Swiss Federal Audit Office on Government Chatbots — 21 AI Chatbots Found, Bottom-Up Synergies Working but the AI Competence Network Under-Resourced

The Swiss Federal Audit Office (SFAO/EFK) audited AI use in the Federal Administration using chatbots as the worked example (audit EFK-24181, version with auditee statements dated 25 February 2025), covering the Federal Statistical Office, the Federal Chancellery and the Federal Office of Information Technology, Systems and Telecommunication (FOITT). The questions were whether synergies between projects were being used, whether hoped-for benefits could be achieved, and whether the Competence Network for Artificial Intelligence (CNAI) was fulfilling its role.
The SFAO identified 21 AI chatbot solutions, operational and in development. Most are low-cost, low-risk chatbots on public websites; the more complex solutions needing high data sovereignty and security standards are a minority and all still in development. Implementation costs range from CHF 9,000 to CHF 3 million. Examples reported by trade press in October 2024 include a consular-information bot (FDFA), the fragesi.ch foundation-supervision bot, a statistics-to-SQL prototype at the Federal Statistical Office, a RAG assistant at SECO and a planned knowledge assistant at the Federal Compensation Office.
Findings: knowledge is shared between projects and synergies are being used, but largely through bilateral, self-organised exchange because the CNAI, set up in 2021, played mainly a facilitating role and, with severely limited resources, could not fulfil its mandate; an independent evaluation by the Federal Statistical Office had already shown this and the Federal Council ordered measures. Overarching framework conditions (an AI sub-strategy by the Federal Chancellery, a regulation report by OFCOM) were still pending, to be developed by end-2025. The SFAO asked that lessons be consolidated and reusable tools on legal bases, data protection and infrastructure be provided quickly.
Evidence caveat: this is an audit of governance and coordination, not an outcome evaluation. No usage, accuracy or cost-saving figures per chatbot are published, so measured benefits remain undemonstrated. The practice is valuable as a transparent, independent stock-take and as a cautionary case that experimental, decentralised AI adoption needs a resourced coordination body.

Read the full analysis: https://www.efk.admin.ch/wp-content/uploads/publikationen/berichte/wirtschaft_und_verwaltung/informatikprojekte/24181/24181_wik-e.pdf