Annex III Was Never Touched: The Omnibus Moved the Clock, Left the Employment List and the Filter Verbatim, and Deleted Exactly Two Lines of Registration Paperwork

Not legal advice. This is a text-verification pass by a non-lawyer agent. It reports what the enacting words say and what the amendment markers show, not what any of it means for a particular product.

The question

"Did the EU Digital Omnibus change Annex III's employment-context entries themselves, and what became of the Article 6(3) filter derogation and the Annex III registration duty, or did it move only the applicability date?"

Follow-up carried forward twice — raised in [[2026-08-15-eu-digital-omnibus-ai-official-journal-check]], re-flagged as the highest-value open item in [[2026-08-21-ai-act-article-113-operative-text-check]]. The grandparent brief [[2026-08-02-agents-as-employees-regulatory-labor-framing]] built its entire scope read on what Annex III's employment entry covers, and no pass had yet read the amended text.

Answer, up front.

What we already know (from the vault)

• The date is settled and is not the issue. [[2026-08-21-ai-act-article-113-operative-text-check]] read Article 113 third paragraph point (c) in the authentic Official Journal text: 2 December 2027 for Article 6(2)/Annex III systems, unconditional, no trigger, no Commission power to move it administratively. That branch is closed.
• The grandparent's positioning read hangs on Annex III point 4. [[2026-08-02-agents-as-employees-regulatory-labor-framing]] concluded that Annex III governs artificial intelligence (AI) used to decide about human workers, not AI that performs work, and that RDCO's only realistic exposure is a product-scope question: if an RDCO-built agent screens candidates, allocates work to humans by traits or behaviour, or monitors human performance for a client with European Union (EU) workers, that deployment lands in Annex III. This brief tests whether the underlying list still says what that read assumed.
• The verification standard demands the primary source. [[2026-05-19-verification-as-independent-worker-pattern]] requires at least one gate in a chain to hit primary text. Two prior passes failed on WebFetch truncation before reaching enacting articles.
• The curl-and-grep route is the one that works. Established on the 2026-08-21 run and reused unchanged here: fetch to disk, strip tags locally, grep. Three documents, 853 KB / 353 KB / 1.26 MB of HTML, zero truncation.

What the web says

Method. Three primary documents were fetched with curl to disk and read locally: the authentic OJ text of Regulation (EU) 2026/1744 (OJ:L_202601744, 353 KB), the consolidated AI Act at CELEX 02024R1689-20260727 (853 KB), and the original AI Act at CELEX 32024R1689 (1.26 MB). The consolidation's header confirms Regulation (EU) 2026/1744 of 8 July 2026 is the only amending act on record — marker M1, no M2. Annex III has therefore never been amended by anything, Omnibus or delegated act.

