Demand Letter assistant

The Legal Aid Society of San Bernardino (LASSB) partnered with Stanford students in the 2024–25 AI for Legal Help policy lab class to develop an AI-powered tool that supports staff in drafting reasonable accommodation demand letters for tenants with disabilities.

These letters, required under federal and state fair housing laws, request accommodations such as allowing emotional support animals or physical modifications to a rental unit. The letters are often urgent and emotionally sensitive, request changes such as allowing emotional support animals, deadline extensions, or physical modifications. The task is repetitive and legally structured, yet time-intensive, and attorney time is often a bottleneck.

The AI agent conducts an interview with the client post-intake, extracts relevant facts, and generates a draft demand letter for attorney review. The system aims to accelerate drafting, improve consistency, and allow attorneys to focus on review and final tailoring.

Why the Demand Letter task?

LASSB identified this task as a high-impact opportunity: the process is frequent, repetitive, and highly structured, yet time-consuming for attorneys and staff. Demand letters are often drafted under time pressure, and intake workflows vary widely. LASSB sought a solution that would maintain attorney oversight while accelerating the drafting process, improving consistency, and allowing staff to serve more clients.

Prototype Process

Over two academic quarters, the student team scoped and iterated on a prototype AI agent that interviews clients after intake, gathers relevant details about their disability and requested accommodation, and generates a draft letter for attorney review. The team worked closely with LASSB attorneys to ensure trauma-informed language, legal accuracy, and appropriate tone. Multiple iterations of sample letters and interaction flows were tested, with attention to AI behavior, hallucination risks, and usability.

Prototype Details

The final proposed workflow positions the AI agent between intake and attorney review. After clients are screened and deemed eligible, the AI agent conducts an interview, produces a summary, and generates a draft letter.

• Trauma-informed Interviewing: The AI gathers client information through a structured, respectful dialogue that avoids retraumatization.
• Draft Letter Generation: The tool produces a professional, legally accurate draft based on the client’s disability, the requested accommodation, and supporting details.
• Attorney Review Workflow: All outputs are routed to a human attorney for revision and approval before sending.
• Safety & Quality Controls: AI hallucination risk, tone, and accuracy were evaluated through a custom rubric developed with LASSB attorneys.

Attorneys maintain full control over the final version, with AI outputs designed to save time, not replace legal judgment. The team also proposed options for summarizing client inputs and integrating attorney review interfaces.

Evaluation Plans

To evaluate the AI tool, the team developed a set of benchmarks and rubrics focused on legal accuracy, tone, trauma-informed language, and completeness.

Feedback was gathered from LASSB attorneys, staff, and subject matter experts, who emphasized the importance of professional tone, factual correctness, and respectful client communication.

Testing revealed areas of strength and persistent challenges, including occasional repetition and lack of clarity.

Next Steps

Next steps include refining the AI model’s prompts, training with additional documents and examples, improving usability for attorneys and clients, and integrating multilingual support.

Longer-term goals involve connecting the agent with LASSB’s LegalServer system and developing broader evaluation methods to assess client outcomes and attorney satisfaction.