The data protection officer walks into the IT admin's office with one question: which AI tools are we actually using? He opens the proxy logs and sees chatgpt.com, Copilot and a handful of domains he does not recognise. Half the staff are working from home today, and he sees nothing of their traffic.
The answer is not one method but a choice about what to measure. Every common source sees a different part of the use. The source that sees the most is the place where the work happens: the browser. And the measurement that stays useful longest is the continuous one, with no user id.
Which methods are there, and what does each one miss?
| Method | What it sees | What it misses |
|---|---|---|
| Network monitoring and DNS | Domains visited from the office network | Remote work and phones; free versus business account on the same domain; AI features with no domain of their own |
| Browser and endpoint approaches | What is opened in the browser, at the office and at home, including extensions | Unmanaged phones; and often records too much: path, content and user, unless you deliberately switch that off |
| Surveys and questionnaires | Why people choose a tool, and what they miss in the approved one | What people are not willing to report; out of date the moment it closes |
| Expense claims and invoices | Paid subscriptions and licences | Everything free, and everything someone pays for privately |
| SaaS management and SSO logs | Tools that run through the organisation's sign-in | Everything without SSO, which by definition is the shadow AI |
The bottom two are the least suited to finding shadow AI and the best suited to something else: they describe what the organisation itself has set up. That is the list of approved tools, and you need that too.
The top two look at traffic, and the difference is where they stand. The network only sees who is at the office. The browser travels with the person. That is the same shift described in the browser became your biggest security blind spot: the work has moved into the browser, and most security still looks at the network.
What does almost every method miss?
Four categories fall through the cracks everywhere, and together they are larger than what you do see.
Remote work and phones. A laptop on a home connection is outside the office network. An unmanaged phone is outside everything. Only a measurement that sits on the device itself goes home with it.
Free accounts. On the network, a free and a business ChatGPT account look identical: the same domain. On the invoice, the free account does not appear at all. Yet the difference in terms between those two is the biggest privacy difference there is.
AI features inside existing tools. Your meeting tool has been summarising calls since last month; your PDF reader has acquired an assistant. There is no new domain and no new purchasing decision, so none of the five methods fires. No measurement helps here, only a fresh assessment when something changes: an update, new terms, an acquisition.
Browser extensions. An extension with read access to every page sees more than most AI services, and it appears in no log as a visited domain. Browser extensions are AI tools too explains why they belong in the register.
Why does a one-off inventory age?
An accountancy firm surveys all its teams in the spring and arrives at eleven AI tools. Two months later the network logs show a translation service nobody mentioned, the case management software has gained an AI summary, and the tax team is using a specialised tool that did not exist when the survey ran. The list of eleven is not wrong. It describes the spring.
That is not solved by surveying more often. Why an AI tool catalog ages describes the three mechanisms: tools arrive daily, AI appears inside software you already have, and tools change owners and with them their terms. A snapshot, however often you take it, remains a snapshot.
What does work is turning the question around. Not: which tools exist, or which ones people might use. But: which services were opened in our organisation this month. That list is short, a few dozen, and it keeps itself current as long as it comes from observation.
How do you do it without a user id?
By recording only what the governance question needs, and nothing more. The question is: which AI services are used, how often, and is there a decision on them. For that answer, the domain of the service is enough. Who opened it, which page, and what was on that page add nothing to the decision and do change the character of the measurement.
That is why the BeeSensible extension records only the registrable domain, at most once per day per device, with no user id, no path and no page content. It does so only when the organisation has switched on the AI module and the analytics setting. The observed domains are matched against the catalog of 865 AI tools, and an admin sees per tool whether a decision exists; new tools start on "no decision yet".
