Which AI tool should we introduce? I hear that question in almost every first meeting. The more important one is rarely asked: can the AI actually find what your company knows?
AI has arrived in Swiss SMEs. Around a third are now actively using it, up from just over a fifth a year ago (AXA/Sotomo, 2025). The tools are getting better and cheaper. What has barely changed is the foundation they work on: the knowledge inside the business.
🐛 The problem
An SME's knowledge rarely sits where an AI can find it. Some of it lives in the heads of experienced staff. Some is buried in inboxes, old project folders, or three versions of the same proposal. And some was never written down at all, because everyone knew who to ask.
When you point an AI at that, one of two things happens. Either it finds nothing and gives a generic answer. Or it finds the wrong version and answers confidently. Neither failure is immediately obvious. A study of 319 knowledge workers adds a second effect: the more people trust an AI, the less they check its results (Lee et al., 2025).
Students at OST saw this dynamic directly. In a workshop they built their own knowledge base and an AI advisor to go with it. What hadn't been uploaded, the advisor didn't know. One sentence stuck: what the AI can't find doesn't exist for it.
🦋 The pragmatic solution
AI readiness therefore starts not with the tool, but with knowledge management. Context engineering is knowledge management with a new consumer. The proven methods matter more, not less. Four questions help with the sorting:
🦋 What may the AI know? Classify information — public, confidential, secret — and define who may see which answers.
🦋 Where does the AI get its information? Name sources, assign a responsible person to each, and delete anything outdated.
🦋 When do we trust it? Answers with cited sources only, explicitly allow "I don't know", and test with fixed benchmark questions.
🦋 What must people be able to do? Check, correct, and approve results. That skill needs practice, or it atrophies.
What this looks like in practice
Picture a machine manufacturer with 60 employees. Sales wants to produce proposals faster. Instead of rolling out a tool straight away, the team selects the 20 most important documents: product sheets, standard clauses, two good proposals. Each document gets an owner and a review date. Old versions are deleted. Only then does the AI get access — and every answer shows its source.
After four weeks something unexpected surfaces. The most frequent question from the field team concerns a discount rule that appears in none of the documents. It exists only in the sales director's head. The AI couldn't find it, and that is the most valuable result of the first month: the gap is now visible and will be written down.
What about data protection?
That is precisely why ordering knowledge before choosing a tool pays off. Once information is classified, you know what can go to which model. Public product sheets can be processed anywhere; confidential proposals stay with a provider in Switzerland or the EU. Without that classification, every tool decision is a gut decision.
Small steps, not a big project
"That sounds like a year-long project," I often hear. It doesn't have to be. One area, a handful of documents, one month. After that you can see where the knowledge holds and where the gaps are. Small steps are better than no steps.
Benefits at a glance
🦋 Answers grounded in your company's current state
🦋 Less dependence on individual people, including through retirements
🦋 Clear rules about what the AI may know, before confidential data enters the picture
🦋 Staff who check AI results rather than accept them blindly
🦋 Freedom to choose your tool, because your knowledge isn't locked inside it
Conclusion
Organisations don't become ready for AI by finding the best tool. They become ready by keeping their knowledge in good order.
👉 Where does your organisation stand? An AI potential analysis shows which knowledge is ready and where things still need work. Feel free to get in touch.