AI Companies Shift From Text Training to Virtual Workplaces

A new training paradigm for AI agents

The frontier of artificial intelligence development is moving beyond chatbots. Major technology companies are now investing heavily in reinforcement learning environments—realistic digital simulations of workplaces where AI systems can practice performing entire jobs rather than simply generating text responses.

Two recent developments illustrate this industry-wide pivot. Google is negotiating to invest over $1.5 billion in Mechanize, a startup specializing in virtual work environments for AI training, according to exclusive reporting by Business Insider. Separately, Meta has begun collecting employee keystrokes, mouse movements, and screen activity to help AI systems understand how people actually use computers and workplace software.

Why it matters

This shift addresses a fundamental limitation of current AI systems: while large language models excel at conversation, they struggle with complex, multi-step tasks that unfold over hours or days. Reinforcement learning environments could unlock AI's ability to handle complete workflows autonomously—a capability that would transform how businesses deploy artificial intelligence across operations from software engineering to finance and legal work.

From static datasets to simulated work

The first generation of large language models relied on two primary training methods: massive text datasets scraped from the internet and human contractors rating chatbot responses. That approach produced capable conversational AI but hasn't delivered agents that can independently execute extended, complex work.

Scale AI, a leading training data supplier, reports that nearly half of its new projects now involve reinforcement learning environments. These simulate realistic coding environments, computer use scenarios, and enterprise workflows where AI agents learn through trial and error rather than pattern matching on static text.

"The way models are trained is evolving," wrote Chetan Rane, head of product for agents and RL environments at Scale AI. "Frontier models increasingly need to learn through trial and error in realistic simulated environments."

Building digital replicas of real jobs

Mechanize, founded by AI researcher Tamay Besiroglu, has set an ambitious target: creating simulated environments that "capture the full scope of what people do at their jobs." The startup initially focuses on software engineering, where future coding agents would learn from professional programmers before improving through reinforcement learning in increasingly complex simulated projects.

The company's approach centers on reward signals—feedback mechanisms that tell an AI whether it completed a task correctly, then reinforce successful behavior patterns. Besiroglu explained this allows models to understand task completion rather than simply generating plausible-sounding responses.

Meta's employee activity tracking serves a similar purpose: capturing real-world examples of how people navigate workplace software, use keyboard shortcuts, and complete everyday computer tasks. This data would train AI agents on actual human workflows rather than idealized scenarios.

Embedding AI development in business operations

Uber has launched "Agentic Pods"—teams of AI engineers embedded in departments including finance, legal, HR, and marketing to observe employee workflows before redesigning them around AI capabilities.

These initiatives signal that the AI industry's competitive focus has evolved. Rather than racing to build smarter chatbots, companies are competing to construct the most detailed digital replicas of real workplaces where AI agents can practice the thousands of decisions and actions that constitute modern knowledge work.

Business Insider first reported these developments.