Maat: The Agentic Legal Research Assistant for Competition Protection

Problem
This paper addresses the inadequacies of existing AI-driven legal research assistants in the domain of competition law. Current models, including general-purpose assistants like Claude and ChatGPT, as well as specialized tools such as SaulLM-7B and LegalGPT, fail to provide the necessary domain expertise, accurate citations, and reliable case references. The authors highlight that these tools often hallucinate information, which is particularly detrimental in legal contexts. This work is presented as a preprint and has not yet undergone peer review.

Method
The authors introduce Maat, a ReAct (Reasoning and Acting) agent specifically designed for competition law research. Maat employs a multi-faceted architecture that integrates various tools tailored to distinct tasks within the legal research process. Key components include:

• Retrieval-Augmented Generation (RAG): This mechanism enhances the reliability of the information by grounding findings in official legal sources.
• In-line Citations: Maat provides comprehensive citations to support its outputs, addressing a critical gap in existing tools.
• Web Search Integration: When database coverage is insufficient, Maat can fall back on web searches to retrieve relevant information.
• User Interaction: The system prompts users for clarification on ambiguous queries, improving the quality of responses.

The iterative design process involved collaboration with competition law experts to ensure that Maat meets the specific needs of legal practitioners.

Results
Maat demonstrates significant performance improvements over baseline models in case-specific tasks, achieving a notable increase in accuracy and reliability. While exact numerical results are not disclosed, the authors report that Maat outperforms all tested baseline assistants in these tasks. Furthermore, Maat’s performance on theoretical question tasks is competitive with the top baseline, indicating its versatility across different types of legal inquiries. The dataset utilized for training and evaluation is publicly available on GitHub, promoting transparency and reproducibility.

Limitations
The authors acknowledge several limitations, including the potential for Maat to still produce incorrect outputs, particularly in edge cases or when faced with highly ambiguous queries. Additionally, the reliance on web searches may introduce variability in the quality of information retrieved. The paper does not address the computational resources required for training and deploying Maat, which could be a barrier for some users. Furthermore, the generalizability of Maat beyond competition law to other legal domains remains untested.

Why it matters
The development of Maat has significant implications for the field of legal research, particularly in enhancing the efficiency and accuracy of competition law analysis. By providing a specialized tool that addresses the shortcomings of existing models, Maat can facilitate better legal decision-making and improve the quality of legal research outputs. This work sets a precedent for future AI applications in niche legal domains, emphasizing the importance of domain-specific expertise in AI model design. The approach taken in Maat could inspire similar frameworks for other areas of law, potentially transforming legal research practices.

Authors: Basant Mounir, Farida Madkour, Amira Abdelaziz, Asmaa Sami
Source: arXiv:2605.27331
URL: https://arxiv.org/abs/2605.27331v1