Evidence note: AI in public administration and planning

Last checked: 19 August 2026. Version française.

This note bounds the synthetic A2 planning-dossier case. It separates administrative checks from the exercise of public authority.

Field anchors

• In a three-month Milton Keynes pilot, staff reported validating applications 60–80% faster. Receipt-to-validation time fell from 15.8 to 7.6 days. Each decision remained with a staff member. This is a close field analogue, but 80% is a task estimate, not a saving for the whole service.
• Cambridge PlanAI generated summaries and a report in 16 minutes for 320 contributions. Staff had spent over 60 hours logging and summarizing, including 18.5 hours on summaries. Planners still read every contribution and make the decision.
• Leeds and Xylo provide a strong governance analogue: visible sources, planner approval, impact assessments, supplier requirements, audit trail, and a public record. The case does not publish a complete causal comparison.
• The UK assisted-decision prototype aims to halve handling time for simple cases. It is an alpha ambition, not an observed result, and its legal context does not transfer unchanged to Switzerland.

Swiss authorities must also apply existing information-security, data-protection, sectoral, cantonal, and procedural rules. Useful official references include the Federal Chancellery, FDPIC, and the Swiss Federal Audit Office.

The worked case uses a narrower boundary: identify documents, rules, and context, but do not recommend the decision, draft its legal grounds, or prioritize people. High task-level scenarios remain useful for planning when labelled as scenarios. Measure administrative checking, assessment, decision, notification, appeal, accessibility, and downstream outcome separately.