ChatGPT, Claude, Copilot and Gemini are not allowed at work if you are going to put personal data in them. Unless you do it this way.

1. The short answer

Never put personal data of the company, of clients or of staff into a personal AI account (including paid ones: Pro, Plus, Max or Go). With a business plan, training turned off, people informed and a short list of approved uses, you can use these tools with peace of mind and without asking each person for consent.

2. The account matters more than the tool

The same tool changes its legal nature depending on the account. With a personal account (including paid ones: Pro, Plus, Max or Go), the provider processes your data as a controller and under its own terms. With a business plan, the data processing agreement is part of its terms, which is what Article 28 GDPR requires when someone processes data on your behalf.

I have read the public terms of the four providers on 29 October 2026. This is what they say, in their own words.

3. Where your data goes and how long it stays

Signing up for a business plan settles the processing agreement. On its own it does not settle where the data is processed or how long it is kept. Four things worth knowing before you sign.

4. The 12 things you cannot do (and what to do instead)

Each rule carries its label. "Law" is an obligation with an article behind it. "Recommendation" is professional judgement on how to comply without drama. First what not to do, then the alternative.

1. 1

   Using personal accounts with company or client data

   Law
   No

   Pasting a contract, a client list or a medical report into ChatGPT Plus, Claude Pro, a personal Copilot or a personal Gemini, even if the company pays for the account.

   Instead

   A business plan with a data processing agreement: Claude Team, Enterprise or API, ChatGPT Business, Enterprise or API, Copilot with a work account or Gemini in an eligible Workspace edition. And, as an organisational measure I recommend, blocking personal accounts on corporate devices.

   Basis Art. 28 GDPR. In consumer Claude for the EEA, commercial use is also prohibited by its terms (section 11).

2. 2

   Uploading data that requires special safeguards to a general-purpose assistant

   Recommendation
   No

   Passwords, keys and tokens. Data on health, ideology, trade union membership, sexual orientation or biometrics. Criminal data. Data you process on behalf of a client without their authorisation for that provider. Information protected by a confidentiality agreement.

   Instead

   An internal list of banned data, by type of data and not by file name. Content is what counts: a quote can contain a diagnosis. Before uploading, trim what adds nothing.

   Basis Arts. 9 and 10 GDPR for special categories and criminal data, which may only be processed under an exception and with safeguards. The GDPR has no "list of data banned from AI": this list is a recommendation built on the requirements of each type of data.

3. 3

   Adding the AI provider as a sub-processor without your client knowing

   Law
   No

   Processing with AI data that you handle on behalf of a client without their prior written authorisation for that provider.

   Instead

   If your data processing agreement contains a general authorisation, notify the client beforehand and give them the chance to object. If it is specific, request written authorisation. The sub-processor takes on the same obligations as you.

   Basis Arts. 28.2 and 28.4 GDPR. AEPD guidelines on contracts between controllers and processors, section 11.

4. 4

   Believing that removing the name is anonymising

   Law
   No

   Deleting the name and leaving the job title, the town, the date of the accident and the amount. That person is still identifiable.

   Instead

   Ask yourself whether someone, with what remains and what they already know, can get back to the person. If the answer is yes, it is still personal data and is treated as such. For most tasks, uploading only what is needed is enough.

   Basis Art. 4.1 and Recital 26 GDPR: data relating to a person who is identifiable "directly or indirectly" is personal data.

5. 5

   Processing data with AI without saying so in your privacy information

   Law
   No

   Using AI providers as data recipients without them appearing, at least by category, in the information you give to clients, candidates and staff.

   Instead

   Update the information: recipients or categories of recipients (technology and AI providers), transfers outside the EEA with their safeguard and how to obtain a copy. Categories are fine, as specific as possible, but have the actual providers identified in case someone exercises their right of access. If the change is relevant to people who have already given their data, communicate it before the new use through a channel that reaches them. No mass mailing is needed.

   Basis Arts. 13.1.e and f and 14.1.e and f GDPR. Time limits in Art. 14.3: one month at most, or earlier if you communicate with the person or disclose their data. Transparency guidelines WP260 rev.01, annex.

6. 6

   Letting AI alone decide something with legal or similarly significant effects on a person

   Law
   No

   Rejecting a candidate, refusing a loan or closing an account based solely on a model's output. A box-ticking "human review", where a person signs off whatever the machine says, does not count.

   Instead

   Article 22 applies when the decision produces legal effects or similarly significantly affects the person: rejecting, refusing, dismissing, closing. Sorting emails or proposing a draft does not. In those cases, let a person with judgement and real power to change the outcome decide. If you genuinely need an automated decision, it is only allowed in three cases: necessary to enter into or perform a contract with that person, authorised by a law with safeguards, or with explicit consent. For contract and consent, guarantee human intervention, the right to express their point of view and the right to contest the decision. If the AI builds profiles, inform people about it even if a person decides.

