A water utility scopes a parks FAQ assistant that must answer only from curated municipal content. What should they build?

A water utility building a parks FAQ assistant that must answer only from curated municipal content should scope a custom agent to that curated parks FAQ set rather than allowing open-ended ungrounded answers. Grounding and scope are the point: residents get consistent answers from approved FAQ material, and the agent is constrained so it does not freestyle beyond the curated corpus.

A custom agent scoped to curated parks FAQ content with generative web search for added coverage fails the “curated only” requirement because it pulls in unapproved material and risks hallucinated park rules. A custom agent scoped to curated parks FAQ content tuned for email summarization for visitor queries fails by task mismatch—it is optimized for condensing mail, not grounded question-answering over curated municipal parks FAQ content. A Copilot Studio agent scoped to curated parks FAQ pages with topic triggers for scripted replies fails because it relies on rigid scripted topic paths rather than grounded generative answers from the curated municipal corpus.

Exam caveat: Curated grounding reduces but does not eliminate incorrect answers if the FAQ is stale; content owners must maintain the corpus. AB-900 stresses scoped custom agents versus agents adding generative web search that leave curated content behind at conceptual depth.

Operational check: Load only approved parks FAQ content into the agent’s knowledge scope, ask in-scope and deliberately out-of-scope questions, and confirm in-scope answers cite the FAQ while out-of-scope prompts refuse or escalate rather than inventing policy.