How Grand Traverse County put an AI-powered public service assistant on its website — anonymous, grounded in verified county documents, and compliant with FedRAMP GCC — in time for residents to use it the same day it launched.
County government runs on business hours. Residents don't. Someone trying to find out when the next Board of Commissioners meeting is, what the FY2026 budget covers, or how to contact the Health Department shouldn't have to wait until Monday morning to find out. And county staff shouldn't spend their day fielding calls that a well-designed AI agent can handle instantly.
The challenge for a county government isn't just deploying an AI chatbot — it's deploying one that only answers from verified, approved public documents. A county AI that fabricates a policy, misquotes a budget figure, or invents a service that doesn't exist creates real liability. The constraint was the same as it always is in government AI: grounded answers only, every response tied to a real source.
CORA Connect was built to solve exactly this problem — a 311-style public service agent that answers resident questions accurately, anonymously, and within the compliance boundary that government demands.
"A county AI that fabricates a policy or invents a service creates real liability. Grounded answers only — that's the constraint that makes public-sector AI trustworthy."
CORA Connect Design Principle — Grand Traverse County ITCORA Connect is built on a Retrieval-Augmented Generation pipeline running entirely within Microsoft's Government Community Cloud. Every resident question goes through the same sequence: retrieve relevant chunks from a verified document index, then generate an answer constrained strictly to what was retrieved. If the knowledge base doesn't contain the answer, CORA Connect says so and routes the resident to the right department.
The data flow for every resident interaction follows the same path:
CORA Connect's knowledge base contains exclusively public county documents — nothing internal, nothing sensitive, nothing that hasn't been reviewed and approved for public distribution. Content is organized across five logical categories in Azure Blob Storage, each with a designated content owner and upload cadence.
Every document goes through a pre-upload review for PII, HIPAA, and law enforcement sensitivity before entering the knowledge base. CORA Connect is strictly isolated from CORA Insight — internal documents are never accessible through the public agent.
One of the realities of building government AI inside Microsoft's GCC environment is that some features available in commercial Copilot Studio simply aren't available in GCC. CORA Connect's roadmap reflects that honestly.
The most notable gap: thumbs up/down feedback is not a native feature in Copilot Studio GCC. Rather than accepting this limitation, the Grand Traverse County team is building a custom feedback mechanism — a lightweight Power Automate flow that captures resident satisfaction signals and routes them to the IT team for knowledge base improvement. This kind of workaround is exactly what government AI practitioners need to share with each other — not the polished demo, but the real implementation detail.
Additional roadmap items include expanded department content coverage, multilingual support, and CSAT analytics integration as GCC feature parity with commercial Copilot Studio continues to improve.
A public AI agent that answers from memory — rather than verified documents — creates liability the moment it gets something wrong. Constraining generation to retrieved content isn't a limitation; it's the feature that makes the agent trustworthy.
Building inside Microsoft GCC means accepting that some commercial features don't exist yet, and that some integrations require creative workarounds. Knowing this upfront lets you design around the gaps rather than discover them in production.
An AI agent is only as good as its knowledge base. Defining who owns each document category — and when they update it — is an organizational challenge, not a technical one. CORA Connect's content cadence governance was designed before the first document was uploaded.
A soft launch with no marketing push generated over 1,400 resident sessions. That's residents actively choosing to use the agent to find county information — validating the core premise before any formal rollout.
We built CORA Connect in Grand Traverse County and have shared the implementation runbook with peer counties across Michigan. The architecture, governance framework, and lessons learned are yours.
Start a Conversation