The state’s small business support platform included a live chat feature designed to help entrepreneurs get real-time answers about permits, licensing, and startup guidance. But with more than 145,000 conversations across five agencies, the team didn’t have a scalable way to analyze what users were asking—or where they were getting stuck.
Without clear visibility into these interactions, it was difficult to identify recurring pain points, respond to user needs, or improve the experience.
To help the State better understand where users were getting stuck, we built a scalable, privacy-conscious way to analyze live chat conversations using AI.
Our goal was to highlight patterns and trends without relying on manual review. By combining proven large language models with a practical implementation approach, we made it easier for teams to act on real user needs, faster.
We built an AI-powered audit tool that uses large language models to analyze chat transcripts and surface recurring questions, confusion points, and user needs. Instead of manually reading thousands of conversations, the team now receives summaries and trend reports that make it easier to identify patterns and take action.
These insights help the State improve:
Large language models were used to turn unstructured chat transcripts into structured, actionable insights. This reduced the need for manual review and accelerated feedback loops between users and service teams. The result is a faster, more consistent way to identify friction points and deliver targeted improvements across digital services.
We leveraged trusted LLMs to move quickly and securely. There was no need to invest in training new models from scratch. We used existing large language models — Claude, OpenAI, and Amazon Q — accessed through secure cloud platforms.
Our team tested multiple options, tuned prompts, and manually validated outputs to ensure they were accurate and relevant. This let us deliver results without training a custom model or introducing new infrastructure.
This implementation demonstrates AI system integration at its most practical, deploying generative AI capabilities within an existing tech stack while meeting state-level requirements for security, governance, and oversight.
We made AI insights easy to integrate and repeat. The audit tool fits into existing workflows and user privacy is a priority every step of the way. Chat transcripts are scrubbed for personal information and then fed through the AI tool using a structured prompt, and reviewed by staff before insights are used to inform platform and service updates.
This repeatable, privacy-conscious model reflects MLOps and ML deployment best practices. Outputs remain reliable, traceable, and usable over time without adding technical complexity.
Applying AI to tens of thousands of user conversations gave the State a clearer picture of where business owners were struggling and a scalable way to respond. The audit tool saves staff time, supports faster decision-making, and helps teams stay focused on real user needs.
Fearless helped prioritize use cases with tangible operational value, built with human oversight, and aligned every step of implementation with long-term service goals.
Our practical approach to AI implementation shows how government agencies can use existing, trusted models to solve real operational challenges, without introducing unnecessary complexity or risk. We prioritized transparency, data privacy, and human oversight at every step, ensuring that insights could be trusted and acted upon.
By embedding AI into an existing workflow, rather than replacing people or rebuilding systems, we delivered a responsible, sustainable solution that helps the state better serve its small business community.
This work is proof that AI doesn’t have to be disruptive to be powerful. When applied thoughtfully, it becomes another tool to help the government deliver faster, smarter, and more human-centered services.