EU AI Act Compliance in 2026: What the Omnibus Changed and What Is Due Now

Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on July 24, 2026 and entered into force on July 27.²⁷ It moved the start of the high-risk obligations by 16 months for stand-alone systems and by 12 months for AI built into regulated products, because the harmonized standards, guidance and national authorities those obligations depend on were delayed.² A compliance plan built before the Omnibus was agreed carries the wrong central date: August 2, 2026 no longer applies to Annex III systems. The transparency duties in Article 50 kept their date and have applied since August 2, 2026.⁴

This article is engineering guidance: counsel should confirm how any specific system is classified, and nothing here is legal advice.

• Aug 2, 2026Article 50 transparency duties for chatbots, synthetic content and deepfakes began to apply⁴
• Dec 2, 2027High-risk obligations begin for Annex III systems in areas such as employment, credit and education²
• 7%of total worldwide annual turnover, or €35 million if higher, the maximum fine for a prohibited practice¹

EU AI Act deadlines after the Omnibus

Three stages of the Act are already in force, one more deadline falls on December 2, 2026, and the high-risk obligations now start on December 2, 2027 and August 2, 2028.¹² The Omnibus set these as fixed dates in the regulation itself.²

The Omnibus left Articles 9, 12, 13, 14 and 15, the core high-risk requirements, untouched.² Its other changes that bear on engineering and procurement:

• AI literacy Article 4 now requires providers and deployers to take measures to support the AI literacy of their staff, and states that no specific level has to be guaranteed for any individual.²⁸
• Small mid-caps SME relief now extends to small mid-cap enterprises: a simplified technical documentation form, a proportionate quality management system and lower fine caps.²³
• Registration A provider that decides an Annex III system is not high-risk must still register it, with less information than before.²⁷
• Bias data A new Article 4a allows special categories of personal data to be processed to detect and correct bias, only where strictly necessary and where synthetic or anonymized data cannot do the job.²
• Cybersecurity A high-risk system that meets the relevant conditions of the Cyber Resilience Act is deemed to meet the cybersecurity requirement in Article 15.²

Does the EU AI Act apply to US, UK and Indian companies?

Yes, if the company places an AI system or general-purpose AI model on the EU market, or if the output of its AI system is used in the EU, wherever the company is established.¹ Article 2(1)(a) covers providers placing systems on the EU market irrespective of where they are located, and Article 2(1)(c) covers providers and deployers in a third country whose system output is used in the Union.¹ A US software company selling a product with AI features to European customers is a provider under the first. A credit model run from Virginia or Bengaluru that scores applicants for an EU subsidiary falls within the plain wording of the second.

The UK is outside the EU, so the Act is not UK law, and UK companies are caught through Article 2 on exactly the same terms as US and Indian ones.¹ A provider established outside the EU must appoint an authorized representative in the EU, by written mandate, before it makes a high-risk system available there.¹ The main exclusions are military, defense and national security uses, systems built solely for scientific research, pre-market research and testing (real-world testing excepted), and open-source systems that are neither high-risk, prohibited nor covered by Article 50.¹

Provider or deployer: your role sets your duties

You are the provider of an AI system if you develop it, or have it developed, and place it on the market or put it into service under your own name; you are a deployer if you use an AI system under your authority.¹ Putting into service includes supplying a system for your own use, so a bank that builds its own credit-scoring model is its provider as well as its deployer, with the heavier set of duties.¹

Roles also move. A deployer, distributor or other third party becomes the provider of a high-risk system if it puts its name on the system, makes a substantial modification to it, or changes the intended purpose of any system, including a general-purpose one, so that it becomes high-risk.¹ The Omnibus sharpened a procurement duty: the provider of a high-risk system and any third party supplying an AI system, model, tool or component used in it must agree in writing the information and technical access the provider needs to comply, and a breach sits in the 3% fine tier.² Contracts with model and tool vendors need that clause before the build starts.

Is your system high-risk? An Annex III walkthrough

An AI system is high-risk if it is a safety component of a product regulated under Annex I that needs third-party conformity assessment, or if its intended purpose falls in one of the eight areas listed in Annex III and no exemption in Article 6(3) applies.¹ Classification follows the intended purpose the provider states, so the test is run use case by use case.¹

• 1. Biometrics Remote biometric identification (identity verification excluded), categorization by sensitive attributes, and emotion recognition.
• 2. Critical infrastructure Safety components in digital infrastructure, road traffic, water, gas, heating and electricity.
• 3. Education Admission, grading learning outcomes, assigning education levels, and detecting cheating in tests.
• 4. Employment Targeted job ads, filtering applications and evaluating candidates; decisions on promotion, termination and task allocation; monitoring performance and behavior.
• 5. Essential services Public benefit eligibility, credit scoring (fraud detection excepted), life and health insurance pricing, and emergency call triage.
• 6 to 8. Public powers Law enforcement; migration, asylum and border control; and the administration of justice and democratic processes.

