A privacy officer at an accounting firm receives the AI register for review. ChatGPT is not in it, "because we never bought it". Copilot is, as a single row. A few questions later it turns out three kinds of ChatGPT and two kinds of Copilot are in use, and almost nothing has been recorded about any of them.
The answer to the question in the title is yes, and it is shorter than the reason. An AI register describes what is used, not what was purchased. And for these two tools, "ChatGPT" or "Copilot" is not a row but a collective name for three or four situations that differ on the register's own fields.
Why does a tool you never bought belong in the register?
Because the risk does not start with the invoice, it starts with the use. A tool nobody pays for and twenty people open processes the same data as a tool with a contract. The difference is that for the first one, nobody has looked.
The register is where that looking lands. Without a row there is no owner, no decision, and nothing to show employees at the moment they open the tool. That is why a register has four statuses rather than two: a tool that has been noticed but not yet assessed is in it as "no decision yet". That is an honest status, not an approval, and it beats absence. How the statuses work is in four statuses for AI tools.
What exactly counts as self-started AI use, and why it almost always runs through a free account, is in what is shadow AI. For this piece one point is enough: the free account is the default route for use without a purchase, and that makes the question of which account sits behind a tool the core of the row.
Why are free and business separate rows?
Because they differ on exactly the fields a register exists for. Same tool on the screen, different terms underneath; the account type weighs more than the tool.
| Variant | Account | Training on input | Data processing agreement | Reach and connections |
|---|---|---|---|---|
| ChatGPT Free, Plus, Pro | Personal, outside the organisation's administration | Possible, unless the user turns it off | No, consumer terms | Whatever someone pastes or uploads, plus memory, shared links and apps the user connects themselves |
| ChatGPT Team, Enterprise, Edu | The organisation's, with an administrator | Excluded by default | Yes | Whatever someone pastes or uploads, plus apps and connectors the administrator allows |
| Consumer Copilot | Personal, no work account | No enterprise data protection | No | The open web and whatever someone pastes |
| Microsoft 365 Copilot | Work account, paid add-on | Not for the underlying models | Yes, under the Microsoft terms | Mail, files, calendar, Teams and SharePoint, with the user's permissions |
There is a third Copilot, Copilot Chat with a work account: enterprise data protection, but no automatic access to mailbox and files. In the register that is its own row, between the two in the table. The details per variant are in Microsoft Copilot inherits your access and ChatGPT and work data; this piece only uses what was established there.
Read the table by column and you see why one row cannot work. The free row gets a different decision from the business row, because the risk differs. The free row needs an alternative (the business one), the business row a condition (which data may and may not go in). And the review runs differently: consumer terms change without the organisation receiving a letter.
Which fields differ for these two compared with other tools?
For most tools in a register, vendor, purpose and owner are most of the work. For ChatGPT and Copilot, the differences sit in four other fields.
Account. Who owns the sign-in? With a personal account, the chat history, the memory and every shared link sit with the employee, and all of it walks out of the door when they leave. The organisation has no hand in it. Record in the row whether the account belongs to the organisation and, if not, the route to the business variant.
Training. With personal accounts it is a setting the user has to switch off; with business variants, input is outside training by default. Record which of the two applies, and keep it separate from storage: not training says nothing about how long a prompt is kept. That distinction is worked out in what "no training on your data" really means.
Data processing agreement. If the tool processes personal data on behalf of the organisation, Article 28 GDPR calls for a data processing agreement, and that exists only on the business variants. For the free row the field is therefore empty, and that is a fact that steers the decision. When the agreement is needed is in do you need a data processing agreement for an AI tool.
Reach and connections. This field is missing from most registers and for Copilot it is the most important one. ChatGPT sees what someone pastes, uploads or connects through an app. Microsoft 365 Copilot sees everything the user is allowed to open, including the folder that was opened to "everyone" three years ago. The risk then sits not in the tool but in the permission structure, and the register should point to where that is being cleaned up.
What does this look like in the register?
Back to the accounting firm. After an afternoon's work, the single Copilot row has become three and the missing ChatGPT has become two.
Microsoft 365 Copilot is allowed for the forty licences, owned by the head of operations, with a reference to the project reviewing SharePoint permissions. Copilot Chat with a work account is allowed. Consumer Copilot is not allowed, with Copilot Chat as the alternative. ChatGPT Team, in use by the advisory team, is limited use with one condition: no client data. And free ChatGPT is on no decision yet, until the advisory team's owner has decided whether Team becomes the route for everyone.
None of that is an incident. The register now describes what happens, and every row says who is responsible for it.
One limitation belongs here. An observation of use tells you which domain was opened, not which account: chatgpt.com is chatgpt.com. The account type shows in licences and administration, and in the agreement the organisation makes. In BeeSensible's catalogue, the consumer and the business tier of the same tool are separate entries with their own risk score, and the free tier points to the business tier as the safer alternative. The organisation decides which status each row gets; the extension reports only the domain of an opened tool, without a user id. More on the AI tools page.
