What the Federal Council was told
On 18 September 2026 the Federal Council was briefed that the federal administration is scaling back the implementation of its AI strategy. According to the federal government's announcement, the Federal Chancellery re-prioritised the implementation plan and attached a 2026–2028 roadmap. Six of eight supporting measures continue, funded at CHF 2 million a year.
The surviving measures include a generative AI system available to all federal employees, a federal-wide assessment of AI's potential for process optimisation, and basic documents on the legal framework for using AI. The planned AI handbook is dropped entirely. The planned central "KI-Marktplatz", a place to find and reuse AI solutions, is replaced by an expanded version of the existing AI project database. Some other measures will move more slowly or at smaller scale. The Chancellery stays coordinator, and the measures will be reviewed by the end of 2028 at the latest.
The cut belongs to a wider squeeze. The same announcement ties it to the Federal Council's 18 February 2026 decision that some federal digitalisation projects must slow down. Netzwoche's report describes the same package of continued and dropped items.
The strongest case against trimming
Critics of the cut have a serious argument. A government that asks the private sector to adopt AI responsibly has to show it can do the same. Handbooks and shared marketplaces are the unglamorous infrastructure that stops 116 separate projects from reinventing procurement, risk assessment and documentation. The federal announcement counts 116 AI projects across departments, including pollen forecasting and automated detection of solar installations in aerial images. Without common guidance, each office risks making its own compliance judgments. Those judgments will be uneven, and some will be overcautious.
A second worry is signalling. Switzerland markets itself as a trusted, innovation-friendly location, and a visible retreat from the state's own AI programme could read as a lack of seriousness.
Why the cut is mostly defensible
The decision is less alarming than the headline suggests, for three reasons.
First, the dropped items are the most easily substituted. A handbook written by a central office is quickly outdated in a field that changes every few months. The Chancellery keeps a competence network of AI practitioners who share knowledge across offices, which does most of what a handbook was meant to do. Replacing a new marketplace with an expanded project database avoids building a second platform. Offices still get a place to find existing solutions at lower cost. It is a reasonable use of limited funds.
Second, the measures that survive are the ones that deliver value directly. A generative AI system for all employees, and a systematic assessment of where AI can actually improve processes, are the parts of the strategy that change how the administration works. The assessment also answers a parliamentary request, Postulate 24.3582. Spending scarce money on use rather than documentation is consistent with an evidence-first approach.
Third, a roadmap with a hard review date at the end of 2028 is better than an open-ended programme. It forces the government to ask what worked, rather than treating every measure as permanent.
The real risk is on the legal track
The part that should concern readers is not the handbook. It is the legal framework. The Federal Council decided on 12 February 2025 that Switzerland will incorporate the Council of Europe AI Convention into Swiss law. As BAKOM's overview of AI policy states, the FDJP, working with DETEC and the FDFA, is to prepare a consultation draft by the end of 2026. That draft is not yet out as of this writing.
Switzerland chose this route deliberately. Rather than copy the EU AI Act, it favours targeted amendments to existing, sector-specific law, with general rules limited to areas touching fundamental rights. This is the right instinct for a small, open economy. Sectoral regulators such as the data protection authority already understand their domains, and a horizontal statute would add compliance burden on firms that already face the EU's rules when they sell into the single market. Voluntary commitments and industry solutions are meant to fill the gaps.
That design only works if the administration has the capacity to do it well. A "tailored rules" approach means many small amendments across many statutes, and each one needs analysis of how AI is actually used in that sector. That is skilled, slow work. The federal AI strategy's retained measure on the legal framework for the administration's own AI use is part of that capacity, and it should be protected from further cuts.
What to watch
- The consultation draft. If it arrives by the end of 2026 and keeps its focus on transparency, data protection, non-discrimination and supervision, Switzerland keeps a credible, light-touch model.
- Whether the slower measures slip further. The announcement says some measures will proceed more slowly or on a smaller scale. The 2028 review should publish what was delayed and why.
- Whether the cut spreads. If the same pressure reaches the regulators who will apply the new rules, a light-touch approach could become no oversight at all.
Bottom line
Trimming a handbook and a marketplace is not a retreat from AI. A state that spends CHF 2 million a year on the tools staff actually use, and reviews the results in public, is arguably practising the proportionate approach it asks of industry. The test comes in the next few months, when the legal draft is due. A lean strategy is only a good strategy if the law that follows it is ready on time and well made.