For large companies outside the public sector, the exposure sits mostly in areas 4 and 5: recruiting tools, workforce analytics, lending and insurance pricing.¹ A provider relying on Article 6(3) must document its reasoning before launch, produce it on request and register the system.¹² For areas 2 to 8, conformity assessment is internal control with no notified body, so your own evidence is the whole case.¹

What Article 50 requires today

Article 50 has applied since August 2, 2026 and requires disclosure whenever people interact with an AI system or see content it generated or manipulated; for most companies it is the part of the Act due now.⁴⁵

The other Article 50 duties took effect on August 2, 2026 with no grace period.⁹ The final Code of Practice on Transparency of AI-generated Content, published June 10, 2026, covers marking and detection for providers and labeling for deployers; about 190 companies and organizations had signed it by the end of July 2026.⁶ Under the Commission's guidelines, signing the code is one way to show compliance with the marking and labeling duties; a company that does not sign must show equivalently adequate means.⁵ In engineering terms: a disclosure component in every conversational surface, a marking step in every generation pipeline, and a detection check in release testing.

The high-risk requirements, mapped to what engineering builds

Each high-risk requirement in Articles 9 to 15 turns into a specific artifact an engineering team produces and keeps current, and a regulator's first request will be for those artifacts.¹ Built into the delivery pipeline, they cost little to maintain; reconstructed a month before an audit, they cost a great deal and prove less.

Around those seven requirements sit the provider's process duties: a quality management system, conformity assessment, the EU declaration of conformity, CE marking, registration and monitoring after launch.¹ The post-market monitoring plan is now part of the technical documentation, and the Commission owes a template for it by September 2, 2027.² Serious incidents must be reported within 15 days of awareness, two days for a widespread infringement or serious disruption of critical infrastructure, and 10 days for a death.¹ Those clocks need an on-call runbook written before launch. In the systems we build, the trace, the evaluation gate and the override queue ship in the first release, and our AI development teams generate the technical file from the pipeline itself.

How the EU AI Act treats AI agents

An AI agent is an AI system under the Act, which has no separate category for agents, so it is classified by its intended purpose like any other system.¹ A customer service agent that talks to people carries the Article 50(1) disclosure duty today. An agent that screens candidates or decides on credit is an Annex III system, and if it profiles the people it assesses, no Article 6(3) exemption is available.¹ A company that connects a general-purpose assistant to its hiring workflow has changed that system's intended purpose and becomes its provider.¹

The high-risk requirements map closely onto agent architecture. Article 14's stop control is the agent's kill switch. Article 12's logs are its step-by-step traces. Article 15's duty to resist attempts to alter a system's use or outputs is where prompt injection through the documents and emails an agent reads belongs, and its feedback-loop rule is relevant to agents that learn from their own memory.¹ Our guide to AI agent guardrails maps these controls in detail. Supervision depends on who built the agent: if a model provider builds an agent on its own general-purpose model, the AI Office supervises it in most cases; an agent your company builds on a third-party model stays with national authorities.²

EU AI Act penalties

Fines reach €35 million or 7% of total worldwide annual turnover, whichever is higher, for prohibited practices, and €15 million or 3% for most other breaches, including Article 50.¹ The penalty regime has applied since August 2, 2025, and Member States set the detailed rules.¹

Authorities weigh gravity, duration, the people affected, company size and cooperation.¹ Until December 2, 2027, the live exposure for most enterprises is the prohibitions and Article 50.

An EU AI Act readiness plan in five workstreams

Readiness takes five workstreams, run in this order because each depends on the one before.

1. 1Inventory and classify List every AI system you build, buy or embed, including agents and AI features in vendor products. Record your role, the intended purpose, the Annex III area and any Article 6(3) reasoning; counsel signs off the classification.
2. 2Close what is already due Screen every system against Article 5, including the two bans that apply from December 2, 2026. Ship Article 50 disclosure and output marking, and record the AI literacy measures you run for staff.
3. 3Put the evidence in the pipeline Generate traces, evaluation results against declared thresholds and dataset records from the system on every release, so the technical file stays current.
4. 4Design oversight and incident response Build the override, review and stop controls; name the people who hold oversight and train them; write the serious-incident runbook against the 15-day, 10-day and two-day clocks.
5. 5Fix contracts and change control Write information and access terms into supplier contracts. Record for every release whether it is a significant change, and pre-determine the retraining you will assess at conformity assessment. Appoint an EU authorized representative if you provide high-risk systems from outside the EU.

The Omnibus bought time for high-risk systems. A company that uses the next 14 months to build traces, evaluation gates, oversight controls and supplier terms into its delivery pipeline will reach December 2027 with the technical file written by its own systems. A company that waits will write that file by hand, under a deadline, for a system it can no longer freeze.