A municipal open-data office asks whether a generative model is needed or a rules-plus-search design suffices for structured dataset lookups. What should audit challenge?

When a municipal open-data office asks whether a generative model is needed or a rules-plus-search design suffices for structured dataset lookups, audit should challenge whether AI is necessary given use-case complexity that a simpler controlled design may meet. Predictable schema lookups (“which CSV holds pothole counts”) are often better served by governed catalogs, APIs, and search with deterministic filters. Generative models add hallucination and citation risk without proportional benefit when the task is retrieval of known structured assets.

Whether embedding model upgrades improve precision for structured dataset filtering is the wrong starting point: it presumes an AI pipeline is already justified and ignores the threshold question of whether a simpler catalog or search design can handle predictable lookups. Whether response caching lowers latency for repeated open-data portal queries is a performance tuning issue, not a needs analysis; it does not address whether generative AI is necessary or whether rules-plus-search already suffices. Whether provenance links connect generated summaries to source dataset versions for public reuse addresses transparency after generation, but it does not answer whether generative AI is needed for structured lookups or whether a simpler controlled design would meet the use case.

Exam caveat: necessity and proportionality are core AI governance themes—capability does not equal appropriateness. Operational check: document task predictability, error cost of wrong dataset pointers, compare a catalog/search prototype against GenAI on the same queries, and approve GenAI only where ambiguity or synthesis clearly exceeds simpler designs.