What AI can't find doesn't exist for it

What does your AI know about your company? Usually less than you think, and the little it knows isn't always the right thing.

AI has arrived in Swiss SMEs. Around a third now use it actively, up from just over a fifth a year earlier (AXA/Sotomo, 2025). The tools keep getting better and cheaper. The foundation they work on has barely changed: the knowledge inside the company.

🐛 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 in mailboxes, old project folders or three versions of the same quote. Some was never written down, because everyone knew whom to ask.

Point an AI at this and it either finds nothing and answers in generalities, or it finds the wrong version and answers convincingly. Neither is obvious at first. On top of that, people check AI output less the more they trust the AI, as a study of 319 knowledge workers shows (Lee et al., 2025).

Students at OST recently saw how direct this link is. On violet dragonfly they built their own knowledge base and an AI advisor to go with it. Whatever they hadn't uploaded, the advisor didn't know. That is intentional: an advisor that answers only from verified knowledge doesn't fill gaps with inventions. The same holds in a company, except there's more at stake than an exercise.

🦋 The pragmatic solution

AI readiness therefore starts with knowledge management, not with the tool. Context engineering is knowledge management with a new consumer, and established methods carry more weight as a result. Four questions help to sort things out:

🦋 What may the AI know? Classify information, for example as public, confidential or secret, and decide who may see which answers.

🦋 Where does the AI get its knowledge? Name the sources, assign one responsible person per source and delete what is outdated.

🦋 When do we trust it? Allow only answers with a source, explicitly permit "I don't know" and check with a fixed set of test questions.

🦋 What must people be able to do? Check, correct and approve results. That takes practice, otherwise the skill withers.

What verified knowledge looks like

Take a machine builder with 60 employees. Sales selects the 20 most important documents: product sheets, standard clauses, two good quotes. Each document gets a responsible person and a review date. The Wingmate, the digital companion on violet dragonfly, answers on this basis, and every answer shows its source. If a question gets tricky, it goes to the specialist together with the full conversation.

Analysing the conversations shows where knowledge is missing, so the team can fill gaps deliberately instead of documenting on suspicion. After a few weeks, a discount rule might surface there that appears in no document and exists only in the sales manager's head. Knowing about that gap is worth more than any quick answer.

Does this need an IT project?

No. Documents go into the knowledge base by upload, and existing repositories can be connected via open interfaces (MCP). The language model is selectable, including an open-source model, and the content is not used for training. The effort lies in selecting, checking and maintaining knowledge. That work belongs to the subject-matter experts, not to IT.

This keeps the human at the centre. AI amplifies, it never replaces.

Benefits at a glance

🦋 Answers based on your company's current knowledge

🦋 Less dependence on individual people, including when they retire

🦋 Clear rules on what the AI may know, before confidential data comes into play

🦋 Employees who check AI results instead of adopting them blindly

🦋 Visible knowledge gaps instead of invented answers

Conclusion

An AI is only as good as the knowledge it is allowed to find. At violetdragonfly.io you can see how an AI answers from well-organised knowledge, with a free daily allowance of Wingbeats.