiers with objective criteria, reviewed by experts from the occupation | Not listed on its site |
| Scale AI | Enterprise systems and desktop work, from one large vendor | Data and evaluation | Enterprise Atlas environments, Professional Office and desktop tasks | Expert rubrics, a rubric-based agent judge, process and outcome verifiers | SOC 2 Type II, ISO 27001, FedRAMP High |
| Surge AI | Ready-made enterprise and long-horizon tool use tasks | Expert data | Off-the-shelf enterprise and long-horizon tool use environments, custom RL environments | Expert rubrics | Not listed on its site |
| Turing | Clones of enterprise apps and MCP environments | Expert data | UI clones of enterprise apps, MCP environments, off-the-shelf and custom datasets | Rubric-aligned tasks, trajectory traces | SOC 2 Type II |
Fleet
Fleet describes itself as an applied research lab dedicated to scaling out environments, training gyms that model work for AI.
How does Fleet work?
- Its training gyms are simulations of how real businesses run, where agents practice work while people supervise, evaluate and steer their behavior.
- Its legal projects use realistic matters to test how well agents research, reason and back up their conclusions.
- Legal experts write grading criteria tied to the case file and the governing law, and review AI work for errors in reasoning, facts and citations.
Where it fits
Fleet’s public site describes its gyms through its careers pages, and does not say whether it sells ready-made environments or custom builds.
Best for: Teams that want simulated workplaces with human supervision and expert grading.
Sources: fleetai.com, fleetai.com/careers/domain-expert-legal
Halluminate
Halluminate is a data research lab building benchmarks and RL environments for knowledge work, starting with financial services and consulting.
How does Halluminate work?
- Tasks operate a real desktop environment with files, office applications and a data room, while calling tools for email and chat, in one trajectory.
- Every problem is written and reviewed by practicing finance deal professionals and scored by a mix of rubric grading and deterministic checks, audited in an independent QA pass.
- Its Westworld Finance Diligence Bench (Aug 11, 2026) has 88 problems covering a full acquisition due-diligence process, with trajectories that reach hundreds of steps.
Where it fits
Halluminate’s depth is in deal and diligence work. Its site does not say whether it sells ready-made environments or custom builds.
Best for: Agents for finance and consulting work, with tasks written by deal professionals.
Sources: halluminate.ai, halluminate.ai/blog/due-diligence-bench
Mercor
Mercor builds benchmarks and RL environments on real professional work through its APEX program, drawing on its network of experts.
How does Mercor work?
- Expert-built tasks run inside clones of enterprise apps and test whether agents can navigate hundreds of files, hold context and finish long-horizon professional work.
- APEX-Agents measures long-horizon, cross-application tasks for investment banking analysts, management consultants and corporate lawyers. Its 31 worlds and 240 tasks are open source, with the Archipelago service that runs and evaluates trajectories.
- Datasets are pre-built, peer-reviewed and ready to license, and custom projects cover agent environments and new verticals.
Where it fits
Mercor’s accounting index (Jul 31, 2026) sets tasks in a fictional company at month-end close, with its own accounts, software and files.
Best for: Teams training agents on banking, consulting, legal or accounting work who want open tasks to start from.
Sources: mercor.com/apex/off-the-shelf-data, trust.mercor.com
Patronus AI
Patronus AI is a behavioral research lab for AI. Its knowledge work environments recreate a job, with the documents, tools and messages a professional works with and a task that ends in a deliverable, alongside environments in coding, computer use, science, alignment and more.
How does Patronus AI work?
- Built with practitioners from the job, who ground the case, the documents and the checks in how the work is done. Experts who did not build a task review it before it ships.
- Real job functions, from product managers to financial analysts, including corporate finance, banking, audit and accounting.
- Finance tasks run in spreadsheets, terminal-style tools and document apps, and many span several apps and long sequences of steps where the state changes along the way.
- In finance work, code checks any number that can be checked, practitioners write the criteria where judgment matters, and experts test each grader against worked solutions.
Where it fits
Patronus AI publishes on how models and agents behave and where they fail. In MEMTRACK: Evaluating Long-Term Memory and State Tracking in Multi-Platform Dynamic Agent Environments (NeurIPS 2025 Workshop), an agent works through an organization’s timeline across Slack, Linear and Git, and the best model tested reached 60% correctness. FinanceBench, its open-book benchmark of 10,231 questions about public companies, has 350+ citations and 82,000+ Hugging Face downloads of its open sample as of November 2026 and is used by post-training teams as an environment. Ready-made environments and custom builds are both offered, and most work is custom.
Best for: Teams training agents on multi-app professional work who want finance graders tested against worked solutions.
Sources: patronus.ai/domains/knowledge-work, patronus.ai/research/memtrack
Plato
Plato builds worlds for models to gain experience in. Each is a working simulation of the tools, software and systems a job runs on, built around the tasks and workflows people do in it.
How does Plato work?
- Worlds are built around the exact tasks and workflows of a job.
- Computer Work Val v1.0 (Oct 27, 2026) is an environment data stream and validation set of 300 worlds.
- Its site says the environments can be used for both testing and training.
