Mercor acquires DeepTune to build AI training worlds
CALIFORNIA, UNITED STATES — Mercor has agreed to acquire DeepTune, which has recreated hundreds of enterprise applications as practice environments for AI agent training through reinforcement learning.
According to a company press release, the deal pairs Mercor‘s five-million-expert network with DeepTune’s enterprise software infrastructure.
DeepTune’s founding team moves to Mercor’s New York office; the company raised $43 million in a Series A led by Andreessen Horowitz earlier in 2026. Terms of the acquisition were not disclosed.
DeepTune’s enterprise environments fill Mercor’s training stack
Effective AI agent training requires three components: software environments where work occurs, task definitions specifying what agents must accomplish, and performance verifiers that measure whether objectives were met.
DeepTune has spent two years recreating hundreds of enterprise applications — from spreadsheets to CRM platforms — with the operational fidelity needed to train agents on real-world workflows.
Mercor’s expert network supplies the task and verification components; DeepTune provides the environment layer — the missing piece for a complete AI training stack.
In the July 9 announcement, Foody wrote that the constraint in AI agent development has ‘shifted to the environments themselves’ — that ‘building environments rigorous enough for that kind of training is incredibly hard.’
Reinforcement learning demands drive Mercor’s environment build-out
Mercor was already a DeepTune customer before the acquisition — a commercial relationship that gave Foody direct insight into the productivity gains from pairing expert evaluation with purpose-built environments.
DeepTune’s environments mimic live enterprise software, allowing models to practice tasks — updating CRM records, building spreadsheets, processing reports — under conditions that reflect actual workplace complexity.
Foody frames the deal as a structural response: reinforcement learning can now teach any well-defined task, making environment quality the primary bottleneck for AI advancement.
Writing in the announcement, Foody argued that reinforcement learning can now teach models ‘almost any task that can be clearly defined and scored’ — making the training environment, not the model, the central challenge for the next wave of AI improvement.
The deal reflects the emerging market for AI training infrastructure — specialist providers competing to build environments that determine which models improve fastest on enterprise tasks.
For BPO, AI agents trained on enterprise software signal a shift from rules-based automation toward adaptive, task-learning systems capable of replacing human processing at scale. Mercor’s position — between frontier AI labs and enterprise workflows — gives the combined platform significant leverage in the next wave of AI-driven workplace automation.
