EU AI Act omnibus: what is delayed, what is not

On 7 May 2026, the Council of the EU and the European Parliament reached a provisional agreement on the Digital Omnibus{target="_blank" rel="noopener"} to push the EU AI Act's high-risk obligations under Annex III to 2 December 2027 — a 16-month delay from the original 2 August 2026 deadline. The headlines were immediate, and inside enterprise procurement they read like a reprieve. Several voice AI projects we are advising on this month came back with a single question: do we still need to do this now?

The honest answer, for anyone deploying voice AI into the EU market, is yes. The omnibus delayed the heaviest part of the regime. It did not delay the part that hits voice AI deployments first. Article 50 — the transparency layer that compels you to disclose the AI nature of a voice interaction and to mark synthetic audio — is still on track. It applies from 2 August 2026, with a final grandfathering cliff on 2 December 2026 for any system placed on the EU market before August.

This post is the corrective. We separate what the omnibus moved from what it left alone, and we translate that into concrete actions for enterprise voice AI buyers — the kind we run through the AI placement diagnostic before any deployment commitment.

This guide is shipped by the team behind Dilr Voice — enterprise voice AI live in 40+ countries with Article 50 disclosure primitives shipped at the runtime layer. Or see DATS, our five-stage AI consulting methodology.

For the structural reading of which voice AI use cases fall in scope at all, our earlier piece on the EU AI Act and voice AI obligations still holds — the omnibus changed the calendar, not the categories. What changed is the sequencing of risk: Article 50 now sits in front of Annex III, not alongside it.

What the omnibus actually moved

The Digital Omnibus is a simplification package, not a rewrite. The Council press release on 7 May 2026 frames it as a recalibration of timing and burden, not of scope. Travers Smith's analysis{target="_blank" rel="noopener"} is the clearest practitioner summary in English of the dates and what survived. Three structural moves matter for voice AI buyers.

Move 1 — Annex III high-risk delayed to 2 December 2027

Standalone high-risk AI systems classified under Annex III now have compliance dates pushed to 2 December 2027. For voice AI, this affects deployments in essential services (utilities triage, banking and insurance customer-facing voice), employment (interview triage by voice), education (admissions and assessment), critical infrastructure operations and biometric voice categorisation. The conformity assessment, technical file, CE marking, EU database registration and post-market monitoring obligations are all on this delayed clock.

For voice AI embedded inside Annex I regulated products — for example medical devices with a voice interface — the deadline moves to 2 August 2028.

Move 2 — Article 50 transparency held its line

The omnibus left Article 50 on its original schedule: in-force 2 August 2026, with a grandfathering window closing on 2 December 2026 for any system already on the EU market before August. After that date there is no carve-out, no transition, no exemption for legacy deployments. Every voice AI interaction with an EU user must, from 2 December 2026, disclose its non-human nature and mark synthetic audio output in a machine-readable way. The Dilr Voice enterprise platform ships these primitives at the runtime layer — disclosure script with audit log, watermarked TTS output, deployer notification toggle — so the obligation is configuration, not engineering.

We covered the operational depth of this in our EU AI Act Article 50 voice AI disclosure compliance guide. The omnibus does not weaken any of it. If anything, it raises the stakes by clarifying that Article 50 is the regime that will be enforced first.

Move 3 — Article 5 prohibitions strengthened, not loosened

The omnibus adds new prohibited categories rather than removing any. Voice AI use cases that rely on deceptive impersonation of a human, manipulative emotional inducement, or exploitation of vulnerable users remain prohibited from the existing in-force date and now sit alongside additional prohibitions on certain synthetic media. None of this was delayed.

The right way to read the omnibus is as a re-sequencing. A year ago, enterprise voice AI buyers could plausibly bundle Article 50 disclosure work into a larger Annex III conformity programme arriving in August 2026. That bundle is now broken. The transparency programme runs ahead — and on its own clock.

What this means for voice AI buyers right now

If you are inside an EU-facing voice AI deployment, the omnibus changes your prioritisation but not your spend trajectory. The compliance work survives; the order changes. Three operational consequences follow, and each maps to a concrete decision.

Consequence 1 — Article 50 controls become a 2026 capex item. Disclosure logic, synthetic-audio marking, deployer-side notification flows, and the human-AI distinction in conversation design all need to be live well before 2 December 2026. That means contracted, tested and in production by Q3 2026 — not scheduled into a 2027 Annex III workstream.

