Ledgers, purchase orders, support queues and email from real businesses, rebuilt as environments for training and evaluating AI agents.
Illustrative. A freight company’s systems, reconstructed as an environment. Each point is a record; the bright path is an agent at work.
Synthetic data is clean. Real companies are not. Agents trained here meet the exceptions and edge cases they will face in production.
Real records keep the disputes, duplicates, late approvals and half-finished workflows that synthetic data smooths away.
Every detour is something an agent has to learn.
A disputed line item takes the invoice through a vendor query, a credit note and a second match.
The ledger holds the payable. The full story lives across systems.
Each environment reproduces a process that actually ran, with the systems, roles and rules the company used.
Operational data that never reached the public web, sourced across industries and systems, with new companies added as we acquire them.
Where the data comes from
We acquire real operational data from distressed, insolvent and financially constrained businesses, then turn it into structured, simulation-ready environments.
We buy the operational systems of businesses in administration, liquidation or restructuring, through the insolvency practitioners and buyers who handle their assets.
Seven years of decisions, recorded across six business systems.
Personal data is replaced with consistent tokens inside a secure enclave before anything else happens. Structure, timing and relationships stay intact.
The people change. The relationships that explain their decisions stay.
We rebuild the company as typed state: the entities, relations, permissions and workflows that moved work between people and systems.
Records become entities. Entities reconnect into the workflows that moved the company.
The result runs as an environment with tasks, tools, graders and deterministic resets, ready for training or held-out evaluation.
The company may have closed. Its hardest decisions are ready to run again.
Seven years of decisions, recorded across six business systems.
The things research, legal and security teams ask us first.
From businesses that are distressed, insolvent or winding down. We acquire their operational systems as part of a lawful sale of company assets, typically through administrators, liquidators or restructuring advisers.
No. Names, contact details, bank and tax identifiers and other personal data are replaced with consistent tokens before data is used to build anything. We keep the structure and timing that make the data useful, not the identities in it.
A reconstructed company state, the tools an employee would have used, tasks drawn from what actually happened, and graders that score outcomes. Where the source allows, environments also include the historical human trajectories as reference demonstrations.
Synthetic environments are generated from assumptions about how work happens. Ours come from records of how it did happen, including the exceptions, delays and workarounds that are hard to invent.
The source systems were private and never published, which makes environments built from them a strong basis for contamination-resistant evaluation.
Yes. Tell us which workflows and systems your agents need to handle. We match them to data we hold, or prioritize acquisitions that fill the gap.
Request access and tell us what you are training. We follow up to scope your needs and share sample environments under NDA.
Tell us what you are training. We’ll share sample environments under NDA.
Get Access