Open Door Legal Intake Agent

Click the image above to see a (somewhat outdated) demo of the agent

The OpenDoor Legal Intake Agent is a prototype of agentic AI system for facilitating the intake of new clients for OpenDoor Legal. OpenDoor Legal's mission is to ensure universal access to justice, and the intake agent aims to help them scale their effors by reducing the time needed to qualify new clients.

Problem

Though we all have a right to an attorney for criminal cases, we do not for civil cases like those relating to housing issues, or family disputes, or dealing with estates. When people who cannot afford an attorney face issues like eviction they often turn to legal aid organizations to help them stay in their homes. In order to get aid from such groups they have to qualify, which means filling out long, often legalistic forms and submitting proof that they meet the organizaiton's acceptance criteria (usually income qualifications and residency requirements).

Filling out these forms is often burdensome for potential clients. Many don't have english as a first language, and not comfortable with technology, or feel intimidated by the associated jargon. Because of this, they often either don't complete the forms when they'd otherwise qualify, or require individual assistance from legal aid staff members to complete the intake and provide complete and correct information. This means people who should be getting help don't.

Solution

AI agents can make it easier and faster to complete intake. A conversational, rather than form-based approach is easier and more inviting for potential clients. Leading LLMs are highly polyglot, can answer questions if users are confused, and can check users responses to ensure data quality. They can also execute workflows typically done by staff, like reviewing documents and scheduling follow-ups. The could help more people access legal aid services and help legal aid organizations scale.

The OpenDoor Legal AI intake agent is a prototype for a system to do this. The agent was created to see how agents could support ODL in achieving it's mission of universal access to justice.

Technical Architecture

The prototype was built with the following tools:

• LangGraph ReAct agent using gpt-4.1-mini
• Sqlite for conversation memory and user persistence
• Custom tools for Salesforce data fetching and synchronization, answer extraction and persistence, dropbox document uploads
• Streamlit and Twilio for chat interface and sms-based follow-up
• Prefect for task scheduling (follow-up conversations and data quality checks)
• DeepGram for voice based input STT

Features & Functionality

The agent prototyp handles nearly all of this autonomously, and includes additional functionality to facilitate the intake process.

• Asks and saves all of the questions on the current form using a conversational script via the chat interface
• Creates or updates Salesforce records for Contacts and Legal Issues
• Schedules follow-up conversations based on user availability
• Initiates sms-based follow-up conversations to collection additional case information and documents
• Handles images and documents sent via the application or sms and persists them in Dropbox

Challenges and Learnings

So far we (ODL and I) learned a lot about the value and challenges of using agents for this kind of work. In no particular order, here are a few of those things:

• Chat is not necessarily the best interface for this kind of application. Many of ODL's clients are not great typists and it can be burdensome for them to have to explain complicated legal issues via typing.
• Voice could be a good interface, but is not a panacea. Users even high-quality, low latency systems non-natural and weren't always successful using them for input
• It's very important to understand the workflow and systems a given organization is using. It was only after extensive interviews with Front Line Partners that I realized we didn't want to just replicate the existing process, since it was inefficient and driven by the shortcomings of the form / existing tech.
• Simulation is great for end-to-end integration testing and evaluation. We heavily leaveraged LLMs to create simulated personas of potential clients to test the many edge cases and different client types.

Results & Impact

The results of the prototype have mostly been learnings, but ODL as a result of the work ODL received a six-figure grant from Salesforce to productionize the application.