The EU Delayed the AI Act Obligations Your Vendor Builds. The Ones You Deploy Landed on August 2.

On 27 July 2026, Regulation (EU) 2026/1744 entered into force and moved the AI Act's high-risk deadline from 2 August 2026 to 2 December 2027. Most enterprises read the headline, exhaled, and moved the whole work package sixteen months to the right. That was the wrong read. The obligations that were postponed are the ones somebody has to build. The obligations that survived took effect on 2 August 2026, two of them land on you directly as the deployer, and they carry the same maximum penalty as the ones that moved.

The split is not perfectly clean and I will be precise about it below. The deferred regime does contain deployer duties, and half of Article 50 binds providers rather than you. But the deferred set is overwhelmingly build work measured in quarters, and what survived is a short set of decisions measured in days. That asymmetry is the trap.

Sixteen days have passed. Most companies I talk to have not noticed.

What actually moved

The AI Omnibus was published in the Official Journal on 24 July 2026 and entered into force three days later. It is a real and substantial deferral, and it is worth stating precisely what it did:

• Annex III high-risk systems (Article 6(2)): the standalone high-risk categories, covering employment and HR, creditworthiness, education, essential services, law enforcement, biometrics, critical infrastructure and justice, moved from 2 August 2026 to 2 December 2027.
• Annex I embedded high-risk systems (Article 6(1)): AI as a safety component inside a regulated product moved from 2 August 2027 to 2 August 2028.
• Systems intended for use by public authorities and already on the market keep their original Article 111(2) runway to 2 August 2030. The Omnibus did not grant that one, it clarified it, confirming the legacy grace applies at the level of the system's type and model rather than the individual unit.
• Article 50(2) machine-readable marking of synthetic content received a grace period to 2 December 2026 for systems placed on the market before 2 August 2026.
• Two new prohibitions were added under Article 5, effective 2 December 2026, covering non-consensual intimate imagery and CSAM generation.

That is a genuine reprieve, and if you are building or reselling an Annex III system you should take it seriously and use it.

If you built your regulatory calendar off guidance published earlier this year, including mine, move the high-risk row to December 2027. Do not move the transparency row. It did not go anywhere.

Now look at what that list has in common. Risk management systems. Technical documentation under Annex IV. Data governance and bias testing on training sets. Conformity assessment. CE marking. Registration in the EU database. Every one of those is a construction project measured in quarters and euros, and almost every one of them belongs to the provider. Where the high-risk regime does touch you as a deployer, through Article 26 human oversight procedures, six-month log retention, and the Article 27 fundamental rights impact assessment, those moved to December 2027 as well.

The Omnibus delayed the expensive things. It did not delay the cheap ones.

What did not move

Article 50 applied on schedule on 2 August 2026, and the Omnibus made no substantive amendment to it. It has four obligation-imposing paragraphs plus a fifth that governs timing, and the duties are not evenly distributed:

Paragraph 1 binds providers. Systems intended to interact directly with people must be designed so those people know they are talking to an AI, unless it is obvious.

Paragraph 2 binds providers. Systems generating synthetic audio, image, video or text must mark their outputs in a machine-readable, detectable format. This is the paragraph that got the grace period to 2 December 2026, and only for systems already on the market.

Paragraph 3 binds deployers. If you run an emotion recognition system or a biometric categorisation system, you must inform the people exposed to it that it is operating.

Paragraph 4 binds deployers. If you deploy a system that generates or manipulates image, audio or video content constituting a deep fake, you must disclose that the content is artificially generated. And where AI-generated or manipulated text is published to inform the public on matters of public interest, you must disclose that too, unless the content went through human review or editorial control and a natural or legal person holds editorial responsibility for its publication.

Paragraph 5 sets the timing: the disclosure must reach the person concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure, and it must meet applicable accessibility requirements. A line in your terms of service does not satisfy this. Neither does a badge that appears after the conversation ends.

Half of Article 50 is yours. There is no conformity assessment involved, no notified body, no vendor dependency, and no procurement cycle to hide behind. Which means there is also nothing to point at when a regulator asks why it was not done.

The asymmetry nobody priced

Here is the structural point, and it is the reason the delay creates a risk the deadline never did.

The obligations that were postponed require a build. The obligations that survived require a decision.

A risk management system takes a program, a budget line, a steering committee and eighteen months. It is visible. It shows up in a capital request. Someone owns it and reports on it monthly, which is exactly why it got a deferral: enough companies could demonstrate to Brussels that they were not going to make it.

Adding one sentence of disclosure to a synthetic voice in your IVR takes an afternoon. It has no budget line, no owner, no steering committee, and no line item in a status deck. Nothing in an enterprise gets done because it is easy. Things get done because someone is accountable for them on a Thursday.

