A privacy officer at a housing association gets two questions about the same tool in one week. Marketing wants to use ChatGPT to make job adverts read better. HR wants to use it to pre-sort the fifty applications for a vacancy, which saves a day of reading. Both teams expect the same answer, since it is the same tool.
The answer is different twice over, and that is not inconsistency. One AI tool can be low and high risk at the same time, because the risk does not attach to the tool but to the use, the data and the account. The AI Act does exactly the same: it classifies applications by their purpose, not tools by their name.
Why does the risk not attach to the tool?
An assessment of an AI tool has two halves. The first is about the vendor and the product: what happens to input, whether it is trained on, where processing takes place, whether a data processing agreement exists. That half is the same for every department. The second half is about the organisation: what the team will use the tool for and which personal data comes with that. That half differs per request, and it is the one most often skipped. How those dimensions relate is set out in how do you assess the risk of an AI tool.
Do only the first half and you end up with one verdict per tool. Do both and you end up with one verdict per use.
What changes with the purpose?
This is where the AI Act comes in, and the phrase that carries everything is "intended purpose". The regulation calls an AI system high-risk when the purpose it is intended for is listed in Annex III. Point 4 of that annex covers employment: filtering applications, evaluating candidates, deciding on promotion or dismissal, tracking performance. Drafting a job advert is not on it.
For the housing association that means marketing's request does not touch Annex III. HR's request does, with a nuance. A general chatbot is not intended by its provider as a selection system, so a recruiter who pastes individual CVs into it is not using a high-risk system in the sense of the law. If the organisation builds the model into the recruitment process and lets it decide who gets an interview, it has created a high-risk system itself and moves from deployer to provider. What that brings with it is in the EU AI Act for employers.
That nuance is not a free pass. Either way the recruiter is making a decision about people with a tool that was not assessed for it, and the GDPR and the organisation's own policy apply.
What changes with the data?
The GDPR asks the question from the other side: not what the system is, but which personal data is processed and on what basis. A job advert contains none. Fifty applications contain names, addresses, dates of birth, work histories and sometimes a line about health or family. What comes into play then, from lawful basis to human involvement, is in using AI safely in HR.
Same tool, two processing operations. For the second you want to know whether a data processing agreement exists, whether input is retained and whether it is trained on. Which brings in the account.
What changes with the account?
The recruiter uses ChatGPT through a free account that was already on her laptop. Marketing has a business licence through the department. That is not a detail: the free account runs under consumer terms, without a data processing agreement, and at most vendors with training on input as the default unless the user turns it off. The business account arranges all three differently. Why the line between account types weighs more than the line between brands is in free versus business AI accounts.
For classification under the AI Act the account makes no difference; the purpose decides. For the risk under the GDPR it is the difference between a processing operation you can account for and one you cannot. That is why free and business versions of the same tool sit as separate rows in a register.
What does that look like for one tool?
Put the three variables side by side and one tool splits into a few rows.
| Use | Data | Account | AI Act | GDPR | Decision |
|---|---|---|---|---|---|
| Drafting a job advert | No personal data | Business licence | Not high-risk | No processing of personal data | Allowed |
| Summarising cover letters | Candidates' personal data | Free account | Not high-risk, but a decision about people | Processing without a data processing agreement | Not allowed through this account |
| Ranking candidates as a fixed step in selection | Personal data, possibly special categories | Business licence | Becomes high-risk once built into the process | High-risk processing, DPIA in scope | No decision yet, assess first |
The table answers the question in the title. The first row is low risk, the third is high, and it is the same tool. A register with a single row for ChatGPT has one of the three wrong.
How did it go on at the housing association?
The privacy officer gives marketing the go-ahead, with the business licence as the condition. HR gets a different answer: summarising applications is not allowed through the free account, and ranking candidates is first assessed as a separate application, with the question of who holds oversight and what the vendor has documented about how it works. Until then the status stays "no decision yet".
While working that out, an inventory shows that alongside Copilot, which was already allowed, Perplexity, ChatGPT and three specialised tools are in use, two of which are not yet in the register. That is not an incident. It shows that a one-off inventory falls behind quickly, and that the conversation about one tool usually starts with which tools there are at all.
BeeSensible links a catalogue of AI tools to the use observed in the organisation: only the domain, at most once a day per device, with no user id. Free and business versions of the same tool sit in it as separate rows, each with its own risk score and an EU AI Act indication. That score is advice about the tool; a catalogue cannot fill in the use and the data, which is why the organisation sets the status itself, with a condition under "limited use" that marks out the use. How that works is on the AI tools page.
