Most real AI deployments, from a CV-screening tool to a customer-service chatbot, now answer to two rulebooks at once: the GDPR and the EU AI Act, Regulation (EU) 2024/1689. They are separate laws with separate enforcers, and meeting one does not discharge the other. This article maps where they overlap, what the 2026 AI Omnibus changed, and how a small or mid-sized business can build one programme instead of two. It is general information, not legal advice.
The calendar moved in 2026. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026, according to Praxikon and Usercentrics. It pushed back the main high-risk deadlines, and Usercentrics notes that it did not change the lawful-basis requirements for processing personal data. The AI Act's calendar slipped; the GDPR duties that already apply to AI did not. Teams that read the delay as breathing room risk a gap in exactly the place regulators are already looking.
Where the two regimes meet
Four areas show the overlap most clearly. In each, meeting the AI Act requirement does not discharge the GDPR one, and the reverse is also true.
| Topic | GDPR | EU AI Act |
|---|---|---|
| Lawful basis and purpose | Art. 5 and 6: every processing activity needs a lawful basis and a specified purpose, including training and deployment | Applies without prejudice to the GDPR; no lawful basis of its own |
| Transparency | Art. 13 and 14: tell people what is processed and why | Art. 50: disclose AI interaction and mark synthetic content; instructions for use for high-risk systems |
| Automated decisions | Art. 22: limits on solely automated decisions with legal or similarly significant effects | Art. 14: effective human oversight of high-risk systems |
| Data quality and sensitive data | Art. 5(1)(c) and (d), Art. 9: minimisation, accuracy, special-category limits | Art. 10: data governance; Art. 10(5) permits special-category data for bias detection |
The most contested question is the lawful basis for training. In Opinion 28/2024, the European Data Protection Board (EDPB) said legitimate interest is not a default basis for developing or using AI models. The controller must demonstrate a specific interest, necessity and a balance that holds against data subjects' reasonable expectations, as CMS summarises it. We cover the enforcement side in GDPR Is Already Fining AI Mistakes.
The Omnibus also widened the special-category basis in Article 10(5) from high-risk providers to all AI systems and general-purpose models, under a strict-necessity standard, according to Gibson Dunn and Orrick. Orrick stresses that it creates no duty to perform bias detection and requires safeguards such as pseudonymisation and deletion. Read the final text before relying on it.
Impact assessments: one exercise, three lenses
The GDPR requires a data protection impact assessment (DPIA) under Article 35 where processing is likely to result in a high risk to individuals. Systematic profiling and automated decision-making are the textbook triggers, so many AI deployments need a DPIA already (see our DPIA article for AI chatbots).
The AI Act asks for related but different work. Providers of high-risk systems must run a continuous risk management system under Article 9. Certain deployers, including public bodies and private entities providing public services, must carry out a fundamental rights impact assessment (FRIA) under Article 27 before first use. The Act links the exercises: deployers use the provider's instructions for use to support their DPIA (Article 26(9)), and a FRIA can complement a DPIA that already covers the same ground (Article 27(4)).
The difference lies in the question each one asks. A DPIA asks what could go wrong for people's privacy and data rights. A risk management system asks whether the AI system is safe and performs as intended across its lifecycle. A FRIA asks which fundamental rights, such as non-discrimination, could be affected for the groups the system touches.
A single, well-structured assessment can answer all three. Document the purpose, data flows and lawful basis once. Then add a section on accuracy and bias testing, a section on affected groups, and a section on human oversight and residual risk. Review it at each material model change, because the GDPR treats a DPIA as a living document and the AI Act expects risk management to run for the whole lifecycle.
