AWS re:Invent 2025 - How Mary Technology is building the legal Fact Layer for agentic AI on AWS

Summary of AWS re:Invent 2025 Presentation: "How Mary Technology is building the legal Fact Layer for agentic AI on AWS"

Overview

Mary Technology, a legal tech firm based in Sydney, presented how they are addressing the challenges of using large language models (LLMs) for legal document review and dispute resolution workflows. The CEO, Dan, highlighted four key problems with using LLMs in this domain and explained Mary's approach to building a "fact layer" to power more reliable and explainable AI-assisted legal work.

Limitations of LLMs for Legal Document Review

1. Training Data Availability: Legal documents often contain sensitive information that cannot be publicly shared, limiting the training data available for LLMs.
2. Multiple Correct Answers: In legal disputes, there are often multiple valid narratives and perspectives, which LLMs struggle to capture effectively.
3. Information Loss from Compression: LLMs compress documents into embeddings, tokens, and summaries, losing important nuances and details critical for legal analysis.
4. Extracting Meaningful Facts: LLMs have difficulty extracting and contextualizing key facts from legal documents, such as identifying specific individuals, dates, and their relevance to the case.
5. Lack of Confidence and Explainability: LLM-generated outputs may be fluent and plausible, but lawyers need to be able to verify and justify the facts and conclusions presented.

Mary Technology's Approach

1. Fact-Centric Processing: Mary treats facts as "first-class citizens," building a specialized pipeline to extract, structure, and enrich facts from legal documents.
2. Fact Modeling and Metadata: Each fact is represented as an object with detailed metadata, including provenance, relationships, and explanations for how it was derived.
3. Verification and Narrative Building: Mary provides a comprehensive review and verification experience, allowing legal teams to assess the facts, build narratives, and collaborate on the case.
4. Integrating with Downstream AI: The structured fact layer can be seamlessly integrated with other AI systems, such as OpenAI, to power more reliable and explainable legal workflows.

Technical Details and Results

• Mary's fact extraction and processing pipeline leverages advanced natural language processing and machine learning techniques to handle unstructured legal documents.
• The platform can extract and contextualize a wide range of entities, events, dates, and other key information from legal documents.
• Mary has achieved a 75-85% reduction in time spent on document review for its customers, which is often a major bottleneck in litigation.
• The platform has received a 96/100 Net Promoter Score, indicating a high level of customer satisfaction.

Business Impact and Applications

• Mary's solution addresses a critical pain point for law firms and legal departments, enabling them to streamline document review and fact-finding processes.
• By providing a reliable and explainable fact layer, Mary empowers legal teams to make more informed decisions, build stronger cases, and deliver better outcomes for their clients.
• The platform's integration capabilities allow it to seamlessly fit into existing legal workflows and leverage other AI technologies, further enhancing its value proposition.

Customer Testimonial

"Mary has been a game-changer for our firm. It's reduced the time we spend on document review by over 75%, allowing us to focus on the strategic and analytical work that really matters. The fact layer is a revelation - it gives us confidence in the information we're working with and helps us build airtight cases for our clients."