Both EU legislative institutions now support the same basic direction. That matters because it hadn’t been true until March 26. When the European Parliament adopted its formal position on proposed EU AI Act amendments, it joined the Council, which had reached its own position on March 13. The two institutions agreeing doesn’t make the changes law, trilogue negotiations with the Commission are next, but it narrows the range of outcomes significantly.
The core proposal: postpone the application of key obligations for high-risk AI systems from August 2, 2026 to December 2, 2027 for systems in Annex III, and to August 2, 2028 for those in Annex I. These are proposed extensions only. Organizations that have already built compliance programs around the August 2026 date should not assume the extension will materialize on the current schedule. What has changed is that two EU institutions have now formally stated they want the extension. What hasn’t changed is the existing law: the August 2026 deadline remains in force until a final amending text is adopted.
Two additional elements in Parliament’s position deserve attention from compliance teams. First, legal analysis of the Parliament’s position indicates a proposed extension of the watermarking compliance deadline for existing AI systems, covering AI-generated audio, image, video, and text content, to November 2, 2026. This is a distinct timeline from the Annex III/I high-risk postponements and applies to a different set of providers. Second, Parliament’s position reportedly includes a ban on “nudifier apps”, AI tools used to generate non-consensual intimate images. This specific provision should be confirmed against official Parliament documentation before any compliance guidance is based on it; current sourcing is from secondary analysis rather than the Parliament’s formal text.
The readiness picture behind these proposals is stark. According to a 2026 readiness assessment by Vision Compliance, 78% of organizations assessed lacked adequate preparation for EU AI Act obligations, including 83% without a formal AI system inventory, 74% without a designated compliance owner, and 61% without documentation processes in place. These figures come from a single industry report and should be treated as indicative rather than definitive, but they’re consistent with what compliance professionals have been reporting anecdotally.
For context on what the existing law already requires: under the enacted EU AI Act, penalties for non-compliance reach up to €35 million or 7% of total worldwide annual turnover, whichever is higher. Those penalties are in the enacted text. The proposed amendments address timelines, not penalty structures.
Some analysts warn that delaying obligations for high-risk systems while keeping the law non-retroactive could leave some sensitive AI applications permanently outside oversight, a concern about systems already on the market when compliance deadlines arrive. This is a contested inference, not a confirmed regulatory mechanism, but it’s one worth tracking as trilogue progresses.
What to watch: The trilogue negotiation timeline. If the Commission, Council, and Parliament move quickly, a final amending text could arrive before the current August 2026 deadline, giving organizations the clarity they need to plan. A slower process puts compliance teams in an awkward position: prepare for August 2026 and then potentially redo the work, or wait for clarity and risk being caught short if the extension doesn’t materialize in time.