Of all the European companies that considered using artificial intelligence and then dropped the idea, 70.89% gave the same reason: a lack of relevant expertise. The figure is Eurostat’s, from 2025 data. Behind it come uncertainty about the legal consequences (52.52%) and fear of breaching data protection rules (48.83%).

That ranking tells me more than the headline does. Technology can be bought in an afternoon and gets cheaper every year. What cannot be bought is what the company already knows, spread across inboxes, shared drives, an ERP and three people who have been there fifteen years.

Adoption, meanwhile, is moving fast: in Spain it went from 11.31% to 20.27% of companies with ten or more employees in a single year. What is not moving at the same speed is the state of the material those systems are fed.

Information is not knowledge

Digitising documents was the step before this one, and we have all taken it. Storing is not knowing. For a document to become corporate knowledge it has to meet four conditions at once, and that is where almost everyone falls short:

  • Context. Who wrote it, for which case, and within what limits. A price with no record of which client it applied to or on what terms is not a price: it is a loose number that somebody will eventually use badly.
  • Currency. Someone has to be able to say it is still live. An old version of a policy is not harmless because it is old: it is dangerous because it reads exactly as convincingly as the current one.
  • Accessibility. The person who needs it finds it at the moment they need it, not two days later by internal email. Knowledge that arrives late no longer changes any decision.
  • Governance. A named owner who maintains it, corrects it and authorises its use. With no name attached, a document ages on its own and nobody notices until someone quotes it in a meeting.

Without all four, an organisation can hold twenty thousand files and still run on individual memory.

The part nobody wrote down

The most valuable knowledge is the kind that was never documented: who knows which client to phone before sending anything, which exception can be accepted without asking, which supplier promises deadlines they will not meet. It is built over years and leaves in an afternoon. It is rarely noticed the next day; it shows up three months later, when somebody decides badly without knowing a precedent existed.

Our rule across the group’s companies is simple and, at first, unpopular: no internal assistant answers a question about a price, a contract or a policy without stating which document it came from and what date it carries. If it cannot cite it, it says it does not know. We prefer an honest “I don’t know” to an answer that sounds right, because the answer that sounds right and is wrong never surfaces in internal review — it surfaces in front of a client. That requirement lives in the technology all our companies share, not in each project separately; left to each team’s judgement, it survives two releases.

No traceability, no trust

Under the EU AI Act, as the European Commission summarises it, high-risk systems must log their activity so that results can be traced. An internal knowledge assistant almost never falls into that category and is therefore obliged to do nothing of the sort. We write the logs anyway. The day an answer goes wrong, the question will not be whether keeping them was mandatory, but who can reconstruct where it came from.

And there is an uncomfortable part. Integrating a company’s knowledge does not mean opening it up: a system that gives everyone the same answer has quietly deleted the permissions the organisation had in writing. Confidentiality levels get replicated, not flattened. Who signs off those decisions is something we cover in another article, and what happens when the system does not know that it does not know, in where a digital worker actually breaks.

A company’s knowledge is rarely lost all at once. It is lost because nobody ever decided whose it was.