Municipal budget chatbot prototype reports better answers, with incomplete validation

Start with the finance director, communications lead, clerk and resident-service manager. Ask which recurring budget questions are difficult, who validates fiscal-year distinctions, and how residents currently obtain corrections. A bounded engagement could test a public-document assistant against the existing website and staff-assisted service. The value hypothesis is easier access to verifiable answers; this prototype does not support guaranteed savings, head-count reductions or increased public trust. Establish demand before proposing a purchase. For a smaller municipality, identify the staff time needed to maintain the corpus and answer escalations. Include residents who use assistive technology and those who prefer a non-chat channel in discovery.

Build a read-only proof of value using approved budgets and a separately maintained answer set. Evaluate retrieval, calculations, fiscal-year selection, projected-versus-actual distinctions and multi-turn ambiguity independently. Show the exact supporting passage and provide a fallback when evidence is missing. The prerequisite is a versioned corpus with a finance owner and representative resident questions. Compare the assisted workflow with ordinary document search, using current models rather than assuming the historical ranking persists. Test prompt injection, log retention and tool access. Measure error severity, response latency and cost, and assess whether finance reviewers can detect incorrect answers. Neither code availability nor citation presence is sufficient acceptance evidence.

Assign finance as content owner and IT as service operator. Prepare documents, construct the evaluation set, train service staff and define a correction route before a limited pilot. Dependencies include document accessibility, fiscal-year metadata, reviewer availability and maintenance funding. Proposed acceptance criteria are correct year and value on every critical test, accurate citations on the agreed sample, a demonstrated escalation path and no material accessibility regression against the existing service. These are proposed criteria, not study results. Track net staff effort and resident task completion alongside adoption. Pause publication of unsupported answers, and rerun checks after budget revisions, prompt changes or model upgrades.