Ales Frelih

You may have read the headlines in May 2026: the EU is delaying the AI Act. Relief all around. Some compliance teams packed up their urgency and moved on.

Here is what those headlines missed: the delay is not yet law.

As of today, 45 days before the 2 August 2026 deadline, the proposed change is still just a proposal — a political agreement, not a signed regulation. If it is not formally published as law before 2 August, the original deadlines come back into force. Nothing changes.

This article tells you exactly what the proposed change says, what it does not touch, what is confirmed for August regardless, and what you should be doing right now.

—

## What happened in May 2026?

In November 2025, the European Commission proposed a package of amendments to the EU AI Act. The goal: give businesses more time to comply with the most demanding rules — specifically those covering “high-risk” AI systems.

After months of negotiations between EU institutions, a provisional political agreement was reached on 6 May 2026. Member State representatives confirmed it on 13 May 2026.

That sounds official. It is not. In EU law, a political agreement is a handshake, not a signed contract. To become law, the change still needs:

– A formal vote in the European Parliament

– Sign-off by the Council

– Publication in the EU’s Official Journal

None of those steps have been completed yet. Until they are, the original AI Act deadlines are the law.

—

## What would the proposed change actually do?

Assuming the amendment is formally adopted before 2 August, here is what it would change:

### More time for high-risk AI systems

The EU AI Act classifies certain AI uses as “high-risk” — recruitment tools, credit scoring, insurance risk assessment, student grading, tenant screening, border management, law enforcement. These systems face the strictest compliance requirements.

Under the original law, those requirements were due from **2 August 2026**.

Under the proposed amendment:

| AI system type | Original deadline | Proposed new deadline |

|—|—|—|

| Standalone high-risk systems (recruitment, credit, law enforcement…) | 2 August 2026 | 2 December 2027 |

| Embedded high-risk systems (medical devices, vehicles, machinery…) | 2 August 2027 | 2 August 2028 |

That is roughly 16 extra months for the first category, 12 months for the second. Significant — if the amendment passes.

### A short grace period for content watermarking

Article 50 of the AI Act requires providers to embed invisible machine-readable markers in AI-generated content (images, video, audio). For AI systems already deployed before 2 August 2026, the proposed amendment gives a four-month grace period: the watermarking obligation would start **2 December 2026** instead of 2 August.

This applies only to systems already on the market. New deployments after 2 August are subject to the watermarking rule from day one.

### A new type of AI is banned

The proposed amendment adds two new prohibitions: AI systems designed to generate non-consensual intimate imagery, and systems that produce child sexual abuse material. These would be added to the list of AI uses that are banned outright in the EU.

—

## What the proposed change does NOT touch

This matters more than the delays.

**The transparency rules still apply from 2 August 2026.**

Article 50 requires businesses to tell users when they are interacting with an AI system — a chatbot, an AI assistant, a voice bot. This rule is not delayed. It applies from 2 August 2026 under every version of the law.

**The staff training obligation is already in force.**

Article 4 of the AI Act — requiring businesses to ensure their staff have adequate understanding of AI — has applied since 2 February 2025. The proposed amendment softens the wording slightly, but the obligation itself is not going away.

**The rules for large AI model providers (ChatGPT, Gemini, etc.) are unchanged.**

Those rules have applied since 2 August 2025. No change proposed.

The proposed deadline extension is targeted, specific relief for high-risk AI system compliance. It does not delay the obligations that apply to most SMEs — the transparency requirements and the staff literacy obligation.

—

## The risk that is not being discussed

Here is the scenario that every compliance team should be planning for:

**If the proposed amendment is not formally enacted before 2 August 2026, the original law applies — in full.**

No political agreement. No headline. No delay. Just the law as written.

A law firm analysis published on 27 May 2026 described formal adoption as “expected in the coming weeks, in advance of the 2 August 2026 deadline.” That was three weeks ago. The formal vote has not happened yet.

If you operate any AI system involved in recruitment decisions, credit scoring, insurance underwriting, tenant screening, education assessment, or law enforcement — you need a plan for August 2 that assumes the proposed delay does not come into force.

This is not pessimism. It is risk management. The downside of planning for the original deadline and then getting more time is: you are over-prepared. The downside of planning for the delay and then not getting it is: you are in breach of regulation on day one.

—

## Microsoft just published a free compliance toolkit

While EU institutions negotiate, Microsoft published a practical resource on GitHub — the Agent Governance Toolkit — an open-source framework for organisations deploying AI agents.

It includes a structured EU AI Act compliance checklist, covering identity, execution controls, human oversight, and reliability requirements. It maps to the OWASP Agentic Top 10 risk list.

For SMEs without dedicated legal teams, a structured checklist from Microsoft is a solid starting point — far more actionable than reading 144 articles of regulation, and credible enough to use in an internal audit.

The fact that a company the size of Microsoft is publishing practical EU AI Act checklists in June 2026 tells you something: the August deadline is real, and the technology industry is taking it seriously.

—

## What to do right now — a short action list

Whether the proposed amendment passes or not, your list is the same:

**1. Implement Article 50 transparency before 2 August.** If you use AI in any customer-facing interaction — a chatbot, an email assistant, a recommendation engine — you need a clear disclosure process. Tell users they are interacting with AI. This deadline is not changing.

**2. Document your Article 4 staff training.** AI literacy has been legally required since February 2025. If you have not run training and kept records, do it now. If you have done it, make sure you have proof.

**3. Check whether you use high-risk AI.** Run through the Annex III list: recruitment, credit, insurance, education, law enforcement, border management. If you use AI in any of those areas, you need to understand your obligations — even with a potential extension, December 2027 is not far off.

**4. Watch for the formal adoption announcement.** The signal you are waiting for is publication in the EU’s Official Journal. Until that happens, the original law is the law.

**5. Use free resources to assess your exposure.** You do not need a law firm to take the first step. The free checklists below take 15 minutes and tell you where you stand.

—

## The bottom line

The proposed deadline extension for high-risk AI systems may well pass. It probably will. But it has not passed yet, and it does not affect the obligations that apply to most businesses using AI — transparency disclosures and staff training.

Plan for 2 August 2026. Use any extension you get as an opportunity to do the work properly, not as a reason to stop.

—

## Free resources to get started

**Free Compliance Check (German):** [Bin ich betroffen? EU KI-Verordnung Selbstcheck](https://frelih.gumroad.com/l/slwzch)

A downloadable checklist covering Article 50 and Article 4 obligations — for German-speaking SMEs.

**Free Compliance Check (French):** [Suis-je concerné? Check EU AI Act](https://frelih.gumroad.com/l/bghydm)

The same resource for French-speaking markets.

**Full Compliance Kits** for teams that need policy templates, internal training documentation, and disclosure statements:

– [Complete Compliance Kit — Deutsch — €149](https://frelih.gumroad.com/l/mjsaqg)

– [Complete Compliance Kit — Français — €149](https://frelih.gumroad.com/l/xdyenu)

—

*Sources: Gibson Dunn Client Alert, 27 May 2026; Microsoft agent-governance-toolkit (GitHub, June 2026); EU AI Act (Regulation EU 2024/1689); European Commission proposal, November 2025.*

*This article is for informational purposes only and does not constitute legal advice. Consult qualified legal counsel for advice specific to your organisation.*