For voice AI specifically the controls cluster around four artefacts: (1) an opening disclosure script with audit-trail evidence, (2) machine-readable marking on TTS output, (3) a deployer-facing notification toggle, and (4) a retention regime aligned with your voice AI data retention obligations under GDPR. None of these are heavy engineering, but all four need governance owners — exactly the kind of decision the AI execution office engagement is designed to ratify. The teams that delay this on the basis of headline reprieve will discover in Q4 2026 that the runtime work itself is fine; the documentation, RACI and supervisory-authority correspondence are what take time.

Consequence 2 — Annex III preparation is not optional, it is sequenced. Pushing Annex III to December 2027 does not remove the gating events: conformity assessment, technical documentation, CE marking, EU database registration. For a voice AI system used in essential services or employment screening, the lead time on these is 9–12 months. If you are buying or building in that scope today, your internal timeline still says 2026 — you are now working to a Q3 2027 deadline for the documentation pack, not a 2026 panic.

This is also where the omnibus quietly hardens the bake-off. With more time, supervisory authorities will issue more guidance and procurement teams will get sharper. Vendor selection now needs to anchor against a 2027 evidence file — the kind covered in our enterprise voice AI vendor evaluation checklist. The bar will be higher when you arrive.

Consequence 3 — Penalty exposure on Article 50 is real and immediate. Under Article 99, breach of Article 50 transparency obligations carries an administrative fine of up to €15 million or 3% of worldwide annual turnover, whichever is higher. This is the medium fine tier — distinct from the €35M / 7% prohibited-AI tier in Article 5. It is large enough to materially exceed the cost of any voice AI programme.

For a UK enterprise serving EU customers, the relevant overlay is the parallel UK regime. The ICO AI Code of Practice obligations from May 2026 and the FCA AI governance expectations for voice AI operate on a similar clock for UK-side risk. Buyers running pan-European voice deployments are not freed from documentation work by the omnibus — they are freed from doing the heaviest part of it in the same quarter as the transparency cut-over.

Is this obligation delayed? A decision tree. The tree below collapses the omnibus into a single procurement-ready read for any voice AI obligation in scope.

What every voice AI deployment needs in the file by 2 December 2026. The discipline we recommend to every client running an EU-facing voice AI programme is to assemble a single evidence file by Q3 2026 — modelled on the AI tool inventory ICO, FCA and EU AI Act all want, but tuned to Article 50 disclosure as the primary obligation. Audited disclosure logs, marked-output samples, a deployer notification policy, and a fall-back transcript of the opening seconds of every call. None of this is delayed. All of it underwrites the case that you took transparency seriously when the supervisory authority arrives.

Action map — what voice AI buyers must do, by obligation. The table below condenses every operative obligation into a single procurement view.

If you want to talk through which row applies to a specific deployment, book a 30-minute scoping call — no deck, confidential.

Sequencing matters. The buyers we see making the right call this month are treating Article 50 as a 2026 spend item with a fixed-fee compliance scope — the kind of bounded read produced by our DATS five-stage AI methodology — and treating Annex III as a 2026 programme-design item with a 2027 delivery deadline. The buyers we see making the wrong call are deferring both — which leaves them in November 2026 trying to retrofit disclosure logic into a voice runtime that was never built for it. That retrofit is the kind of failure mode we wrote about in AI voice pilot purgatory: the programme stalls because the foundation work was delayed, not because the technology missed.

The pattern the omnibus surfaces is the broader one underneath every enterprise voice AI deployment: regulation is a runtime constraint, not a paperwork constraint. The vendors who treat Article 50 as configuration — disclosure script, marking primitive, audit log — will hit December 2026 with the file ready. This is the architectural premise behind our voice AI agents product: compliance lives where the call lives. The vendors who treat it as documentation will not. This is the same architectural argument we made in our analysis of voice AI hallucination as a procurement gate: the obligation has to be expressible in the runtime, or it is not real. Procurement teams should treat the omnibus as a re-prioritisation signal, not a deferral signal — exactly the framing baked into our voice AI procurement framework for May 2026.

Going deeper? Read the EU AI Act voice AI obligations guide, the Article 50 voice AI disclosure compliance guide, or our broader enterprise AI voice governance framework. To translate this into your stack, see Dilr Voice enterprise platform.

Written by the Dilr.ai engineering team — practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.