EC Article 50 guidelines: a voice AI deployer checklist

On 8 May 2026 the European Commission published a 40-page draft of its guidelines on the transparency obligations under Article 50 of the EU AI Act. The consultation closes on 3 June 2026 — 13 days from now — and the obligations apply with full force from 2 August 2026. The penalty ceiling sits at €15 million or 3% of global annual turnover, whichever is higher.

This is the first Commission instrument that interprets the full scope of Article 50. The Article itself is a short list of duties; the draft guidelines are where the operational detail lives — when disclosure is required, what counts as "obvious", and how synthetic audio must be marked. For the foundational text, see our earlier guide to Article 50 voice AI disclosure obligations. This post is narrower and more urgent: it translates the draft guidelines into a deployer-side checklist for enterprises running voice AI in the EU market.

The window that matters is not 2 August. It is 3 June. Enterprises that wait for the final text will have missed the only point at which these obligations are still negotiable.

This checklist is shipped by the team behind Dilr Voice — enterprise voice AI with disclosure logic, recording-consent management and audit trails built into the runtime. Or see DATS, our senior-led AI consulting practice for regulated deployments.

Most enterprises treat the AI Act as a single compliance event with one owner. The draft guidelines make that posture untenable for voice. A deployer who assumes the vendor "handles Article 50" will discover at audit that two of the four obligations were theirs all along — and that the contract said nothing about either.

What the draft guidelines actually change

Article 50 has been on the statute book since the AI Act entered force. What changed on 8 May 2026 is interpretation. The draft guidelines resolve three questions that the bare text left open — and each resolution has a direct operational consequence for a voice deployment.

Disclosure must be at first interaction, per person, and unburied. The guidelines fix the timing: an interactive AI system must disclose its non-human nature "at the latest at the time of first interaction" with each natural person exposed. "Per person" matters for voice. A caller who is transferred, or who joins a conference line midway, has not heard the opening disclosure — the guidelines require repeated or persistent disclosure in exactly those cases. And the disclosure must be "clear and distinguishable": a line buried in terms and conditions, or in a layered IVR menu, does not satisfy the obligation. For a voice agent this is straightforward to engineer if it is designed in — and expensive to retrofit if it is not, a point we cover in our guide to change management in AI voice deployment.

The "obvious" exception is narrower than buyers assume. Disclosure can be skipped only where the AI nature is "obvious" to a reasonably informed, observant and circumspect person — the average-consumer test. The guidelines weight that test against the actual audience: where vulnerable groups (children, elderly or disabled callers) are in scope, the bar for "obvious" rises sharply. For a consumer-facing voice line handling, say, energy-vulnerability or collections calls, "obvious" is almost never available as a defence. The honest reading: most enterprise voice deployments must disclose, and the exception is a documented judgement call, not a default.

Synthetic audio marking has no prescribed standard yet. This is the non-consensus point. The guidelines confirm that synthetic content must be marked in a machine-readable format under Article 50(2) — but they explicitly state that no single marking technique currently meets the Act's tests of effectiveness, interoperability, robustness and reliability simultaneously, and that a combination is therefore required. For synthetic audio specifically, the guidelines prescribe no audio-specific architecture at all. The technical detail is deferred to a separate Code of Practice. The deployer consequence is uncomfortable: you cannot wait for a prescriptive audio standard, because there is not going to be one before 2 August. You have to specify marking expectations yourself, in the contract, now.

The voice AI deployer checklist

The draft guidelines reward enterprises that separate the four obligations and assign each an owner. Conflating them is the single most common error — and it leaves real gaps. Here is the deployer-side split, mapped to the action each obligation demands. The same logic underpins our AI execution office, where the disclosure runbook is owned as a standing programme rather than a one-off review.

Two of these four are yours. If your voice programme runs AI voice sentiment analysis — and most enterprise deployments now do — Article 50(3) is a deployer obligation you cannot delegate to the vendor. The notice is cheap to add to an opening script and expensive to omit from an audit file.

Decide whether disclosure applies — then document it

The decision is mechanical once the obligations are separated. The tree below is the version we use in deployment reviews. The point of writing it down is not the answer; it is the audit artefact. When a regulator asks why a particular line skipped disclosure, "we judged it obvious" is not a defence — "here is the documented average-consumer assessment, dated and signed" is.

Fix the contract before the obligation bites

Most voice AI contracts signed before late 2025 contain no Article 50 language at all. That is the highest-leverage gap to close. At the next renewal — or sooner, by amendment — a deployer should require: a provider warranty that 50(1) self-disclosure is engineered and configurable; a written 50(2) marking statement for synthetic audio; and indemnity allocation for transparency failures. This is procurement work, not legal box-ticking, and it sits naturally inside an AI placement diagnostic — a fixed-fee assessment that ranks every gap by exposure before any deployment commitment. For the wider governance scaffold, see our enterprise AI voice governance framework, and for how this slots into a vendor map, our voice AI procurement framework.

What this looks like in a regulated deployment

Consider a mid-market UK insurer running an outbound voice agent for claims follow-up — a representative pattern, not a named client. Before the guidelines, the programme assumed the platform vendor "owned the AI Act". The deployer checklist surfaced two gaps in an afternoon: the agent ran emotion detection to route distressed callers (a 50(3) deployer notice, missing) and the opening script disclosed AI nature only on the first leg, not after warm transfer (a 50(1) timing failure). Both were fixed before 2 August at near-zero cost — a script edit and a runtime flag. The cost of finding them at audit instead would have been a remediation programme and a defensible-position scramble. The difference was sequencing: separating the four obligations early, while the consultation window was still open. The same discipline applies whether the vehicle is a privatised AI build through DATS or an off-the-shelf voice agent — the deployer obligations do not transfer.

Closing the gap before 2 August is cheaper than closing it after. Try Dilr Voice with disclosure logic in the runtime, book an AI placement diagnostic, see our DATS methodology for regulated builds, or read about our approach to placing AI inside compliant enterprise systems.

Two further points worth tracking. The guidelines confirm a transitional period to 2 December 2026 for generative AI already on the market — relevant to synthetic-audio marking, but not to the interactive-disclosure and deployer obligations, which bite on 2 August. And the broader EU AI Act omnibus did not move the Article 50 clock. If you have not yet built one, the disclosure obligations should also be reflected in your AI tool inventory — the register that ICO, FCA and the EU AI Act now all expect. If you want a second read on where your specific estate sits, speak to our operators before the consultation closes.

The wider context is sobering. McKinsey's The State of AI 2025 found that roughly 88% of enterprises now use AI, yet only about 6% capture material EBIT impact. Compliance gaps are one of the quiet reasons programmes stall short of value: an unmanaged transparency exposure becomes a deployment freeze the moment legal notices it. The enterprises that treat the 8 May draft as a planning input — not a press release — convert a regulatory deadline into a procurement advantage. The consultation is still open. Use it.

The full draft is published on the European Commission's digital-strategy guidelines page, and a clear practitioner read of the ten key takeaways is set out in this Inside Global Tech analysis.

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.