A new colleague at an insurer asks her team lead whether she may use Perplexity to look up policy terms. The team lead knows Copilot is allowed, because there is a licence for it. About Perplexity she knows nothing, so she says: "Better not, for now." The colleague uses it at home anyway, on her own laptop.
The answer to which AI tools employees may use is a decision the organisation takes. It is not printed on the tool. An organisation that has not decided is not a strict organisation; it is one where every team lead works it out alone.
A tool is not "safe" or "unsafe" on its own. The same chatbot can be used to rewrite an internal email or to summarise a client file, on a free account or on a business licence with a data processing agreement underneath. Those are four different situations, and the tool is the same in all four.
So "is this allowed?" is a question for the organisation, not for the vendor. The vendor tells you what happens to your data. What you want to do with that is a weighing you make yourself: against the work people do, against the data that passes through it, and against the agreements already in place. The AI Act asks organisations that deploy AI systems to make sure their staff are sufficiently AI literate. Knowing which tools the organisation has allowed, and why, is the most practical place to start.
#How do you get to a list?
Four steps, in this order. The temptation is to start at step two.
Take stock of what is already there. Not the tools you know, but the tools being opened. That list is almost always longer than expected, and it contains tools nobody ever took a purchasing decision on: a transcription tool someone tried in a meeting, a browser extension, an AI feature inside software that was already there.
Give every tool a status. Four are enough: allowed, limited use, no decision yet, not allowed. What each one means is set out in four statuses for AI tools. What matters most is that every tool has exactly one, including the tools you still need to look at.
Write the condition down. "Limited use" says nothing without the sentence after it. "Only without personal data" is a condition. "Internal meetings only, with the participants informed" is a condition. "Use with care" is not. What may and may not go into a prompt is covered independently of the tool in what not to share with AI; the condition per tool comes on top of that.
Bring the list to the work. A list in a PDF on the intranet is read once, on the first day. The decision has to be visible at the moment someone opens the tool, because that is when the question arises. This is what a list looks like when it sits with the use instead of in a document:
A few example decisions, with the tools deliberately unnamed, because the decision differs per organisation.
#Why does a banned list not work?
Because a ban does not take the work away. Someone who has to summarise a forty-page report still has that report when the tool is off limits. The work then finds another route: a private account, a personal phone, a tool nobody knows about. That is exactly the use that falls outside any view and outside any data processing agreement, as described in blocking AI creates shadow AI.
A list with alternatives works differently. "Not this one, that one" is a decision a person can follow without stopping their work. The approved alternative does have to be a real alternative: the same kind of work, roughly as easy. Replacing a chatbot with "ask a colleague" is not an alternative.
Research by KPMG and the University of Melbourne among 48,000 workers in 47 countries (2025) found that 58 percent deliberately use AI for work and 57 percent keep that hidden from their employer. A list that only bans gives those 57 percent no reason to come into view. A list that says what is allowed instead, per banned tool, does.
#Is a free account different from a business account?
Yes, and this is the distinction most often missing from a list. A free or personal account runs under consumer terms: input can be used for training unless the user switches that off, there is no data processing agreement, and the organisation cannot configure anything centrally. A business licence of the same tool excludes training by default and comes with a data processing agreement. OpenAI states that split explicitly for Team, Enterprise and the API, and the other major vendors draw the line in a similar place.
For the list this means that "ChatGPT: allowed" is not a decision. "ChatGPT on the business licence: allowed; ChatGPT on a free account: not allowed, use the licence" is one. The difference between those two accounts is worked out in free versus business AI accounts.
New AI tools appear every day, and existing software gains AI features without anyone taking a decision. So a list is never complete. That is not a problem, as long as you know the status of a tool that is not on it yet.
That status is: no decision yet. It is neither a ban nor an approval, and it should be visible to whoever opens the tool, with a request to enter only public information for the time being. What is still needed is a short route to a decision: who looks at it, within what time, and on what basis.
A municipality had allowed Copilot and considered the matter closed. An inventory showed Perplexity, ChatGPT and three specialised tools in use next to it, two of which were recorded nowhere. That was not an incident. It showed that a list drawn up once quickly falls behind what people do, and that "no decision yet" is a status you need.
BeeSensible links a catalogue of 865 AI tools to the tools being opened in the organisation. An administrator sees which of those still have no decision and records that decision, with a condition or an alternative; the decision then appears to the employee at the moment they open the tool. The usage data carries no user id and no content. How that works in the product is on the AI tools product page.
How to write those decisions into a policy people will actually read is covered in writing an AI policy. The other pieces on this subject are collected on the AI governance topic page.
The next time a colleague asks whether a tool is allowed, the answer is not "better not" but "check the list, and if it is not on there, that is its status". It is a smaller answer, and it keeps the use in view.