This is the failure mode I have watched in every regulated rollout: organisations execute the obligations that look like projects and miss the obligations that look like footnotes. The AI Act just handed every European enterprise a set of obligations that look like footnotes and are priced like projects.

The penalty tier is the tell

If you want to know how the legislator weighted transparency, read Article 99.

Article 99(4) sets the €15,000,000 or 3% of total worldwide annual turnover tier, whichever is higher. That paragraph covers the provider obligations in Article 16, the authorised representative obligations in Article 22, importer and distributor obligations in Articles 23 and 24, the notified body articles, and deployer obligations under Article 26. Subparagraph (g) of that same paragraph reads: transparency obligations for providers and deployers pursuant to Article 50.

Article 50 is not in a lesser tier. It sits in the same paragraph, at the same cap, as the high-risk deployer duties that were just postponed by sixteen months. The EU deferred the deadline on one and left the other live at an identical price.

For small and medium enterprises the cap is the lower of the two figures rather than the higher, and the Omnibus extended that same treatment to the new small mid-cap category it created. For everyone reading this at enterprise scale, 3% is the operative number, and it attaches to the undertaking, which on the prevailing reading means consolidated worldwide group turnover rather than the revenue of the European subsidiary that happened to deploy the system.

There is a fair objection here, and you will hear it from your own general counsel, so it is worth answering. Part of what drove the delay was that member states have not all finished standing up their authorities, which means enforcement capacity for Article 50 is uneven right now and in some markets the designated market surveillance authority is not yet ready to act on anything. That is true. It is a reason to plan calmly rather than to panic, and it is not a reason to do nothing, for three reasons. Where authorities are designated, Article 50 is live and enforceable today. Article 50 breaches are unusually self-documenting, because the recorded call, the archived page and the published PDF all persist, so an omission this quarter becomes evidence in a proceeding two years from now rather than a moment that quietly passes. And the fine was never the only exposure. Customer contracts increasingly warrant AI Act compliance, consumer and unfair commercial practice claims do not wait for a market surveillance authority, and works councils run on their own timetable.

Why your inventory will miss it

Now the part that turns this from a legal briefing into an operating problem. The asymmetry above explains why the surviving obligations get no owner. There is a second problem, independent of the first, which is that they never reach the list at all.

The delayed obligations attach to systems your company calls AI. The surviving obligations attach to systems nobody in your company calls AI.

Your AI governance committee has a register. On that register are the things that were procured as AI: the copilot, the agent platform, the model gateway, the document intelligence pilot. Those are the systems that would have triggered Annex III analysis, and those are the systems that just got sixteen months.

Article 50 exposure does not live there. It lives here:

• Contact centre sentiment scoring. Sold as quality management, bought by the CX team, renewed by procurement as a telephony line item. If it infers the caller's emotional state from their voice, which is biometric data, Article 50(3) applies and the caller has to be told at the start of the call, not in the recording notice you already have, which discloses recording rather than emotion inference. Two refinements matter here. Sentiment scoring run only over chat text or call transcripts is not emotion recognition under the Act's definition, which requires biometric data, so a keyword sweep will over-collect. And if the same product scores your own agents' emotions, which most quality management suites do by default, you are no longer in disclosure territory at all. See the next bullet.
• Emotion inference on your own employees. This is the one that turns an inventory into an incident. Article 5(1)(f) prohibits inferring emotions from biometric data in the workplace, and the Commission's guidelines on prohibited practices read "workplace" to include recruitment. It has been in force since February 2025, and it sits in the 7% tier, not the 3% one. So if your interview analytics infer a candidate's emotions from face or voice, or your quality management suite scores your agents' frustration from their tone, the answer is not a disclosure notice. The answer is to turn it off and take advice. Biometric categorisation of candidates that stops short of emotion inference stays in Article 50(3) and does need the notice. Sit with the contrast, because it is the whole argument in miniature: the Annex III high-risk analysis of your screening tool was deferred to December 2027, and neither the Article 50(3) notice nor the Article 5 prohibition moved with it.
• Synthetic humans, internal and external. L&D has been generating narrated modules with cloned voices and generated presenters for two years, and marketing has been doing the same in brand video. The boundary matters here: a deep fake requires resemblance to something that exists plus a false appearance of authenticity. A visibly synthetic presenter who resembles nobody in particular generally falls outside Article 50(4). A cloned voice of your actual CFO, a de-aged executive, or a resurrected likeness is squarely inside it.
• Published text on matters of public interest. This is the one that surprises executives most. Your sustainability report, your safety notices, your policy positions, your newsroom. If AI generated or materially manipulated that text, disclosure applies unless you clear an exemption with two limbs: human review or editorial control, and a natural or legal person holding editorial responsibility for the publication. The second limb is the easy one. Your company is a legal person and its name is already on the cover. The first limb is where enterprises actually fail, because the Commission's guidance asks for deliberate examination of the substance by someone with relevant knowledge, and a spell-check or a formal sign-off does not count.

