Flight simulators for office work

Pilots don’t learn to land in a storm by reading about it. They practice in a simulator: a realistic cockpit, a scenario that goes wrong in a specific way, and an instructor who knows what should have happened. AI labs now train their agents the same way, and that is why they want your company’s records.

The internet was never the whole story

Large language models were trained on public text, and that supply is finite. Epoch AI’s best estimate is that the stock of public human-written text will be fully used somewhere between 2026 and 2032. OpenAI’s co-founder Ilya Sutskever put it more bluntly in December 2024: “we have but one internet.”

Company records matter for another reason: they hold a kind of text the web lacks. The internet shows the finished product: the published article, the polished proposal. It almost never shows how work actually happens: the support ticket from first complaint to final fix, the Slack thread where a decision gets argued out, the change order that reshapes a project. That part of work was never public, so no model has learned it from the web.

Agents need somewhere to practice

The major labs now sell agents that carry out multi-step tasks inside business software. OpenAI’s Frontier platform, launched in February 2026, connects agents to CRM tools, ticketing systems, data warehouses and internal apps. Answering a question is easy to test. Working a ticket through five systems and getting it right takes practice, and an agent has to learn it by doing.

So labs train agents in reinforcement-learning environments: simulated workplaces where a model attempts a task, is scored on the result and tries again. One founder compared building one to “creating a very boring video game.” The research firm SemiAnalysis counts more than thirty-five companies building them, and notes that labs reuse an environment for future models once it is built.

Why real records make better simulators

A flight simulator is only as good as its scenarios. An environment built from a real company’s records has a realistic cast, real problems and a record of what actually happened, which serves as the answer key. Synthetic workplaces invented from scratch tend to give themselves away; MIT Technology Review notes that real archives avoid errors such as references to dates that don’t exist.

The buyers are specific about what they want. Surge AI wants archives covering “a range of situations, from disagreements to judgment calls,” so that agents can practice work like debugging IT systems or drafting legal briefs. Mercor bought Deeptune, a builder of simulated workplaces, in July 2026. Handshake bought Cleanlab in January 2026 to build evaluations and environments. micro1 says its buying commitment is meant to widen the industries and workflows its environments cover.

What makes records valuable

• Variety over volume. Buyers price an archive on how many distinct, usable tasks it yields, not its size in gigabytes. A varied Slack can be worth more than a much larger archive of routine messages.
• The full loop. Records that connect, such as a ticket linked to the code change that fixed it, the email that prompted it and the CRM record of the customer, let a simulator follow work end to end.
• Breadth of systems. Handshake tells applicants that its highest offers typically include data from eight or more named applications.
• Real judgment. Disagreements, escalations and decisions are what an agent most needs to learn, and what synthetic data struggles to invent.

What happens to your records

Buyers say they remove personal information before records are used, and the programs in our buyer index describe how. The records then become the scenarios inside an environment, or the tasks in an evaluation, rather than text a model simply memorizes. Because environments are reused, your records may help train more than one model. Ask each buyer which uses its license covers.

What this means for an owner

Our view: the records your company produced by doing its work are material AI labs can’t get from the public web, which is why buyers are paying for them. The companies best placed are those with years of history spread across several connected systems, and teams whose work happened in documents, tickets and conversations.