Case Study: Designing an AI Assistant for Document-Driven Legal Workflows

Overview

This case study explores the evolution of an AI-powered assistant designed to help legal professionals navigate complex, document-heavy workflows. Originating from a concept unveiled at LegalWeek, the project evolved into a full-featured intelligent agent that enables users to query structured data grids, summarize clauses, and interact naturally with legal content through a conversational interface.

I led the design strategy, end-to-end UX, and cross-functional alignment to ensure the assistant was powerful, trustworthy, and intuitive for non-technical users.

My Role

• Principal Product Designer
• Led concept-to-execution UX across chat, grid, and assistant integration
• Conducted early research at LegalWeek and synthesized 40+ user stories
• Collaborated with engineers, AI researchers, and legal SMEs
• Delivered a prototype and design patterns for future expansion

Problem

Legal teams often review hundreds or thousands of documents per matter. This task is high-stakes, time-consuming, and difficult to scale with traditional tools. Teams lacked:

• Context-aware tools to interact with extractions and summaries
• AI that could clarify ambiguous queries and cite sources
• Interfaces that balanced powerful functionality with clarity and speed

Discovery at LegalWeek

At LegalWeek, we tested an early concept with legal professionals: "What if you could ask a question and get a summary across multiple contracts, complete with citations?"

Lawyers responded with excitement but raised concerns about:

• Trust and hallucination risk
• Needing to control which documents the AI considered
• Wanting summaries plus access to raw excerpts

These conversations laid the groundwork for the following design principles and confirmed the real demand for intelligent document querying.

Design Goals

• Make legal querying as simple as asking a question — without sacrificing control
• Support real-time summarization across structured and unstructured content
• Handle ambiguity responsibly and clearly
• Ground answers in documents with confidence cues and citations
• Empower experts and casual users alike through modular interaction design

Key Design Decisions

From Chatbot to Intelligent Agent

Inspired by LegalWeek feedback, we rejected a simple chatbot model. Instead, we designed for:

• Retained memory across sessions
• Structured task execution
• Conversational precision with legal nuance

Clarification Loop for Ambiguous Queries

To mirror how lawyers work:

• The assistant detects vague prompts
• Responds with clarifying questions
• Only delivers results once the query is clear

Interacting with the Legal Grid

• Users click cells to get summaries + full-text excerpts
• Filters limit AI scope
• Prompts can summarize across rows or the entire dataset

Trust, Citations, and Feedback Loops

• All answers include inline citations
• Confidence scores and "Why this answer?" links build transparency
• Feedback prompts help fine-tune future iterations

Prototyping & Testing

I prototyped several UX flows:

• Ask → Clarify → Summary with sources
• Cell click → Nested excerpt + summarization
• Toggle-based prompt builder for repeat tasks

Test sessions with legal teams helped us refine the following:

• UX writing for clarifications
• The visual weight of trust indicators
• Layout of summary vs source content

Results & Next Steps

• Validated core workflows with 12+ legal teams
• Designed a reusable interaction pattern for LLM + data grid systems
• Developed prompt templates and scoped long-term features like:
  • Saved prompt libraries
  • Persona-based tone settings
  • Project-based chat memory

Conclusion

This project began as a bold idea shared at LegalWeek and evolved into a robust UX framework for integrating AI into legal workflows. It demonstrates my ability to translate early signals into validated features, lead system-level design thinking, and shape product strategy across AI, enterprise UX, and high-trust domains.