Practice voter outreach and canvassing with WeRelate.
For the 2024 presidential campaign, the Technical Projects Group team at Schmidt Entities collaborated with Indivisible, a progressive grassroots organization, to design an LLM chatbot called WeRelate. It was created as a resource for first-time canvassers, enabling them to practice conversations with voters and build confidence in their outreach efforts.
AI-Powered Volunteer Training Tool
The Technical Projects Group at Schmidt Entities embarked on an initiative to explore the potential of large language models in informing voters through campaign talking points and political messaging. This exploration included evaluating the strengths, weaknesses, and risks of using an LLM trained on political candidates’ talking points and policies. During this process, we identified a significant gap in volunteer and mobilizer training, prompting the creation of a chatbot designed to provide dynamic, scenario-based conversations. This tool aimed to better prepare individuals for canvassing and direct voter engagement. I also researched the landscape of AI tools used in messaging and voter outreach as part of the initial research phase.
The problem
In grassroots campaigns, rapid volunteer turnover creates friction for campaign managers trying to maintain momentum, while helping onboard and train first-time volunteers. How might technology enable faster, more consistent training at scale?
User flow demonstrating the Conversation Checklist, Chat, and Ask for Help features
User research and insights
To design a solution that met the needs of volunteers, I led comprehensive research sessions in collaboration with a team from Trestle. These sessions involved interviews with field organizers and experienced volunteers to understand their processes for voter outreach, including:
Following voter outreach protocol.
Preparing to speak to voters and have meaningful conversations.
Learning scripts and effectively conveying talking points.
Our research revealed key pain points in the training and preparation process. We chose to focus on first-time and beginner volunteers, who often felt the most unsure about how to engage with voters. Our goal was to help them feel more confident and better equipped to have meaningful, authentic conversations with potential voters.
To deepen my understanding, I participated in canvassing and door-knocking in Corona, Queens, as a first-time volunteer with NYC Votes. This experience provided first-hand insight into the process of speaking to voters including door-knocking, following Get Out the Vote protocols, and using the Minivan mobile app.
Iterative Design and Development
Persona and Prompt Design
The engineering and product teams collaborated to experiment with prompts that configured chatbot personas. This required close teamwork to refine the LLM’s output and ensure alignment with specific voter demographics and personas.
To validate the chatbot’s utility, I led additional research sessions where participants provided input to the chat bot and reviewed its output to share reactions and recommendations. These insights helped refine the chatbot’s design and ensured it met the needs of volunteers.
Scenario Development
Using feedback from testing sessions, the team iteratively refined the chatbot prompts. The final product supported training volunteers in four voter outreach scenarios:
Get Out the Vote
Persuasion
Volunteer Recruitment
Event Recruitment
I also spearheaded the creation and documentation of all user-facing content. This documentation was reviewed weekly, ensuring alignment between the product direction and code development.
User Interface Design and User Acceptance Testing
I was responsible for all aspects of user interaction and interface design for mobile and desktop platforms. Design decisions were informed by team feedback, brainstorming sessions, and usability testing. This iterative design process led to the creation of high-fidelity wireframes delivered to the front-end engineers.
Testing and Launch
Before launch, the team conducted rigorous testing across various scenarios to identify and address bad responses and bugs. Key steps included:
Creating detailed testing scripts to guide the team’s assessments.
Stress-testing the chatbot and helper feature during the User Acceptance Testing phase.
Recording issues and translating them into engineering tickets for resolution.
After addressing all critical bugs and completing a penetration test, the chatbot was successfully launched in partnership with Indivisible.
Outcomes and Reflections
This project demonstrated the potential of leveraging AI to empower volunteers and enhance voter outreach efforts. Through collaborative design, rigorous testing, and user-centered research, we delivered a tool that effectively supports canvassing training and builds confidence in first-time volunteers. This experience also reinforced the value of iterative feedback loops in refining digital products for impactful results.
While our product did not reach enough live users due to a partnership falling through late in the process, we prepared our infrastructure with measurement and feedback loops in mind. We set up analytics tracking, allowing us to monitor key user behaviors such as onboarding completion, conversation script engagement, and resource usage. To measure satisfaction and surface qualitative feedback, we also designed and integrated an NPS survey to capture how confident volunteers felt after using the tool. These metrics were chosen to align with our goals: increasing volunteer confidence and readiness for voter outreach.
While we weren’t able to fully validate the product in the field, the project sharpened my thinking around prompting and testing LLMs, information delivery and building confidence using a trainijng tool. It also highlighted the importance of early alignment with external partners and how product strategy often requires contingency planning for things outside the team’s control.