Somewhere in your business, someone has pasted a customer email into a free chatbot to get help with the reply. They were not being careless. They were trying to get through the inbox faster, with a tool they already use at home. That is what "shadow AI" usually looks like in a small company: not a rogue IT project, but well-meaning people using consumer tools for work because nobody has offered them anything better.

How common it is

The best recent numbers come from the UK, but there is little reason to think Swiss offices are very different. A Microsoft survey of more than 2,000 employees, reported by Euronews in October 2025, found that 71 per cent had used AI tools their employer had not approved. Over 40 per cent of those said they did so because they already knew the tools from their personal lives. Only 30 per cent said they worried about the privacy of the company or customer data they entered. The uses were mundane: replying to emails, drafting reports and presentations, and occasionally finance tasks.

Swiss voices are raising the same concern. Swisscom's 2026 threat report, covered by Swissinfo, named unauthorised AI tools in everyday work among the specific risks that AI brings when oversight is weak.

What the risk actually is

The problem is not that staff use AI. It is where the data goes and what happens to it afterwards. The UK National Cyber Security Centre, in a September 2026 blog on the subject, points out that information entered into consumer AI services may be stored, retained or used to improve the service, outside the organisation's control, unless specific privacy settings are in place.

For a Swiss SME, that matters for a concrete legal reason. The Federal Data Protection and Information Commissioner (FDPIC) has confirmed that the revised Data Protection Act applies directly to AI-supported processing, without waiting for any AI-specific law. If an employee pastes a client's personal details into a tool whose terms allow the provider to reuse them, the business, not the employee, has to answer for that data flow.

There is also commercial confidentiality to think about: pricing, tenders, contract drafts and unreleased product details are not personal data, but you would not want them in someone else's training set either.

Why a ban rarely works

The instinctive response is to forbid it. In a small company, that tends to push the behaviour out of sight rather than stop it, because the tools are on everyone's phone. The UK NCSC takes the same view. It does not recommend that people stop using AI, and argues that organisations should aim to reduce the risk rather than assume it can be removed. Its most useful line is also its simplest: "You cannot manage what you do not know."

The better approach is to give people a sanctioned tool that meets their needs, set clear rules about data, and make it easy to ask questions.

A one-page AI use policy for an SME

You do not need a long document. You need something staff can remember. A workable version covers these points:

  1. Name the approved tools. One or two business-grade AI services, set up under a company account with terms that exclude your inputs from model training, and with data residency you have checked. Everything else is unapproved for work data.
  2. Sort data into three groups.
    • Green: public information and your own generic text. Any approved tool.
    • Amber: internal documents, non-sensitive customer correspondence. Approved tools only, signed in with a company account.
    • Red: health, financial, HR and identity data, passwords, client secrets under NDA. Never into a general chatbot; only through a purpose-built system that has been reviewed.
  3. Strip identifiers by default. If a task can be done with "the client" instead of a name, do that.
  4. Check the settings once, centrally. Someone should confirm history, training and sharing settings on the company accounts, rather than leaving it to each user.
  5. Keep a human on outputs that leave the building. AI can draft; a person sends.
  6. Make it safe to admit past use. Ask staff which tools they have been using and for what. Treat the answers as requirements, not confessions.
  7. Review every six months. Tools and terms change quickly.

The OWASP guidance on sensitive information disclosure in AI applications makes a similar point from the technical side: educate users not to enter sensitive data, and prefer services with clear terms on retention, use and deletion, including an opt-out from training.

Turning shadow use into a roadmap

There is an upside here. Shadow AI is a free survey of where your team feels the most friction. If three people are using a chatbot to summarise supplier emails, that is a candidate for a proper, supervised automation with defined data flows. If someone is drafting quotes in a consumer tool, that tells you which document workflow to fix first.

Handled this way, the conversation shifts from "who broke the rules" to "what do you need, and how do we provide it safely". That is a better conversation for morale and for data protection alike.

Want to know where your team is already relying on AI, and which of those uses deserve a proper solution? The free AI Readiness Audit is a good place to start.

Sources

  • Euronews, "Most UK employees use AI at work without permission, Microsoft survey finds", 14 October 2025. link
  • Swissinfo (Keystone-SDA), "Telecoms firm warns of sharp rise in cyber threats in Switzerland", 27 April 2026. link
  • UK National Cyber Security Centre, "The hidden risks of shadow AI", 7 September 2026. link
  • Federal Data Protection and Information Commissioner (FDPIC), "Update - Current data protection legislation is directly applicable to AI", 8 May 2025. link
  • OWASP Gen AI Security Project, "LLM02:2025 Sensitive Information Disclosure", 2025. link