The AI Act timeline after the Omnibus
Dates drive priorities, so the table sets out where things stand in November 2026. The Omnibus moved the high-risk obligations and kept the earlier milestones. Sources differ on some details of Article 50, so confirm the final dates in Regulation (EU) 2026/1744 before fixing project plans. For a fuller walk-through, see our 2026 to 2028 compliance calendar.
| Obligation | Applies from | Note |
|---|---|---|
| Prohibited practices | 2 Feb 2025 | Already in force |
| AI literacy (Art. 4) | 2 Feb 2025 | Softened: measures to support literacy, with no guaranteed level per individual |
| General-purpose AI model duties | 2 Aug 2025 | Unchanged |
| Transparency duties (Art. 50) | 2 Aug 2026 | Marking of AI-generated content for generative systems already on the market gets until 2 Dec 2026 |
| New bans on non-consensual intimate imagery and child sexual abuse material | 2 Dec 2026 | Added by the Omnibus |
| Stand-alone high-risk systems (Annex III) | 2 Dec 2027 | Previously 2 Aug 2026 |
| High-risk systems in regulated products (Annex I) | 2 Aug 2028 | Previously 2 Aug 2027 |
| High-risk systems used by public authorities | 2 Aug 2030 | Per Orrick and Praxikon |
Three points stand out. First, the Article 50 transparency duties are already live for most systems and overlap directly with GDPR notice duties. Second, small mid-caps, meaning companies with fewer than 750 staff and turnover up to EUR 150m or a balance sheet up to EUR 129m, gain simplified documentation and tailored penalty caps, per Orrick. Third, the AI Act's maximum fines stay at EUR 35m or 7% of global turnover for the most serious breaches (prohibited practices), according to Usercentrics, against the GDPR's ceiling of EUR 20m or 4%.
A practical integrated approach
An integrated programme avoids doing the same work twice. A workable sequence:
- Build one inventory. List every AI system and AI-enabled feature with its purpose, data categories and vendor, and link each to its GDPR record of processing.
- Assign roles twice. For each system, decide whether you are a controller or processor under the GDPR and a provider or deployer under the AI Act. Duties differ by role.
- Classify risk once. Screen each system against the Act's prohibited and high-risk categories and against the DPIA triggers in the same pass.
- Run a combined assessment for high-risk or high-impact uses, as described above, and keep it under version control.
- Align notices. Update privacy notices and in-product disclosures together, so that Articles 13, 14 and 50 tell one consistent story.
- Design real human oversight. Where decisions affect people, make sure the reviewer can understand, question and override the output. That supports both Article 22 GDPR and Article 14 of the AI Act.
- Tighten contracts. Require vendors to supply instructions for use, training-data provenance statements and cooperation on incidents.
The bottom line
The Omnibus bought time for the AI Act, not for the GDPR. Organisations that use that time to build a single inventory, a single assessment method and a single set of notices will find that the 2027 and 2028 deadlines add little new work. Those that wait will meet those deadlines while still holding unresolved GDPR questions about training data and automated decisions, which are the questions regulators are asking today.
A good first step is to see where your own AI tools stand against the GDPR. GDPRGard's audit tool is a quick place to start.
Sources
- Praxikon: Digital Omnibus and the postponement of high-risk obligations to December 2027
- Usercentrics: EU AI Act deal, Digital Omnibus now in force
- Orrick: EU AI Act update, Digital Omnibus finalises 8 compliance changes
- Gibson Dunn: EU AI Act Omnibus agreement
- Stephenson Harwood: EU Digital Omnibus deal
- CMS: EDPB Opinion 28/2024, key takeaways
- EU AI Act, Regulation (EU) 2024/1689, EUR-Lex
- GDPR, Regulation (EU) 2016/679, EUR-Lex
The positions above are summarised from law-firm and industry commentary; the text of Regulation (EU) 2026/1744 itself was not reviewed for this article. Cross-references to articles of the AI Act and the GDPR, in particular Articles 26(9) and 27(4) of the AI Act, should be checked against EUR-Lex before you rely on them. Analysis of how the two regimes fit together is ours.
Also read:
- The high-risk deadline moved to December 2027. The GDPR overlap didn't move at all.
- Forget the AI Act β GDPR is already fining companies for AI mistakes
- The AI Act Was Delayed, Not Cancelled: Your 2026 to 2028 Compliance Calendar