AI for Justice: How Lone Star Legal Aid built Juris to deliver faster, fairer results
- Legal aid is ahead on AI: 74% of legal aid organizations use AI today, nearly double the 37% adoption rate across the broader legal field.
- Juris centralizes trusted knowledge: Lone Star Legal Aid's AI tool returns fast, source-cited answers using retrieval-augmented generation and semantic chunking.
- Built to be practical and affordable: A phased, two-year build-and-test approach kept infrastructure near $2,000/year (with nonprofit credits) and required about 300 staff hours.
Millions of people who qualify for civil legal help never get it. Legal aid teams feel that pressure every day - thin budgets, rising demand, and knowledge spread across too many systems. That urgency is why legal aid is moving faster on AI than the rest of the profession.
Lone Star Legal Aid (LSLA), serving eastern Texas, ran into the same friction many of you face: research time lost to pricey platforms, scattered PDFs, and "tribal knowledge" locked in individual files. So they built Juris - an AI-enabled knowledge hub that returns source-cited answers and gives staff a single place to search what matters.
Why legal aid is leading
New research shows 74% of legal aid organizations already use AI in their work. The broader legal market sits around 37%. Scarcity forces focus: if a tool reduces duplicated effort and shortens time to a reliable answer, it gets adopted.
The problem Juris solves
Before Juris, LSLA's materials lived in research databases, internal drives, and static repositories. Even vetted documents weren't centrally accessible. The result: slower research and inconsistent quality that strained both attorneys and clients.
Self-help portals and court help centers face the same issue - fragmented sources, expensive licenses, and uneven access to authoritative guidance. A verifiable, consolidated hub is the fix.
How Juris delivers trustworthy answers
Retrieval-augmented generation (RAG): Juris grounds answers in the text it retrieves, not in model guesswork. It cites the sources used, so attorneys can verify and move with confidence.
Semantic chunking: Documents are split into meaning-based sections (e.g., a heading and all the paragraphs under it). When a user asks a question, Juris pulls only the most relevant sections, keeping context intact and reducing hallucinations.
Technical architecture that holds up in practice
Juris pairs Azure OpenAI for secure, stateless model access with a custom web interface that places a PDF viewer right next to the chat. Users can jump from an answer to the exact citation without switching screens.
The system runs on Azure App Service with continuous deployment via GitHub. That means consistent performance and quick updates without heavy ops work.
Build in phases, test hard, refine
- Concept: Define the core problem to solve and the must-have outcomes.
- Evaluate platforms: Compare open-source and commercial options against security, cost, and maintainability.
- Prototype: Prove the concept quickly; measure accuracy and usefulness with real questions.
- Adversarial testing: Push the system to fail. Detect hallucinations. Validate citation reliability.
- Iterate: Shift from size-based to semantic chunking, improve the UI, expand sources.
- Pilot: Prepare for a controlled rollout to subject matter experts and finalize refinements.
Cost, staffing, and sustainability
Thanks to Microsoft nonprofit credits, infrastructure runs near $2,000 per year. Development took roughly 300 staff hours (about 0.5 FTE, plus 0.3 FTE spread across two years). Ongoing maintenance is minimal.
Phase two, funded by a Legal Services Corporation technology initiative grant, is aimed at faster, more consistent research; lighter lifts for frontline and admin teams; and a modular framework others can adapt.
A repeatable playbook for legal service organizations
- Start small: Pick one high-value use case and one clean corpus of documents.
- Bring users in early: Attorneys should test prompts, check citations, and flag gaps.
- Protect build time: Carve out dedicated hours so staff can ship, not just "squeeze it in."
- Ground every answer: Use RAG with citation viewing in the same interface.
- Measure what matters: Accuracy, time to answer, and rate of verified citations.
- Plan for reuse: Build a modular core you can extend to operations and client-facing tools.
What's next
LSLA will continue rolling out Juris in phases and build sister tools on the same foundation. They're also sharing lessons through AI peer learning efforts so other organizations can replicate the model and improve it together.
The real scale comes from collaboration - shared playbooks, pooled datasets, and co-designed tools that raise quality while lowering cost. If you're exploring funding or partnerships, start with the Legal Services Corporation's resources at lsc.gov.
Practical next steps for your legal team
- Audit your top five repeat research questions and the sources you trust most.
- Stand up a small RAG prototype with 50-100 high-value documents and a citation viewer.
- Run a two-week adversarial test with your SMEs; ship fixes weekly.
- Track time saved and citation verification rates before broader rollout.
Curious about structured upskilling for your attorneys and staff? Explore curated AI learning paths by job at Complete AI Training.
For technical context on model hosting and security, review Microsoft's Azure OpenAI Service overview at azure.microsoft.com.
Your membership also unlocks:
