EU AI Act Amendments 2026: What Regulation (EU) 2026/1744 Changed

What actually happened

The European Parliament and the Council adopted Regulation (EU) 2026/1744 on 8 July 2026 in Strasbourg. It was published in the Official Journal on 24 July and entered into force on 27 July, on the third day after publication rather than the customary twentieth, because 2 August was too close to leave the text hanging. Its formal job is narrow: amend Regulation (EU) 2024/1689 — the AI Act — along with the Basic Aviation Regulation and the Machinery Regulation. Everyone calls it the Digital Omnibus on AI.

A regulation amending a regulation applies directly in all 27 member states. No transposition step, no national implementing act to wait for, no drift between Dublin and Warsaw. The AI Act you read last month is a different document today.

It is not a repeal, and it is not the deregulation some of the lobbying asked for. The risk tiers survived. The prohibited-practice list got longer, not shorter. What moved was the calendar for the single most expensive chapter, plus a handful of obligations that got cheaper to discharge.

The one date that did not move: 2 August 2026

This is the part the delay headlines bury, and it is the part that lands first. Article 50, the transparency chapter, still applies from 2 August 2026. If you are reading this the week it published, that is days away, not quarters.

• Tell people they are talking to a machine. An AI system interacting with natural persons has to make that clear, unless it is obvious from the context.
• Mark synthetic content in a machine-readable way. Audio, image, video and text generated or manipulated by AI must carry a marking a machine can detect.
• Label deepfakes. Content that resembles real people, places or events and could mislead has to be disclosed as artificially generated.
• Disclose AI-generated text published to inform the public on matters of public interest, unless a human took editorial responsibility for it.
• Emotion recognition and biometric categorisation: deployers must inform the people exposed to the system.

One concession was granted. Article 50(2) — the machine-readable marking of synthetic output — gets a grace period to 2 December 2026 for systems already on the market before 2 August 2026. New systems get no such runway. The rest of Article 50 arrives on 2 August with no transition at all, and the new enforcement articles switch on the same day.

The nine changes that matter

Original law on the left, amended position on the right. Article numbers are to the AI Act as amended.

• 1. High-risk timing (Article 113). Was: stand-alone Annex III systems from 2 August 2026, AI embedded in regulated products under Annex I from 2 August 2027. Now: 2 December 2027 for Annex III, 2 August 2028 for Annex I Section A. Legacy high-risk systems already in use by public authorities get until 2 August 2030. The Commission's November 2025 draft had tied these dates to a conditional standards-readiness trigger; the final text dropped it, so these are unconditional calendar dates that cannot slip again without a fresh legislative procedure.
• 2. New prohibitions (Article 5). Was: eight prohibited practices, applicable since 2 February 2025. Now: two more — generating or manipulating intimate material without freely given, specific, informed, unambiguous and explicit consent, and child sexual abuse material. Applicable from 2 December 2026, in the top penalty tier of up to €35 million or 7% of worldwide annual turnover. New Article 5(1a) splits provider and deployer liability: a provider is caught where that generation is the intended purpose, or is reasonably foreseeable and reproducible without significant technical modification. Providers are expected to document real technical safeguards — refusal training, prompt guardrails, content filtering, abuse detection.
• 3. Transparency marking (Article 50(2)). Was: applicable 2 August 2026, no transition. Now: same date for new systems, with a grace period to 2 December 2026 for systems placed on the market before 2 August 2026.
• 4. AI literacy (Article 4). Was: providers and deployers shall ensure a sufficient level of AI literacy among staff. Now: take measures to support the development of AI literacy, with explicit language that no specific level has to be guaranteed. Softer in substance, and it applies from 27 July 2026. It still applies to deployers, which means it applies to law firms.
• 5. Bias-detection data (new Article 4a). Was: a narrow legal basis for high-risk providers to process special categories of personal data to detect and correct bias. Now: extended to all providers and all deployers of AI systems and general-purpose models, under a strict-necessity standard with safeguards. If GDPR was blocking your fairness testing, this is the unlock.
• 6. High-risk classification and machinery (Articles 6, Annex I). Was: a broad safety-component test that pulled in a lot of ordinary product AI. Now: new Article 6(1a)–(1c) excludes systems used for non-safety functions — user assistance, performance optimisation, service efficiency, automation, convenience, quality control — and reinstates an objective criterion for what counts as a safety component. AI in machinery moved from Annex I Section A to Section B, so Articles 9–15 and 17–25 no longer apply to it directly; equivalent requirements come via delegated acts under the Machinery Regulation, due by 2 August 2027.
• 7. Registration survived (Article 6(3), Annex VIII). Was: providers self-assessing a system as not high-risk must still register it in the EU database. The Commission proposed deleting this. Now: kept, on the reasoning that it is crucial for market surveillance, with two data points removed from Annex VIII Section B to slim the filing. Authorities keep their access.
• 8. Sandboxes (Article 57). Was: member states had to have at least one AI regulatory sandbox operational by 2 August 2026. Now: 2 August 2027, plus a new Union-level sandbox run through the AI Office with priority access for SMEs.
• 9. Enforcement and penalties (Articles 75, 75a–75d, 99). Was: supervision split, with the AI Office focused on general-purpose models. Now: the AI Office gets exclusive competence over AI systems built on a general-purpose model by the same undertaking, and over AI integrated into very large online platforms and search engines. New Articles 75a–75d give it investigative powers including on-site inspections and sealing premises, the ability to accept binding commitments, a non-compliance procedure with periodic penalties of up to 5% of average daily turnover, a five-year limitation period, and defence rights. Article 99 gains a capped fine tier for small mid-caps (a new Article 3 definition, per Commission Recommendation (EU) 2025/1099) and explicit room for warnings and non-monetary measures. Simplified quality-management obligations widened from micro-enterprises to SMEs generally. All from 2 August 2026.

The EU just named agentic AI, and did not define it

The most interesting line in this regulation is not in any Article. It is in a new Annex XIV, a table of nomenclature codes used to designate what a notified body is competent to assess. Alongside vertical codes for application areas, the amendment introduces horizontal AIH codes for the underlying technology. AIH 0401 is agentic AI.

As far as we can tell, that is the first time the phrase appears in binding Union law. Its sibling code is the tell: AIH 0205 covers AI systems that learn from their environment, excluding agentic AI. Someone drafting a table of administrative codes decided agentic systems were a distinct enough class to carve out.

Be careful about what this is. An administrative code is not a legal definition. There is no Article defining an agent, no risk tier keyed to autonomy, no obligation that scales with how many steps a system takes without a human in the loop. Union law has written the word down and left the substance for later.

That gap is the whole problem for anyone building agentic systems in a regulated field. The obligations that will eventually attach are being drafted while the systems are being shipped, which means the only defensible position is to constrain what an agent is able to do rather than to check what it did. That is the argument we made in governance by construction, and it is why Justinian scopes an agent's action space to a jurisdiction before it is allowed to act rather than filtering the answer afterwards. Thomson Reuters put agentic adoption at 15% of professionals in 2026, with 77% expecting it to be central to their workflow by 2030. The rules will land in the middle of that curve.

What changes for a law firm that uses AI

Nearly every client alert on this regulation is written for a compliance officer at a company that builds AI. If you are a firm that uses it, the vocabulary is against you, so start with the distinction. A provider develops an AI system and puts it on the market under its own name. A deployer uses one under its own authority. Most law firms are deployers, and deployer duties are thinner but they are not zero.

• AI literacy is yours, and it is live. Article 4 binds deployers as well as providers. The amendment softened it to supporting the development of literacy rather than guaranteeing a level, effective 27 July 2026. That is a lower bar, not an absent one, and there is still no exemption for small firms.
• If you screen CVs with AI, you deploy a high-risk system. Recruitment sits in Annex III point 4. The obligations that came with that were due 2 August 2026 and are now due 2 December 2027. Sixteen extra months is the single most valuable thing this regulation gives an ordinary law firm.
• If you publish AI-drafted commentary, disclose it. Article 50(4) covers text published to inform the public on matters of public interest, and the exemption runs on human editorial responsibility. Firm-blog content generated at volume without a named human reviewer is exactly the case it describes. From 2 August 2026.
• If you run a client-facing chatbot, say it is a machine. Article 50(1), from 2 August 2026, unless it is obvious.
• If you build internally, you may have become a provider. Put a tool into service under your own name and the full provider stack can attach, including registration where Article 6(3) applies. Firms building their own workflow layers on top of a model should get this assessed rather than assumed.
• Fairness testing got easier. New Article 4a gives deployers a lawful basis to process special-category data strictly to detect and correct bias. Useful if a client has been asking whether your intake triage disadvantages anyone and GDPR was the reason you could not check.
• Extraterritoriality has not changed. Article 2 reaches providers and deployers outside the Union where the output is used inside it. A firm in Dubai or Beirut serving EU clients is inside the perimeter. The Act follows the output, not the postcode.

Eight questions for your legal AI vendor

We are a legal AI vendor, so read these knowing we have to answer them too. That is rather the point — every one of these has a specific, checkable answer, and a vendor that cannot give you one is telling you something.

• For anything we do with you, are you the provider of an Annex III high-risk system, and which point of Annex III?
• Which Article 50 obligations does the product discharge for us, and which stay with us as deployer? Name them individually.
• Do you mark AI-generated output in a machine-readable way today, and if not, will you before 2 December 2026?
• Where Article 6(3) self-assessment applies to you, are you registered in the EU database, and can we see the entry?
• What does your Article 4 support consist of — actual training and materials, or a PDF and a webinar?
• Which of the two new Article 5 prohibitions could our use of your tool foreseeably trip, and what technical safeguards do you document against them?
• Who is your authorised representative in the Union, and which market surveillance authority supervises you?
• If your system is built on your own general-purpose model, is the AI Office now your exclusive supervisor under the rewritten Article 75?

What did not change

• General-purpose AI obligations. Articles 51 to 55, in force since 2 August 2025, untouched. Systemic-risk thresholds and model-provider duties stand.
• The original prohibitions. Social scoring, subliminal manipulation, certain biometric categorisation and the rest have applied since 2 February 2025 and were not reopened.
• Annex III itself. The classification list is unchanged. A system that was high-risk in June is still high-risk. Only the date it has to comply by moved, and assessments already done still stand.
• The top and mid penalty tiers. Up to €35 million or 7% for prohibited practices, and the tier below it, are as they were.
• The architecture. Risk tiers, conformity assessment, the notified-body route, the database, market surveillance. The EU is still the only jurisdiction on earth with a comprehensive horizontal AI law, and this amendment leaves that intact.

Our read

The delay is real and it is narrow. One chapter moved, for good reasons that have more to do with the absence of harmonised standards and conformity-assessment infrastructure than with any change of heart. Reading this as the EU backing off gets the next two years wrong.

For vendors, the pressure went up. From 2 August 2026 the AI Office can inspect premises and seal them, impose periodic penalties, and take binding commitments. Registration survived a deletion attempt. Machine-readable marking of synthetic output becomes a product requirement with a hard date. None of that is a lighter touch; it is the same substance with better enforcement and a later invoice on the most expensive part.

The dropped conditional trigger is the underrated detail. The November draft would have let the high-risk dates float on standards availability. The final text fixes them. Anyone who built a plan around further slippage should rebuild it: 2 December 2027 and 2 August 2028 now require a whole new legislative procedure to move, and nobody wants to run this one twice.

We are not an EU-established company. We build from Beirut and our clients are concentrated across the Middle East and Europe, which puts us squarely inside Article 2 and outside the comfort of thinking this is somebody else's regulation. So we read the consolidated text on the day it published rather than waiting for a summary, and we would rather publish the dates plainly than sell anyone a compliance panic. The honest position on 28 July 2026 is that most firms have one deadline four days out and one twenty-eight months out, and confusing the two in either direction is the expensive mistake.

• Regulation (EU) 2026/1744 on EUR-Lex
• European Commission AI Act Service Desk
• Freshfields — the final Digital Omnibus on AI
• Gibson Dunn — postponed high-risk deadlines and other key changes
• Thomson Reuters — 2026 AI in Professional Services Report

• Governance by construction — constraining what an agent can do
• How HAQQ handles security and compliance
• Justinian — jurisdiction-scoped legal AI
• Why human review is a designed feature
• How to become an AI-native law firm