Where it fits
Plato suits teams that want whole-job simulations rather than single apps. Verifiers grade each task against objective criteria, and two experts from the occupation review every task.
Best for: Teams that want simulated jobs with their tools and systems for training and testing.
Sources: plato.so, plato.so/research/computer-work-val
Scale AI
Scale AI provides data, evaluations and RL environments to AI labs, governments and enterprises. Enterprise Atlas, its RL environments product, trains agents on long-horizon work across enterprise systems and desktop environments.
How does Scale AI work?
- Environments built with a real company’s operating history, anonymized and rehydrated, plus high-fidelity reconstructions of enterprises for functions such as finance, accounting, loan operations, compliance and customer support.
- Professional Office tasks are long-horizon computer use tasks grounded in professional artifacts across 11 domains, verified by a rubric-based agent judge.
- Expert-designed long-horizon, cross-functional tasks and rubrics, each paired with process and outcome verifiers (Feb 27, 2026).
Where it fits
Its product page offers ready-made sets with sample requests. Its security page lists a SOC 2 Type II report, ISO/IEC 27001:2022, DoD IL4 provisional authorization and FedRAMP High authorization.
Best for: Teams that want enterprise system environments from a large data vendor, or that need FedRAMP High.
Sources: scale.com/rlenvironments, scale.com/security
Surge AI
Surge AI sells training data, evals and RL environments built by the experts behind its lab work, off the shelf or as custom programs.
How does Surge AI work?
- Enterprise agent tasks cover multi-step work inside the systems companies run on, and its long-horizon tool use environments span office documents, enterprise workflows and live external services.
- Frontier work tasks navigate multi-file environments to complete professional deliverables in tax, accounting, finance, healthcare, data science and engineering, graded against rubrics.
- Its simulated workplace post (Nov 3, 2025) gave models 150 jobs inside a simulated online retailer.
Where it fits
Surge AI suits teams that want a ready-made catalog or a custom program with a large data vendor. Custom programs include reinforcement learning and custom RL environments.
Best for: Teams that want off-the-shelf enterprise and long-horizon tasks from a large data vendor.
Turing
Turing builds RL environments and datasets for AI labs, including environments for enterprise knowledge work across industries, functions and roles.
How does Turing work?
- Interactive replicas of enterprise and consumer apps, plus MCP environments with APIs, tool calls and policies, schemas and seed data built by subject matter experts.
- Rubric-aligned tasks in accounting, contract analysis and market reasoning that evaluate judgment, planning and correctness.
- Harnesses capture trajectory traces automatically (RL environments).
Where it fits
Turing suits teams that want app clones and tool environments for enterprise flows, off the shelf or custom. Its data program site lists SOC 2 Type II, with the report available under NDA.
Best for: Teams training agents on enterprise app flows through the UI and through tools.
Sources: turing.com/frontier-ai/enterprise-knowledge-work, lazarus.turing.com
Where Patronus fits
Patronus AI wrote this guide and is one of its entries. It builds knowledge work environments with practitioners from each job, for functions from product management to corporate finance, banking, audit and accounting, alongside environments in other domains. Ready-made environments and custom builds are both offered, parts can be bought on their own, and it confirms how tasks run on a team’s own training harness when it scopes the work.
Its research on agents at work includes MEMTRACK: Evaluating Long-Term Memory and State Tracking in Multi-Platform Dynamic Agent Environments and FinanceBench, which post-training teams use as an environment. Experts test each finance grader against worked solutions, and it lists SOC 2, HIPAA and TISAX, with data ownership set in each agreement. See how it builds knowledge work environments.
Frequently asked questions
What is an RL environment for knowledge work?
It recreates a job. It holds the documents, tools and messages a professional works with, a task that ends in a deliverable and a grader that scores the result. Our page on RL environments for knowledge work agents covers how Patronus AI builds them.
Which companies build RL environments for knowledge work?
Fleet, Halluminate, Mercor, Patronus AI, Plato, Scale AI, Surge AI and Turing, among others, listed here alphabetically. The table above sets out what each one sells and who it suits.
What does long-horizon mean in RL environments?
It describes tasks that run for many steps, often across several apps, where the state changes along the way. Halluminate’s diligence tasks reach hundreds of steps, Mercor’s APEX-Agents tasks cross applications, and Scale AI and Surge AI sell long-horizon task sets.
How is knowledge work graded?
Usually with rubrics written by professionals, plus code checks where a number can be verified. Halluminate mixes rubric grading with deterministic checks, Mercor uses expert-authored rubrics with an LM judge, and Patronus AI has experts test each finance grader against worked solutions.
How is knowledge work different from computer use?
Knowledge work agents are judged on the result, such as the memo or the reconciled account, while computer use is about operating the software. An agent needs computer use to do knowledge work. Our guide to computer use RL environment companies covers that side.
Where can I compare RL environment companies across domains?
Our guide to RL environment companies covers the wider field, including open hubs and platforms.
How was this list made?
Patronus AI wrote it and is one of the companies on it. The entries are listed alphabetically, not ranked, and each entry links its sources so you can judge for yourself.
Last updated