On 21 October 2026, Coreper, the committee of the member states' permanent representatives, agreed the Council's negotiating mandate for the GDPR part of the Digital Omnibus. If you're planning AI projects on your own customer data for 2027, that mandate sends a different signal from the one the Commission sent a year earlier.
The Digital Omnibus GDPR is the EU legislative proposal that amends parts of the GDPR, including the definition of personal data, the rules for AI training and the data breach notification duty, and it is not law yet. In its mandate the Council deletes the relative definition of personal data the Commission proposed, and the European Parliament has no position yet. Until there's a deal and the text is published in the Official Journal, the GDPR applies exactly as you know it.
What did the Commission propose, and what does the Council want?
The European Commission published the proposal on 19 November 2025 as COM(2025) 837, procedure 2025/0360(COD). Article 3 of that proposal amends the GDPR. The Council set out its own version in Council document 14756/26, dated 22 October 2026, with its changes to the Commission text in bold or struck through. The 'GDPR today' column follows the current text of the GDPR (Regulation (EU) 2016/679).
| Topic | GDPR today | Commission (Nov 2025) | Council (Oct 2026) |
|---|---|---|---|
| Definition of personal data (Art. 4(1)) | Any information on an identifiable person | Relative: not personal for an entity that cannot identify the person | Deleted, definition unchanged |
| Pseudonymised data | Remains personal data | Commission decides by implementing act when it no longer is | New Art. 25a: not personal for a third party that cannot identify, still personal for a processor |
| AI development on legitimate interest | Possible under Art. 6(1)(f), with a balancing test | New Art. 88c, with an unconditional right to object | Art. 88 bis, with the ordinary Art. 21 right to object |
| Special category data in AI | Only with an Art. 9(2) exception | New exception for AI development and operation | Only unintended residual data; erase without undue delay |
| Breach notification (Art. 33) | 72 hours, unless no risk | 96 hours, high-risk breaches only | 96 hours, high-risk breaches only, with a national entry point |
| Who sets the criteria | Supervisory authorities and the EDPB | The Commission, by implementing act | The EDPB, by opinion |
The pattern is clear. Where the Commission shrank the GDPR's reach, the Council mostly pulls it back. Where the issue is administrative burden, such as breach reporting, the Council follows the Commission.
Does the definition of personal data change?
In the Council's text, the GDPR definition of personal data stays as it is. The Commission wanted to add three sentences to Article 4(1), the core being that information isn't personal for an organisation that can't identify the data subject with means reasonably likely to be used. The Council struck those sentences from its mandate entirely, along with the matching recital 27.
What replaces them is a narrower rule that only covers pseudonymised data. The new Article 25a in the Council text says pseudonymised data isn't personal data for a third party that can't identify the data subject. If that third party passes the data on to someone who can, both the transfer and the onward processing are processing of personal data. Paragraph 4 then closes the obvious loophole: data a processor handles on behalf of a controller remains personal data for that processor.
That last point lands on every AI project that uses a cloud provider. Most language model vendors act as processors. Under the Commission text you could argue that a vendor seeing only pseudonymised prompts wasn't processing personal data. Under the Council text you can't. We've covered how the Court of Justice already assesses identifiability per party in EDPS v SRB, and how to build a local pseudonymisation filter, so we won't repeat that here.
Training AI on your own data: what changes under the Council text?
AI development based on legitimate interest stays possible in both texts, but the Council doesn't turn it into a separate free pass. The Commission proposed an Article 88c that expressly placed processing for the development and operation of an AI system or AI model under legitimate interest, with a list of safeguards and an unconditional right for data subjects to object.
The Council renames it Article 88 bis and strips it back. Paragraph 1 now only says such processing may be carried out on a legitimate interest under Article 6(1)(f). Paragraph 2 adds that choosing the right lawful ground remains your job as controller. Paragraph 3 requires appropriate safeguards whatever the ground, "including the right to object under Article 21". That's the GDPR's ordinary right to object, where you can carry on if you demonstrate compelling legitimate grounds. The Commission's unconditional opt-out is gone.
For special category data, such as health or ethnic origin, the gap is wider. The Commission wanted a new exception (k) in Article 9(2) for processing "in the context of the development and operation of an AI system". The Council limits it to "incidental processing of residual data, where not intended by the controller". If you find such data in your training, test or validation set, you must erase it without undue delay. If that's technically impossible, or needs manifestly disproportionate effort, you must shield it from further processing, from use in outputs and from disclosure.
In practice: a company that wants to fine-tune a model on customer files containing health data gets no new legal basis from the Council. The exception only covers what slips through despite your filter. Whether training your own AI model makes sense still depends on getting special category data out beforehand.
Don't build a business case on the Commission's version
The Commission text is the most permissive of the three, and it's the one least likely to become law unchanged. A plan that only works if pseudonymised data stops being personal data for your AI vendor, or if AI development gets its own exception for special category data, rests on wording the member states have rejected.
| Plan for 2027 | Safe to build on | Don't build on |
|---|---|---|
| RAG over customer data via a cloud model | Today's GDPR: processing agreement, DPIA, transfer rules | Pseudonymised prompts falling outside the GDPR |
| Fine-tuning on your own data | Legitimate interest with a balancing test, as today | An AI exception for special category data |
| Sharing a pseudonymised dataset with a partner | The C-413/23 P line, assessed per recipient | An implementing act deciding it for you |
| Incident procedure | 72-hour notice to your supervisory authority | The 96-hour limit or the high-risk threshold |
What is safe: a pseudonymisation filter in front of every external call, a fixed step that removes special category data from training sets, a documented erasure procedure and a balancing test per AI use case. You need those measures today, and you'll need them under all three texts. They belong in your DPIA for the AI project, not in a separate omnibus file.
Time saved
Save 4 hours per week on working out which GDPR rules apply to each AI project, now and later
When will the Digital Omnibus GDPR apply?
There's no date. The European Parliament has no position: according to the Parliament's Legislative Observatory, the ITRE and LIBE committees handle the file jointly, the draft report was tabled on 22 June 2026 and amendments followed on 27 July 2026. On 12 November 2026 the status read "Awaiting committee decision". After that come a vote, trilogue negotiations between Council, Parliament and Commission, and formal adoption.
The final provisions in both texts say the regulation enters into force on the third day after publication in the Official Journal. For most of the GDPR amendments, that's also when they start to apply. The new Article 33 breach rule is an exception: in both texts it applies 18 months after entry into force. Neither text yields a calendar date, because everything depends on when Parliament and Council agree. For comparison, the AI Act part of the same omnibus was split off and finished faster, as we described in the AI Act deadline delay. Nothing is settled for the GDPR part.
Two follow-ups are already written into the Council mandate. The EDPB must issue an opinion on pseudonymisation and anonymisation, requested within six months of entry into force, and publish a list of breaches that do and don't result in high risk within nine months of the rules applying. The new text only becomes workable in practice once those documents exist.
Who needs to act now?
A company of 50 to 500 people planning AI on customer data for 2027 has no reason to wait. Start from today's GDPR, choose a set-up where personal data doesn't leave your own environment where you can, and treat the omnibus as a possible easing of breach reporting, not an exemption for AI. The wider rules around AI in the Netherlands, including the AI Act, are covered in our guide to AI legislation in the Netherlands.
If you'd like your AI plan tested against both the current GDPR and the Council text, or want to keep customer data inside your own infrastructure with local AI, we can work that through as part of AI consulting.