AI-Assisted 311 Triage: Building in Public

Years of 311/service-request work taught me where intake actually breaks: ambiguous descriptions, inconsistent categories, duplicates, urgency buried in free text, and routing errors that cost days. It's also exactly the workflow where AI help is plausible and AI overreach is dangerous — a wrong routing suggestion silently accepted is worse than no suggestion at all.

So this prototype is my working answer to a question every GovTech product team is facing: what does responsible AI assistance look like inside a consequential government workflow?

Request intake with map context; AI-suggested category, department routing, and duplicate candidates — each with confidence displayed, not hidden; urgency flags; and a human review step where the agent confirms or overrides in one click, with override reasons captured.

That override-with-reason loop is the design center. It keeps a person accountable for consequential routing, and it turns disagreement into data: the override log is both the audit trail and the improvement signal. Built with AI-assisted development (Claude Code), every line human-reviewed, synthetic requests and open geospatial data only.

The evaluation plan measures trust, not just accuracy: suggestion acceptance rate, downstream routing corrections, duplicate precision, time to triage, and override rate by category — a rising override rate in one category is the model telling you where it's wrong. Error distribution gets checked across geography, because an urgency model that's wrong in one neighborhood more than another is a fairness defect, not a rounding error.

The boundary rule for the whole project: the model recommends; a human owns high-consequence routing and final disposition until the evidence says otherwise. That's not a limitation of the prototype — it's the point of it.