← Back to blog
A small cube device on a stone plinth sits inside a glowing inspection frame beside blank translucent panels, evoking EU AI Act conformity checks.
AI News & Policy·July 3, 2026·8 min read

The EU AI Act's High-Risk Rules Moved to December 2027: What Actually Changes for Builders

The EU AI Act high-risk obligations were due on August 2, 2026, until the AI Omnibus moved Annex III to December 2, 2027. Here is what is required, of whom, and how it compares with the US, UK, and China.

The compliance obligations for "high-risk" AI systems under the EU AI Act were scheduled to become enforceable on August 2, 2026, until the AI Omnibus amendment, in force since July 27, 2026, moved the Annex III high-risk rules to December 2, 2027. The regulation itself has been law since August 2024, but the high-risk deadline is the date the Act stops being a thing legal teams track and starts being a thing engineering teams have to build against. If you're shipping an AI-powered hiring tool, credit-scoring model, or education-assessment product into the EU market, this is the deadline that determines whether your product can keep shipping there.

Most coverage of the AI Act treats it as a monolith: "the EU regulated AI." It didn't. It built a risk-tiered system with different obligations, different deadlines, and different enforcement mechanisms depending on what your system does and who it affects. This piece is about the mechanics (what's actually required, of whom, by when), not the politics of whether the Act is a good idea.

The phase-in timeline, in full

The Act didn't take effect all at once. It's been rolling out in stages since it entered into force (the Commission's AI Act Service Desk keeps the official implementation timeline current), and understanding where the high-risk deadline sits in that sequence matters for scoping what's actually new.

DateWhat appliesScope
Aug 1, 2024Regulation enters into forceClock starts on all subsequent deadlines
Feb 2, 2025Prohibited practices (Art. 5)Social scoring, manipulative/subliminal techniques, untargeted facial-recognition scraping, emotion inference in workplaces/schools, real-time biometric ID in public (with narrow law-enforcement exceptions): all banned outright, no tier
Aug 2, 2025GPAI model obligations (Ch. V)General-purpose AI model providers: technical documentation, downstream transparency, copyright policy, training-data summary; systemic-risk models (≈10^25 FLOPs training compute) get additional model evaluation, adversarial testing, and incident-reporting duties
Aug 2, 2026Transparency obligations (Art. 50); enforcement beginsDisclosure duties for chatbots, deepfakes, and AI-generated content; national and EU enforcement of the prohibitions, GPAI rules, and AI literacy starts
Dec 2, 2026New prohibitions added by the AI OmnibusBans on AI systems that generate non-consensual intimate imagery or child sexual abuse material; generative systems already on the market before Aug 2, 2026 must also have machine-readable marking of synthetic content in place by this date
Dec 2, 2027 (originally Aug 2, 2026)High-risk AI system obligations (Annex III)Providers and deployers of AI used in employment, credit, education, law enforcement, migration, essential services, etc.
Aug 2, 2028 (originally Aug 2, 2027)High-risk obligations for regulated products (Annex I)AI as a safety component in medical devices, toys, lifts, and other CE-marked product categories already regulated under existing EU product law (the Omnibus moved machinery out of this list; AI rules for machinery will come through the Machinery Regulation instead)

So the high-risk deadline originally set for August 2026 (now December 2027) is specifically about Annex III: the list of standalone AI use cases the Act deems high-risk because of their effect on people's rights, not because they're embedded in a regulated physical product. That second category (Annex I) gets a longer runway (now August 2028), mostly because it has to interoperate with existing sectoral conformity regimes for products like medical devices.

What counts as "high-risk" under Annex III

The Act doesn't define high-risk by technology (it's not "any system using a neural network"). It defines it by use case. Annex III lists eight categories, and if your system falls into one of them, the high-risk obligations attach regardless of how sophisticated or simple the underlying model is:

A practical filter: if your AI feature makes or materially informs a decision about a person's job, credit, education, legal status, or access to a public service, assume it's in scope until a lawyer tells you otherwise. This catches a lot of "boring" B2B SaaS (applicant tracking systems, underwriting tools, proctoring software) that doesn't feel like frontier AI but sits squarely in Annex III, which is why AI hiring tools are already running into bias audit laws on both sides of the Atlantic.

A stack of glowing boxes, a magnifier, a stamp and a sealed envelope sit beside a small robot awaiting inspection, evoking provider compliance duties under the EU AI Act.

What providers actually have to do

"Provider" means whoever develops the system or has it developed and places it on the market under their name. If that's you, Articles 8 through 17 require:

  1. A risk management system: an iterative process across the system's lifecycle identifying and mitigating known and foreseeable risks, not a one-time document.
  2. Data governance: training/validation/testing data must be relevant, sufficiently representative, and examined for biases that could affect health, safety, or fundamental rights.
  3. Technical documentation: detailed enough that an authority could assess conformity without access to your source code (Annex IV specifies the contents: design specs, development process, monitoring capabilities, risk management measures).
  4. Record-keeping (logging): the system must automatically log events over its lifetime to enable traceability, at a level appropriate to the intended purpose.
  5. Transparency to deployers: instructions for use covering capabilities, limitations, expected accuracy, and known misuse scenarios.
  6. Human oversight: the system must be designed so a human can effectively oversee it, including the ability to intervene or stop it.
  7. Accuracy, robustness, and cybersecurity: appropriate levels given the intended purpose, tested and documented.
  8. A quality management system at the organizational level, plus a conformity assessment (self-assessment for most Annex III categories, third-party for a smaller subset like remote biometric ID), a signed EU declaration of conformity, CE marking, and registration in the EU database for high-risk AI systems before market placement.

