LawFlow: Collecting and Simulating Lawyers' Thought Processes on Business Formation Case Studies

active

LawFlow is a dataset and benchmark of complete end-to-end legal workflows, collected from trained law students working on real-world business entity formation case studies, used to compare human and large language model workflows and to assess how AI can safely support complex legal tasks.

Endorsements support University of Minnesota, Twin Cities.

People

Dongyeop Kang

Principal Investigator (PI)

Funding Details

Start Date: -
End Date: -
Expected Duration: -
Funding Raised to Date: $74,132
Annual Budget: -
Monthly Burn Rate: -
Current Runway: -
Funding Goal: -
Funding Stage: -
Fiscal Sponsor: -

Project Details

LawFlow is a legal AI benchmark developed at the University of Minnesota that captures full, end-to-end legal workflows for business formation matters. The dataset records how trained law students plan, research, communicate with clients, and draft documents over time, and contrasts these workflows with those produced by large language models. By focusing on dynamic, iterative reasoning rather than isolated input-output pairs, LawFlow enables fine-grained evaluation of where current legal AI systems fall short and how they might better augment, rather than replace, human lawyers.

Theory of Change

By collecting rich, end-to-end records of how lawyers-in-training actually reason through business formation matters and comparing these workflows to those produced by large language models, LawFlow exposes systematic gaps in current legal AI systems and identifies roles where AI can safely augment rather than replace human judgment. These benchmarks and analyses can guide the design and evaluation of future legal AI tools, helping researchers and policymakers develop systems that better align with professional legal practice and reduce risks from over-reliance on inadequately tested AI agents in law.

Grants Received– no grants recorded

Discussion

No comments yet. Be the first to share your thoughts.