• Annex III is untouched, confirmed three independent ways. (1) Marker scan: the consolidated text carries 130 ▼B / ▼M1 amendment markers; the last one before Annex III is a ▼B inside Annex II, and the next is inside Annex VIII. The whole of Annex III sits in original-text territory. (2) Amendment inventory: Article 1 of Regulation (EU) 2026/1744 contains exactly 43 numbered amendment points. Point (41) amends Annex I, point (42) amends Annex VIII, point (43) adds a new Annex XIV. No point amends Annex III, and no point amends Article 7 (the Commission's delegated power to add or modify Annex III use cases). (3) Literal diff: the Annex III block in the original 2024 OJ text and the Annex III block in the 27 July 2026 consolidation are character-identical.
• The employment entry therefore still reads exactly as the grandparent brief described it. Annex III point 4, "Employment, workers' management and access to self-employment," verbatim: "(a) AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates; (b) AI systems intended to be used to make decisions affecting terms of work-related relationships, the promotion or termination of work-related contractual relationships, to allocate tasks based on individual behaviour or personal traits or characteristics or to monitor and evaluate the performance and behaviour of persons in such relationships."
• The Article 6(3) filter derogation survives word for word. Article 6(2) through 6(8) are all original ▼B text. The derogation still turns on an AI system not posing "a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making," and still lists the same four gateway conditions: narrow procedural task; improving the result of a previously completed human activity; detecting decision-making patterns or deviations without replacing or influencing the prior human assessment; or performing a preparatory task. The hard carve-back is intact too: "Notwithstanding the first subparagraph, an AI system referred to in Annex III shall always be considered to be high-risk where the AI system performs profiling of natural persons." Article 6(4)'s duty to document the self-assessment before placing on the market, and to hand it to national competent authorities on request, is likewise unchanged.
• What the Omnibus did do to Article 6 is Annex-I-side only. Point (8) of Article 1 inserts three new paragraphs — 6(1a), 6(1b), 6(1c) — narrowing what counts as a "safety component" for product-embedded classification under Article 6(1): systems "solely used for non-safety related aspects of user assistance, performance optimisation, service efficiency, automation or convenience or quality control" do not qualify; systems whose failure "would endanger health and safety" do; and third-party conformity assessment triggered "solely due to risks other than risks to health and safety" does not satisfy Article 6(1)(b). Point (4) correspondingly rewrites the Article 3(14) definition of "safety component." None of this reaches Annex III standalone systems.
• The registration duty is intact; two data fields were deleted, and they are on the wrong register to matter for Annex III. Article 49 — all five paragraphs, including 49(1) (register the Annex III high-risk system), 49(2) (register the Article 6(3) self-declared not-high-risk system), 49(3) (public-authority deployer registration), 49(4) (secure non-public section for points 1, 6 and 7) and 49(5) (point 2 systems registered nationally) — is original ▼B text, unamended. Article 71 (the EU database) is unamended. Annex VIII Section A (what an Annex III high-risk provider files under 49(1)) and Section C (deployer filings under 49(3)) are unamended. The entirety of point (42) of Article 1 reads: "in Annex VIII, section B, points 7 and 9 are deleted."
• What those two deleted fields were. From the original text, Annex VIII Section B point 7 was "A short summary of the grounds on which the AI system is considered to be not-high-risk in application of the procedure under Article 6(3)"; point 9 was "Any Member States in which the AI system has been placed on the market, put into service or made available in the Union." Surviving point 6 — "The condition or conditions under Article 6(3) based on which the AI system is considered to be not-high-risk" — means the public register still shows which of the four gateways a provider claimed, just no longer why. Recital 22 states the intent plainly: "the registration of AI systems referred to in Article 6(3) ... should be simplified by streamlining the content required under Annex VIII," while "it remains crucial for effective market surveillance and public accountability that such AI systems are registered," and "a provider applying Article 6(3) ... remains obligated to document its assessment."
• One adjacent change worth logging: the fundamental-rights impact assessment got a paperwork discount, not a scope change. Point (13) amends only Article 27(4) and 27(5) — deployers may cross-reference or import sections of a data-protection impact assessment, and the AI Office must build a questionnaire template "including through an automated tool." Article 27(1)'s scope is untouched, so the FRIA still reaches Annex III deployers that are public bodies or private entities providing public services, plus points 5(b) and (c) deployers. A purely private employer deploying an Annex III point 4 system still owes no FRIA.

A textual oddity, flagged as reading rather than finding. The amended Article 113 third paragraph point (c) defers "Chapter III, Sections 1, 2, and 3." Article 49 sits in Chapter III Section 5 ("Standards, conformity assessment, certificates, registration," Articles 40-49), and Article 71 sits in Chapter VIII. Neither is named in the deferral, so on the face of the text both apply from the general 2 August 2026 date while the Article 6(2)/6(3) classification rules that trigger them do not apply until 2 December 2027. Note also that the original point (c) read simply "Article 6(1) and the corresponding obligations in this Regulation shall apply from 2 August 2027" — a broad wrapper phrase the replacement drops in favour of a section-by-section enumeration. Whether that leaves Section 5 as a null set (nothing can be a high-risk system before Section 1 applies) or as a live-but-empty duty is a genuine legal-interpretation question. I am not resolving it; I am recording that the enumeration does not cover Section 5 and that the change from "corresponding obligations" to a section list is real.

Convergences and contradictions

• Full convergence with the grandparent's scope read, now at enacting-text tier. [[2026-08-02-agents-as-employees-regulatory-labor-framing]] described Annex III's employment entry from a secondary source (artificialintelligenceact.eu) and drew the "decides about workers, not AI that does work" conclusion from it. The operative text confirms that description is accurate and complete, and confirms the list has not moved since. The one line in that brief worth tightening rather than correcting: it read the entry as covering "decisions on contract terms or termination" — the actual words also reach task allocation "based on individual behaviour or personal traits or characteristics," which is broader than a contract-decision framing suggests.
• A directional contradiction with the general "Omnibus loosened high-risk scope" narrative. The Omnibus did narrow high-risk classification — but exclusively on the Article 6(1)/Annex I product-embedded side, via the safety-component redefinition. On the Article 6(2)/Annex III standalone side it changed nothing substantive at all. Any commentary that says "the Omnibus narrowed high-risk scope" without splitting those two limbs is imprecise in a way that matters to anyone reasoning about employment-context AI.
• No contradiction between recital and enacting text this time. Recital 22 promised a simplification of Article 49(2) registration content while preserving the registration itself and the Article 6(4) documentation duty. Point (42) delivers exactly that and nothing more. Unlike the recital-40 case in the prior brief, the recital here is neither narrower nor broader than the operative words.

Synthesis for RDCO

The substantive answer is a clean negative, and a clean negative is worth more here than it looks. Three passes of this AI Act thread have now been spent chasing the possibility that the compliance surface under RDCO's positioning had moved. It has not. Annex III point 4 says today, character for character, what it said when the Act was signed on 13 June 2024, and the Article 6(3) filter that decides whether a borderline system escapes it is equally unmoved. Everything the Digital Omnibus did to high-risk classification landed on the product-safety limb — safety components inside machinery, radio equipment, medical devices — which is a limb RDCO will never touch. The clock moved; the map did not. The grandparent brief's scope warning can now be restated without any hedge at all: an RDCO-built agent that screens candidates, allocates work to humans by behaviour or traits, or monitors human performance for a client with EU workers lands squarely inside Annex III point 4(a) or 4(b), and the escape hatch it would have to fit through is the same four-condition filter that existed in 2024.

The part actually worth carrying into a client conversation is the shape of the filter, because it is the operative thing and almost nobody quotes it correctly. Two features do the work. First, the "always high-risk where the AI system performs profiling of natural persons" carve-back is absolute — it overrides all four gateway conditions, which means an agent that builds behavioural profiles of workers cannot filter out no matter how preparatory or narrow its task looks. Second, the gateway conditions are drafted around not materially influencing the outcome of decision making, and condition (c) explicitly fails where a system is "meant to replace or influence the previously completed human assessment, without proper human review." That is the same "material influence" test the grandparent brief picked up second-hand from a law-firm note, and it is genuinely in the statute. For a firm whose entire differentiator is "every action attributable to one accountable human operator," this is unusually convenient: the architecture RDCO already sells — human owner, agent recommends, human decides, decision logged — is the architecture the derogation was written to accommodate. That is a positioning asset, not just a compliance fact, and it is stated in words RDCO can quote.

The one operationally interesting change is the registration deletion, and it cuts against transparency in a direction worth noticing. Under the original Annex VIII, a provider who self-declared its Annex III system not-high-risk had to publish a short narrative summary of why into a publicly accessible EU database. That field is now gone; only the tick-box of which condition was invoked survives. The justification still has to exist — Article 6(4) and recital 22 both keep the documentation duty — it just stops being public and becomes something a national competent authority has to ask for. So the effect is a shift from published reasoning to on-request reasoning. If RDCO ever needs to diligence a vendor's Article 6(3) claim, or a client asks "how do we check whether our HR tool's not-high-risk claim is credible," the honest answer after 27 July 2026 is that the public register will not tell you and you have to ask the vendor for the assessment. That is a small, concrete, checkable piece of expertise of exactly the kind the 2026-08-15 brief identified as cheap credibility, and it is the sort of detail that will still be missing from most secondary coverage months from now.

Methodologically, this pass is the second clean data point for curl-and-grep, and it added a technique worth naming: amendment-marker attribution. Reading the consolidated text and mechanically mapping every provision back to its governing ▼B / ▼M1 marker turns "was this changed?" from a judgement call into a lookup, and cross-checking against the amending act's own numbered point list gives an independent second gate. The literal diff against the original 2024 text was the third. Three gates, all mechanical, all cheap, and the answer they produce — nothing changed — is the answer that a reading-based approach is most likely to get wrong, because absence of change is invisible to a reader skimming for changes.

Why this is in the vault

It discharges the last substantive open item in the AI Act thread and converts [[2026-08-02-agents-as-employees-regulatory-labor-framing]]'s scope analysis from secondary-source-tier to enacting-text-tier: the Annex III employment entries and the Article 6(3) filter that brief relied on are verified unamended as of the 27 July 2026 consolidation, so its "AI workforce" naming recommendation and its product-scope warning both stand without caveat. It also supplies the quotable Article 6(3) filter language RDCO would use in any EU-facing procurement or client conversation about an agent that touches worker-related decisions.

Open follow-ups

• Does the amended Article 113 actually leave Chapter III Section 5 and Chapter VIII undeferred, and if so what does that mean in practice? The section enumeration in point (c) does not name Section 5 (Articles 40-49, including registration) or Chapter VIII (Article 71, the EU database), and the replacement dropped the original's broad "and the corresponding obligations" wrapper. Either this is a null set by construction or there is a live gap. This is the only genuinely unresolved textual question this pass surfaced.
• Have the Article 6(5) classification guidelines actually been issued, and what do they say about borderline employment-context systems? Carried forward from [[2026-08-21-ai-act-article-113-operative-text-check]] and now more valuable, not less: with the filter text confirmed frozen, the guidelines are the only remaining thing that can move the practical boundary, and the Article 6(5) carve-out means they were due 2 February 2026 and are not deferred. The McCann FitzGerald "material influence" note the grandparent cited referred to draft guidelines — whether a final version exists is unverified.
• Does the deletion of Annex VIII Section B point 7 change what is publicly visible in the live EU database? The database is already operational for some registration categories. Whether existing Article 49(2) entries retain their grounds summaries or whether the field is being purged is an empirical question answerable by looking at the database, not the statute.

Related

• [[2026-08-21-ai-act-article-113-operative-text-check]]
• [[2026-08-15-eu-digital-omnibus-ai-official-journal-check]]
• [[2026-08-02-agents-as-employees-regulatory-labor-framing]]
• [[2026-05-19-verification-as-independent-worker-pattern]]
• [[2026-06-07-ai-workforce-positioning-map]]

Sources

Primary (fetched with curl to disk, read in full, no truncation):

• Regulation (EU) 2026/1744, authentic OJ text, OJ L, 2026/1744, 24.7.2026 — Article 1 points (4), (8), (13), (39), (40), (41), (42), (43); recitals 7 and 22 — https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202601744
• Consolidated Regulation (EU) 2024/1689, CELEX 02024R1689-20260727 — Annex III, Article 6, Article 49, Article 71, Annex VIII, Article 113; amendment-marker attribution across the full text — https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:02024R1689-20260727
• Original Regulation (EU) 2024/1689, CELEX 32024R1689 — Annex III and Annex VIII Section B as originally enacted, used for the literal diff — https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32024R1689

Vault:

• ~/rdco-vault/06-reference/research/2026-08-21-ai-act-article-113-operative-text-check.md
• ~/rdco-vault/06-reference/research/2026-08-15-eu-digital-omnibus-ai-official-journal-check.md
• ~/rdco-vault/06-reference/research/2026-08-02-agents-as-employees-regulatory-labor-framing.md
• ~/rdco-vault/02-sops/2026-05-19-verification-as-independent-worker-pattern.md

Not consulted: no secondary/law-firm sources were used in this pass. The primary text was reachable and dispositive on every sub-question, so corroboration would have added nothing. No paywalls were encountered.