An IT administrator at a building services firm is asked by the directors to draw up a list of approved AI tools, "and then we'll put it on the intranet". He opens a blank document and types Microsoft 365 Copilot, because there is a licence for it. Then he stops. Half the estimators use ChatGPT, the project managers have their meetings summarised by a tool he knows only by name, and nothing has ever been decided about either.
A list of approved AI tools does not start from a blank document but from what is already in use. From there you record the account and the data for each tool, give every tool a status with a condition and an owner, put the list where people work and set up a process for new tools. How that list fits into a full AI policy is covered in our guide to writing an AI policy; this piece is about the list itself.
Why not start from a blank document?
Because a list drawn from the administrator's memory only contains the tools that cross his desk. The tools used most are often the ones nobody requested: a free ChatGPT account, a browser extension, the AI feature a meeting tool added last month.
There are three sources, in ascending order of usefulness. The licence and account records show what was bought. A survey shows what people are willing to report. Observing which AI domains are opened, in counts and without names, shows what actually happens. What each method misses is covered in our article on finding out which AI tools employees actually use.
What do you record for each tool?
| Step | What you record | The mistake it prevents |
|---|---|---|
| 1. Start from what exists | The tools that are opened, with the date first observed | A list of what you know rather than what happens |
| 2. Account and data | Which version it is (free, personal, business) and which data may and may not go in | "ChatGPT: allowed" without saying which ChatGPT |
| 3. Status and condition | One of four statuses; for "Limited use" the condition in one sentence, for "Not allowed" the alternative | A yes everyone reads differently, or a no with no way out |
| 4. Owner | The role in the business that uses the tool and carries out the review | Rows left exactly as they were filled in a year ago |
| 5. Publication | Where the decision appears: first when the tool is opened, then in a document | A PDF read once, on someone's first day |
| 6. Process for new requests | Who assesses, within what period, with which questions and what applies in the meantime | Requests left untouched while use continues |
Step two is the one most often skipped. A free or personal account runs under consumer terms: input can be used for training unless the user turns that off, and there is no data processing agreement. The business version excludes training by default; OpenAI documents the distinction itself. On the list that means two rows; why that difference matters more than the name of the tool is covered in free vs business AI accounts.
Which status does a tool get, and what goes with it?
Four statuses, not two: allowed, limited use, no decision yet, not allowed. What each status means and when you choose it is covered in our article on the four statuses for AI tools. For the list, what matters most is what goes with the status.
For "Limited use", that is the condition, in one sentence someone can follow while working. "Only with the business account." "Internal meetings only, and participants are told." "No customer data in the prompt."
For "Not allowed", it is the alternative. A no without an alternative moves the work to a personal account on someone's own phone, and then the tool is not gone but invisible. "Not the free account; use the business licence instead" is a decision someone can follow without stopping their work.
Why is a list without "No decision yet" incomplete?
Because a list of approved tools says nothing about the tools that are not on it, and that is the largest group. AI tools appear faster than an organisation can assess them, and existing software gains AI features without a decision.
Everyone draws their own conclusion about a tool like that. One person reads the list as exhaustive and assumes the rest is forbidden; another reads it as a set of examples and assumes anything not refused is fine. Both readings are reasonable, and the organisation meant neither.
"No decision yet" solves that because it is honest: nobody has looked at the tool yet, so until they do you enter only public information and request an assessment. It is neither a ban nor an approval. It is also your backlog: every tool marked "No decision yet" that is actually in use is an assessment still to be done.
Where do you publish the list?
Where the question comes up, and that is not the intranet. An employee reads the list once and opens a tool hundreds of times. The decision works when it appears as the tool opens: nothing for an allowed tool, the condition for "Limited use", the alternative as the first button for "Not allowed" and, for "No decision yet", a request to enter only public information. Our article on communicating AI policy covers how to do that without burying people in rules.
At the building services firm, the first round of observation shows fourteen AI domains; the administrator had written down four. Alongside Copilot and ChatGPT there are DeepL, Perplexity, the project managers' transcription tool and a tool nobody in the management team had heard of. That is not an incident. It shows that a list from memory is out of date on the day it is finished, and that ten of the fourteen rows start as "No decision yet".
That is how BeeSensible works. The extension reports only the domain of an opened AI tool, at most once per day per device, without a user ID and without content, and sets it against a catalogue of over 1,000 assessed tools. An administrator sees which tools have no decision yet and records the status and the condition; that decision appears in the browser when someone opens the tool, with the allowed alternative as the first button. The product decides nothing itself: the organisation chooses whether an employee may still go on to a tool that is not allowed or is offered only the alternative. There is more on the AI tools page.
