From Manuals to Worlds: Synthesizing Executable Enterprise Environments for LLM Agents

Enterprise applications are often closed source, leaving their internal business logic and workings inaccessible. This creates a challenge for LLM-based agents which must learn how to interact with them to perform application-specific tasks. Learning directly on live enterprise applications however requires repeated interactions and trial-and-error that are generally restricted - while knowledge learned from generic or open-source environments often fails to transfer to application-specific dynamics. Fortunately, substantial information describing enterprise dynamics already exists, scattered across heterogeneous application artifacts such as user manuals, API documentation, and workflow specifications. The real challenge, then, is reconstructing this fragmented information into coherent, structured, and executable environment knowledge. To address this, we introduce Enterprise Emulation System(EES), a framework that reconstructs enterprise dynamics from heterogeneous artifacts into a simulated replica of the actual environment. EES leverages the available artifacts, transforming them into a System Functional Graph (SFG) and an Executable Environment (EE). We evaluate this functionality in a controlled setting on publicly documented EnterpriseArena applications, where EES has no access to the underlying implementations. Our results demonstrate that for Qwen3-8B, providing SFG context yields a 10% relative gain, while training on the EES executable environment followed by SFG-guided inference improves performance by 52% over the base model.