This paper introduces LawFlow, a complete end-to-end legal workflow dataset collected from trained law students based on a real-world corporate incorporation scenario, to support the complex and critical tasks faced by legal professionals, particularly those early in their careers. Unlike existing datasets that focus on input-output pairs or linear thought processes, LawFlow captures dynamic, modular, and iterative reasoning processes that reflect the ambiguity, modification, and client adaptation strategies of legal practice. Using LawFlow, we compare and analyze human and LLM-generated workflows, revealing systematic differences in structure, reasoning flexibility, and planned execution. Human workflows are modular and adaptive, while LLM workflows are sequential, exhaustive, and less sensitive to downstream influences. Furthermore, we suggest that legal professionals tend to prefer AI to perform supportive roles, such as brainstorming, identifying blind spots, and suggesting alternatives, rather than executing complex workflows end-to-end. Consequently, we highlight the current limitations of LLM in supporting complex legal workflows and the opportunities for developing more collaborative and reasoning-aware legal AI systems.
