GovTech Steward — Responsible Public-Sector AI Adoption at KABD

A four-round, advanced public-governance simulation set inside the Kantonales Amt für Bürgerdienste (KABD), the Cantonal Office for Citizen Services in the fictional Swiss canton of Brunnental — a 540-person agency serving 620,000 residents on a CHF 96M budget, with CHF 6.5M ring-fenced for digital transformation. KABD is the most trusted institution in the canton (72% public trust) but one of the slowest: a 14,000-case backlog, a 19-day average processing time and a citizen-satisfaction score of 61/100. On 11 June 2026 a neighbouring canton's automated benefits-eligibility system is suspended after it is found to disadvantage applicants by postal code and language; parliament demands KABD present a responsible AI plan this quarter — clear the backlog and cut the wait without repeating the failure. Playing the agency director (with a digital/AI lead, a data-protection & legal officer, a service-delivery manager and a communications lead), you steward AI adoption through the lens of Reto Steiner's research on public-sector digital transformation and public corporate governance (Public Management Review, 2024): public-sector AI must be governed for public value, accountability, transparency and legitimacy, not merely efficiency. Round 1: diagnose where the delay originates, risk-tier the four candidate use cases (UC-1 document triage, UC-2 benefits-eligibility scoring, UC-3 multilingual chatbot, UC-4 fraud detection) by rights-impact and trust-risk, and state the agency's non-negotiable trust commitment. Round 2: within the CHF 6.5M cap, prioritize one use case to deploy first and set the governance safeguard — human-in-the-loop, transparency, appeal and contestability rights, bias testing and a named accountable owner — confirming compliance with the revised FADP before deployment, and ring-fencing budget for ongoing monitoring (the FADP deadline is this round). Round 3: the data-quality audit reveals the historical training data carries the same postal-code skew that sank the neighbour; decide whether to proceed, narrow, or pause and remediate, then sequence the rollout — pilot scope, monitoring gates, fallback and rollback — avoiding the inequity of piloting on the most vulnerable segment first. Round 4: defend the plan to the parliamentary oversight committee and a journalist in the language of public value, accountability and trust, owning the residual risk rather than defending a metric or blaming the data. The scoring rewards risk-tiered prioritization, a complete safeguard package, data remediation, equitable sequencing and a legitimacy-framed defence — and punishes the five classic errors: leading with the highest-risk benefits-scoring use case, automating rights-affecting decisions with no safeguard, ignoring the biased data, sequencing risk onto vulnerable citizens, and defending on efficiency alone. Final KPIs track citizen trust (%), governance compliance (0–100), backlog cleared (of 14,000) and budget discipline against the CHF 6.5M cap.