Shadow AI in Swiss SMEs: Ban or Channel It?

The question is not whether your employees use ChatGPT, Claude or Gemini. The question is whether you know what data leaves your company – and who bears personal liability if something goes wrong. The answer is not a ban, but channelling. Blanket bans push shadow AI underground. A sanctioned company account, clear data categories and a short policy work better.

Shadow AI is mainstream

The IW Köln brief from September 2026 puts the share of professional AI users relying on tools not provided by their employer at 29.1%. Other surveys show 42% of companies report employees occasionally using unapproved AI tools, 34% frequently. Only 9% fully prohibit AI. The grey zone is large – and growing.

Why bans fail

A ban sounds clean, but it doesn't work. Employees keep using private accounts because those tools are faster than internal approval processes. A prohibition culture creates opacity: anyone afraid of consequences hides their use, making mistakes invisible. The sanctioned alternative must be better than the private workaround – not just permitted, but easier, faster and safer.

The third way: Channel instead of ban

Channelling means steering usage into controlled paths. That includes a sanctioned company account with a data processing agreement (DPA), clear data categories, a named contact person and a short, understandable AI policy.

1. Provide a sanctioned company account that is contractually and data-protection compliant.
2. Define data categories: what may go into an AI tool and what never may.
3. Name a contact person for questions and approvals.
4. Adopt an AI policy that fits on one page.

What the revised Swiss DPA means for you

The revised Federal Act on Data Protection (revDSG) is technology-neutral and applies directly to AI. Crucially, personal liability matters: under Art. 60–63 DPA, fines of up to CHF 250,000 are imposed on the natural person who acted, not the company. Intent is required, and prosecution occurs upon complaint. In minor cases with fines up to CHF 50,000, the company may be convicted instead. This makes the question of who approves a tool and on what contractual basis a leadership issue.

Data categories: What never belongs in a public AI tool

The most effective safeguard is a clear, practical list of prohibited data types. Customer personal data, financial figures, internal emails and strategic information deserve special caution. The rule of thumb: anything that would be existentially damaging if leaked does not belong in a public AI tool – unless a DPA and technical protections such as anonymisation or pseudonymisation are in place.

Bans, laissez-faire, channelling compared

Building an alternative employees will actually use

The first step is not the policy, but a better offer. A sanctioned company account with a clear contractual basis, short approval processes and training on equal footing give employees what they seek: fast help without guilt. When selecting tools, scrutinise the DPA clauses.

For a broader perspective on avoiding strategic mistakes in the global race, the Swiss AI podcast offers guidance on how Swiss SMEs avoid the biggest errors in AI strategy.

First steps without liability traps

Start with an inventory: where is AI already used, which tools, and what data? Then draft a one-page policy together with the teams – that builds acceptance. Update it regularly, because tools change quickly. An honest policy that names a real alternative works better than long documents.