That last item is the one people underestimate the lead time on. Conformity assessment and documentation aren't things you retrofit in the two weeks before a deadline; Annex IV technical documentation alone is a substantial artifact, and building the underlying risk-management and data-governance processes it documents takes real engineering time.

The AI Omnibus left this list of obligations intact. What it changed is the route for smaller companies: SMEs and small mid-cap companies can supply the Annex IV technical documentation through a simplified Commission form, and quality management obligations scale with the provider's size. It also added Article 4a, which lets providers process special categories of personal data where strictly necessary to detect and correct bias, under specific safeguards (summary of the amendments).

What deployers have to do: a separate, lighter set of duties

The Act distinguishes providers from deployers (the organizations that use a high-risk system in their own operations, e.g., an HR department using a third-party screening tool). Deployer obligations under Article 26 are narrower but not nothing, and the AI Omnibus postponed them to December 2027 without changing their content:

That last obligation is the one most likely to surprise product teams at fintechs and insurers specifically (insurers are already juggling NAIC guidance on AI models in the US): it applies even if you bought the underlying model from someone else and did none of the training yourselves.

Penalties

ViolationMaximum fine
Prohibited practice (Art. 5)€35M or 7% of global annual turnover, whichever is higher
Non-compliance with high-risk obligations, GPAI obligations, or other substantive requirements€15M or 3% of global annual turnover
Supplying incorrect, incomplete, or misleading information to authorities€7.5M or 1% of global annual turnover

The percentages track global turnover, not EU revenue: the same structure GDPR uses, and deliberately so. Those are the ceilings for large companies. For SMEs and start-ups, each fine is capped at whichever of the two figures is lower, and the AI Omnibus extended that lower-cap rule to small mid-cap companies for the second and third tiers.

How this compares to the rest of the world

The EU's is the most extensive comprehensive, cross-sectoral, binding AI-specific regime. South Korea's AI Basic Act, in force since January 2026, is the other major one, but it is lighter (fines are capped at KRW 30 million). Most other jurisdictions are either sectoral, principles-based, or still assembling the pieces.

JurisdictionApproachMechanismBinding?
EUComprehensive, risk-tiered horizontal lawAI Act: prohibitions, GPAI rules, high-risk conformity regimeYes, directly applicable across member states
USSectoral + state patchwork, no federal AI-specific statuteExisting agency authority (FTC unfairness/deception, EEOC on algorithmic hiring bias) plus state laws (e.g., Colorado's consequential-decision AI law requiring impact assessments and reasonable-care duties, California disclosure statutes on training data and synthetic content)Yes at state/sectoral level, no unified federal floor
UKPrinciples-based, regulator-led, deliberately not a single AI ActFive cross-sectoral principles (safety, transparency, fairness, accountability, contestability) applied by existing regulators (ICO, FCA, CMA, etc.); AI Security Institute focused on frontier model evaluationPrinciples are guidance; enforcement rides on regulators' existing statutory powers
ChinaRegistration and content-control focusedCAC algorithm registration, generative-AI service measures requiring security assessment and content labeling before public deploymentYes, but scoped narrowly to specific service types rather than all AI use cases

The practical consequence: a product that's compliant for the US market (reasonable-care documentation, some state-level impact assessments) is not automatically compliant for the EU, because the EU regime requires conformity assessment and CE marking as a market-access gate, not just a liability-defense posture. Teams building for both markets end up maintaining two different compliance artifacts even when the underlying risk controls overlap substantially.

What this actually means if you're building right now

If your product touches hiring, credit, insurance, education assessment, or public-service eligibility and you have or want EU users, the practical checklist ahead of the December 2027 deadline is:

  1. Confirm whether your use case sits in Annex III. Don't rely on "we're just a small feature"; the classification is about the decision being made, not the company's size.
  2. If you're a provider, start (or audit) technical documentation and risk-management processes now; these take months to build properly, not weeks.
  3. If you're a deployer using a third-party high-risk system, check whether you owe a fundamental rights impact assessment before first use, independent of what your vendor has already done.
  4. Don't assume US-market compliance work (bias audits, model cards, state-law impact assessments) satisfies EU conformity requirements; the artifacts overlap in spirit but not in form.

The Act's authors built in a two-year runway specifically so this wouldn't be a surprise. Whether that runway was long enough is a fair question, and the AI Omnibus answered it by adding 16 months. That buys time; it doesn't change what has to be built.

#eu-ai-act#ai-regulation#compliance#high-risk-ai#ai-policy#algorithmic-accountability
ShareXLinkedInHacker News

Try it

Free, in your browser, no sign-up.

New posts by email

One email when a new post goes up. No other mail; unsubscribe in one click.

We store your email only to send new posts. See the privacy policy.

Related reading

AI News & Policy
Colorado, California, Texas: Mapping the 2026 US State AI Law Patchwork
AI News & Policy
AI Copyright Litigation, Mapped: Where Fair Use Actually Stands After the Anthropic Settlement
Applied AI
Where AI Actually Works in Government: Citizen Chatbots, Fraud Detection, and the Due-Process Problem