The deadline that didn't move: what August 2 still asks of your customer-facing AI

Version 1.0 · Published 2026-07-06

If you run customer-facing AI in Europe, the past month of headlines has been doing you a quiet disservice. The Digital Omnibus passed, the words “postponed” and “delay” traveled far, and somewhere along the way a lot of teams filed the EU AI Act under “next year’s problem.” For most companies deploying AI toward customers, that is the wrong conclusion.

What moved is the high-risk regime. What did not move is Article 50, the transparency chapter, and that is the part that touches almost every customer-facing AI deployment. By the end of this piece you will know exactly which dates shifted, which one still lands on August 2, 2026, and what to ask your AI vendor before then.

What the Omnibus actually moved

The Digital Omnibus is real, and the postponement is real. The European Parliament endorsed it on June 16, 2026, and the Council gave its final green light on June 29. It takes legal effect once published in the Official Journal, which is expected before August 2.

The dates that shifted are the high-risk ones. Obligations for standalone high-risk systems under Annex III move to December 2, 2027, and for AI embedded in regulated products under Annex I to August 2, 2028, per Gibson Dunn’s analysis of the agreement. One transparency item also got breathing room: generative AI systems already on the market before August 2, 2026 have until December 2, 2026 to meet the machine-readable marking requirement for AI-generated content, which is a provider obligation.

That is the full extent of the relief. Everything else in Article 50 stays exactly where it was.

The deadline that did not move

From August 2, 2026, Article 50 of the EU AI Act applies to providers and deployers alike, and it does not care whether your AI is high-risk. It covers four situations: AI that interacts directly with people, AI that generates synthetic content, emotion recognition or biometric categorization, and deepfakes or AI-generated text on matters of public interest.

For customer-facing AI, the first situation is the one that bites. Any system intended to interact with people must be designed so users know they are talking to AI, disclosed clearly at the very start of the interaction. A label buried in the footer or the terms and conditions does not qualify.

There is an exception when the AI nature is obvious to a reasonably well-informed person, but the Commission’s draft guidelines warn against leaning on it. If your customers include lay users rather than software professionals, assume the disclosure duty applies. That assumption costs you a design decision; the opposite assumption could cost you an enforcement case.

Deployer is not a spectator role

Here is the distinction the headlines skipped: under the AI Act, the company that builds the AI system is the provider, and the company that puts it in front of its own customers is the deployer. If you embed an AI agent in your product or portal, that deployer is you. The postponement moved deployer obligations for high-risk systems, not the transparency duties that come with customer-facing AI.

Concretely, deployers must inform people when emotion recognition or biometric categorization is in play, and must disclose deepfakes and AI-generated text published to inform the public. Just as concretely, the disclosure duty for interactive AI is designed into the product by the provider, but you are the one who enables and configures it in your deployment. A vendor can hand you compliant functionality; only you can switch it on.

And do not forget the obligation that is already live: AI literacy under Article 4 has applied since February 2025. Your staff operating the AI need to know what it does and where its limits are, a duty both parties carry.

The July checklist

The good news is that July is exactly enough time, if you use it as a checklist rather than a countdown. Here is the version we would run:

• Map your Article 50 exposure. List every AI touchpoint your customers meet: agent, search, generated help content. For each, note which of the four Article 50 situations applies.

• Check the disclosure, on screen, yourself. Open your customer-facing AI as a customer would. If it does not say it is AI before the conversation starts, that is your first fix.

• Ask your vendor four questions. Does the product ship Article 50 disclosure functionality, and is it enabled in your deployment? Are AI outputs marked as AI-generated in a machine-readable way, or is there a dated plan to meet the December 2 marking deadline? Can you export per-interaction logs? Will they share the model documentation you need as a deployer?

• Ask about the Code of Practice, this week. The Commission’s Code of Practice on transparency of AI-generated content is voluntary, but signing brings a presumption of conformity with the marking and labelling duties it covers. Vendors that want to be listed as initial signatories must submit by July 22, 2026, 18:00 CEST. Whether your vendor signs tells you something either way.

• Put the role split in writing. Provider and deployer each have homework. If your vendor cannot show you a document that allocates the two roles, you are holding both sets of obligations by default.

How UNLESS handles its half

We should be clear about our own position in this, because we sit on the provider side of the split. UNLESS builds Article 50 disclosure functionality into the Customer Agent, provides technical means to mark AI-generated content, and keeps a per-decision audit trail with timestamps, decision paths, model used, and source citations that a DPO or auditor can read and export.

The role allocation is written down, not implied: Unless acts as provider, our customer acts as deployer, and the compliance appendix spells out who does what. That is the honest scope of any vendor claim in this space. A provider can make August 2 easy for you; no provider can make it disappear for you.

The deadline that matters was never the one in the headlines. Run the checklist this month, put the vendor answers in writing, and August 2 becomes a Tuesday.