   Basis Art. 22 GDPR. Art. 11.2 LOPDGDD for information about profiling.

7. 7

   Crossing the red lines of the AI Act

   Law
   No

   Using AI to infer the emotions of employees or students from biometric data, for example by analysing voice or face in meetings or classrooms.

   Instead

   That tool does not come into the company, except for medical or safety reasons. The other prohibited practices in Article 5 (biometric categorisation to infer sensitive data, social scoring, subliminal manipulation) are not bought "to try out" either.

   Basis Art. 5.1.f of Regulation (EU) 2024/1689, applicable since 2 February 2025.

8. 8

   Leaving staff without AI training

   Law
   No

   Giving access to Copilot or ChatGPT and trusting that everyone "already knows".

   Instead

   AI literacy support measures tailored to what each person uses: a short session, a one-page guide, the tool catalogue with its rules. It does not require an accredited course or a certificate. Keep a record of what was done.

   Basis Art. 4 of the AI Act as worded by Regulation (EU) 2026/1744, applicable since 2 February 2025. The new text clarifies that the obligation "does not require" guaranteeing a specific level for any particular person.

9. 9

   Publishing artificial content without saying so when the law requires it

   Law
   No

   Distributing an image, audio or video generated or manipulated by AI that looks authentic (a deepfake) without disclosing that it is artificial. Or publishing an AI-generated text to inform the public on a matter of public interest without human review or an editorial owner.

   Instead

   Always disclose that deepfakes are artificial. In artistic, creative, satirical or fictional works, in a way that does not hamper the work. For public-interest texts, either label them or have someone review them and take editorial responsibility. If you offer a chatbot under your brand, make sure the user knows they are talking to an AI, unless it is obvious. There is no need to label every email or every post made with AI.

   Basis Arts. 3.60, 50.1 and 50.4 of the AI Act, applicable from 2 August 2026. Art. 3.3: whoever puts the system into service under their own name or trademark may be the "provider" even if they outsource the development.

10. 10

    Using AI to recruit or evaluate staff as if it were just another filter

    Law
    No

    Screening CVs, scoring interviews or deciding promotions with AI without treating it as what it normally is: a high-risk use.

    Instead

    Already today: inform the workers' legal representatives of the parameters, rules and instructions of the algorithm that influences decisions on working conditions or on access to and retention of employment. That does not replace informing each person. And plan for the high-risk obligations, which generally arrive on 2 December 2027, including informing the affected workers and their representatives beforehand.

    Basis Art. 64.4.d of the Workers' Statute, already in force. Annex III and Art. 26.7 of the AI Act, with the dates set by Regulation (EU) 2026/1744.

11. 11

    Ignoring the small print of each licence

    Recommendation
    No

    Signing up for the "Business" plan and assuming it covers health data, processes in Europe or includes every module.

    Instead

    Read three things before signing: restricted uses (some licences limit medical uses or uses with health data), processing region and its price (processing in a single region usually costs more) and which products are left out (in Workspace, Gemini Notebook is not covered by data regions).

    Basis Professional judgement, based on the terms cited in section 3 (consulted on 29/10/2026).

12. 12

    Reusing with AI what comes in through your forms without notice

    Law
    No

    Running contact form messages, job applications or support tickets through an AI assistant with a privacy policy that does not mention it.

    Instead

    Adapt the policy first: recipients, transfers and, where applicable, the use of AI in that processing. It is a half-hour change that avoids uninformed processing.

    Basis Arts. 13 and 14 GDPR. Art. 11 LOPDGDD for layered information.

5. What you can do: 6 myths that hold you back for no reason

Many of the companies that consult me are afraid of things nobody requires. These six are the most common.

6. Get organised in an afternoon: the AI tool catalogue

The law does not talk about a "catalogue". It is my recommendation for turning all of the above into something staff can comply with without reading the GDPR. A table, communicated with acknowledgement of receipt, which can sit alongside the rules on the use of digital devices. That way you organise the measures, train each person in what they use, avoid personal accounts and can prove what has been done.

Example catalogue. The rows are adapted to each company.

Minimum configuration I always ask for

• Training turned off and checked in the admin console, not just in one user's profile.
• Response rating button turned off in Claude. That button allows the conversation to be kept for up to 5 years and used for training.
• Retention periods set for chats, files and projects, and one person in charge of making sure they are met.
• Copilot flex routing reviewed if any contract requires processing in the EU.
• Personal accounts blocked on corporate devices, including the Microsoft policy "Multiple account access to Copilot for work documents".

7. What already applies and what is coming

Almost none of this is in the future. The GDPR has applied since 25 May 2018 and the AI Act already governs everything this article covers, except two dates still to come.