None of these went through the AI governance committee, because none of them were bought as AI. They were bought as a telephony feature, an HR module, a content tool and a marketing service.

The Deployer Transparency Map

You do not need a program for this. You need three answers, they map to the three obligations that are actually yours, and you can have them inside two weeks without involving a single vendor.

Question one: where do we infer a human state?

Search procurement and the CRM, HR and contact centre stacks for the words sentiment, emotion, affect, engagement scoring, attention, tone analysis, and biometric categorisation. Then sort the hits on two axes, because the keyword sweep will over-collect. First, biometric or not. Articles 3(39) and 3(40) both require the inference to be made from biometric data, so voice, facial imagery and physiological signals are in scope, while a sentiment score computed from chat text or a call transcript is outside Article 50(3) entirely and needs no notice. Second, customers or employees: for customers, Article 50(3) requires notice to the person exposed, at first exposure. For employees and candidates, emotion inference from biometric data is prohibited outright under Article 5(1)(f) and no notice will cure it. Route anything in that second box to counsel rather than to your disclosure register.

Question two: where do we synthesise a human?

Any cloned voice, generated presenter, avatar, or manipulated footage of a real or apparent person, across L&D, marketing, IVR and internal communications. Where the output is a deep fake, Article 50(4) requires disclosure. Artistic, creative, satirical and fictional works get a lighter disclosure, not an exemption.

Question three: where do we publish AI-written text about matters of public interest?

Newsroom, ESG and sustainability reporting, safety and product notices, policy and regulatory positions, executive commentary. For each channel, name the person who reviews the substance before publication and describe what that review actually involves. Editorial responsibility is the cheap limb, since your company can hold it. Real review is the expensive one. If you can show that a qualified person examined the substance, you are exempt. If the honest answer is that someone approved a layout, you disclose. The deliverable here is not a name on a masthead. It is evidence that a person who knew the subject actually read it.

Every one of these three is answerable by an existing team using a tool it already owns. The output is a one-page register with four columns: system, article, disclosure text, owner.

One caution, because I have written before about compliance artefacts that describe a reality they do not create. The register is not the defence. The shipped disclosure is the defence. The register is how you prove the disclosure was deliberate rather than accidental, and it is worthless in a proceeding if the fourth column is populated and the third one never made it into production.

The real risk of the delay

I want to be clear that the Omnibus was reasonable policy. Notified body capacity was not there, the standards were not finished, and a deadline nobody can meet produces theatre rather than compliance. Using the sixteen months well is the right response.

The risk is not the delay itself. The risk is calendar contagion.

In most enterprises, the AI Act existed as a single line in a legal or risk calendar reading "AI Act, 2 August 2026." When the date attached to that line moved, everything attached to the line moved with it, including the parts that did not move. The delay did not just defer work. It deferred attention, and attention is the scarcer input.

That is a governance defect, not a legal one, and it has a cheap fix. Split the line. One entry for the high-risk regime, dated 2 December 2027, owned by whoever owns your provider relationships and your Annex III systems. One entry dated 2 August 2026, already live, owned by whoever owns customer communications, HR technology and published content. They are different obligations, on different clocks, held by different people, and they should never have shared a row.

What I would do this week

If you deploy AI anywhere in the EU market, four things:

1. Run the three questions. Two weeks, no vendor involvement, output is a one-page register.
2. Write the disclosure sentences before you find the systems. Draft the three standard notices now, in every language you operate in, meeting the accessibility requirement. When the register comes back, deployment is a copy and paste rather than a drafting project. This is the step that turns a two-quarter effort into a two-week one.
3. Send your providers two written questions. Paragraphs 1 and 2 of Article 50 are their duty, not yours, but the exposure when a customer notices is commercial and reputational and it lands on your brand. Ask what their Article 50(1) interaction disclosure does, and how they have implemented Article 50(2) marking, including whether they are relying on the grace period for systems placed on the market before 2 August 2026. Put the answers in the renewal file. Separately, if you resell or white-label a provider's system under your own name, ask counsel whether you have become the provider for it rather than the deployer, because that changes which half of Article 50 you own.
4. Re-date your calendar entry. Two rows, two owners, two dates.

The high-risk regime will arrive on 2 December 2027 and it will be expensive and it will need every month of the extension you were just granted. Spend those months on it.

Just do not spend the first sixteen days of them, as the companies I speak with mostly have, believing you were granted an extension on something you were not.

Shubhendu Tripathi is an AI and ERP strategy consultant based in Toronto, and the host of The Integration Layer, a podcast on AI, enterprise systems, and the work of making them fit together. Connect on LinkedIn or reach out at tripathis@qubittron.com.