A security manager at an insurer wants to know which AI tools are being used across the organisation and asks IT for an inventory. The same day a reply comes back from HR: this cannot happen without the works council, because then we are tracking employees. Two emails, and the project stalls.
Both are right about their own interest, and both assume one comes at the expense of the other. That is the mistake. The organisation wants to know which tools; the employee does not want a record of what they do. The first question needs no person, so the second concern need never arise.
What does the organisation actually want to know?
Shadow AI is AI use people started on their own: a tool anyone can start online, on which the organisation has taken no decision, and which is actually in use. What that is and why a ban does not solve it is in what is shadow AI. The security manager's question is the ordinary follow-up: which of those tools do we have, and which ones need a decision?
That is a question about services. The answer is a list of domains, a count per domain, and a status per domain: allowed, limited use, no decision yet, not allowed. Nowhere in that list is a name, and nowhere is a name needed. The insurer remains the controller for what happens inside those tools, so the interest is real. It just does not ask for what HR fears.
What does the employee actually not want?
Not that the organisation sees nothing. Most people understand that an employer wants to know whether client data is ending up in a free chatbot. What they do not want is a record of themselves: which tools I use, how often, at what times, and the possibility that this comes back in an appraisal or a dispute.
That is a reasonable concern, and the GDPR takes it seriously: processing about employees needs a purpose, necessity, transparency in advance, and in the Netherlands the Works Councils Act adds to that. What an employer may and may not do is in can my employer monitor whether I use AI. But notice what the concern attaches to: the record of the person. Not the count of the tool.
Where is the difference?
In one field. An observation without a user id says: on this day, this domain was opened in the organisation. An observation with a user id says: this person opened this domain. The first is a count, the second a behavioural record. They can come from the same source and they are two different things, with a different legal basis, a different retention period and a different conversation with the works council.
| Interest | What it asks for | What it does not ask for |
|---|---|---|
| Organisation: sight of shadow AI | Which domains, how often, which status, how it shifts | Who opened the tool |
| Organisation: making the decision land | Showing the decision at the moment someone opens the tool | Recording who saw the notice |
| Employee: no record | No user id, no content, no breakdown per person | That the organisation sees nothing |
| Employee: knowing where they stand | The decision per tool, and the alternative | A ban without explanation |
The two bottom rows and the two top rows fit together. Not because someone struck a compromise; because the organisation's question was never about a person.
Why does oversight of individuals also work against the organisation?
Suppose the insurer opts for a measurement per employee after all. What happens? The 2025 study by KPMG and the University of Melbourne, among 48,000 workers in 47 countries, found that 57 percent of workers hide their AI use from their employer. That is the starting point, before anything has been announced.
Announce that use becomes visible per person and the use moves to the private phone and the private account. The security manager's figures drop, the use does not, and the sight the whole project was meant to deliver is gone. Oversight of individuals costs the organisation exactly the insight it introduced it for. A count without people has no such effect, because there is nothing to evade.
What does this look like in practice?
The insurer sets the measurement up on domain and count, presents the design to the works council, and switches it on. After six weeks the picture looks like this: Copilot is allowed and is opened most. ChatGPT and Perplexity also show up daily, along with three specialised tools nobody knew about: one that summarises policy terms, a translation tool and a transcription tool. Two of them have no decision yet, because they were on no list at all.
No incident, but work. The summarising tool sees clients' policy documents and gets a decision first. The transcription tool follows. Nobody needs to know who used them to decide that, and the works council saw the list before the security manager did. It also shows what a one-off inventory misses: the three unknown tools could not be asked about, because nobody knew they existed.
What does BeeSensible record?
The extension notices when someone opens an AI tool and reports only the registrable domain, at most once a day per device, with no user id, no path and no page content. Every tool starts at "no decision yet"; the organisation sets the status itself. When someone opens a tool that has a decision, that decision appears in the browser, with the approved alternative as the first button, and continuing is always possible. What happened after such a notice is counted per tool, without a user id.
