Witnesses tell House committee internal-facing AI chatbots and data structuring can speed constituent service

An unidentified committee member pressed witnesses on how the House could build an AI-based chatbot that provides accurate, office-specific responses for constituents.

Unidentified Witness (Speaker 2) described New Jersey’s approach: begin with internal-facing chatbots, train models on a restricted corpus of a member’s position statements and drafts, and give staff a tool to synthesize information rather than replace staff. “We loaded it with all of that member's position statements, build drafts, etcetera,” Speaker 2 said, and added that training tools to restrict answers to a specific corpus helps avoid hallucinations.

Speaker 2 pointed to a PopVox and AI for Impact program pilot that built and tested a member chatbot and said the tools allowed New Jersey to reduce call-center response time “from 40 minutes to now 3 minutes” while modernizing call-center operations. The witness emphasized those were staff-facing tools rather than automated constituent-facing bots.

Unidentified Speaker 3 echoed the recommendation, saying offices should consider multiple, separate chatbots for different datasets — for example, one connected to HouseNet for operational information, another to an office’s private data, and a third to a content-management system — and that the critical task is structuring data to enable those services.

Committee members asked whether accuracy and constituent verification could be ensured. Witnesses repeated that limiting a model’s training corpus and keeping chatbots as staff-facing lookup tools are two practical safeguards against inaccurate or off-brand responses.

No formal actions or votes were recorded on the topic during the session.