06-reference/research

eu digital omnibus ai official journal check

2026-08-15·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
eu-ai-actregulationdigital-omnibuscompliance-datesfact-check

Yes and Yes: The Digital Omnibus on AI Published 24 July 2026, and the 2 December 2027 Annex III Date Held

Not legal advice. This is a date-verification pass by a non-lawyer agent. Every date below is tagged with the evidence tier it rests on. Nothing here should be relied on for compliance.

The question

"Did the EU Digital Omnibus on AI publish in the Official Journal before 2 August 2026, and did the adopted text preserve the 2 December 2027 Annex III date?"

Derivative follow-up from [[research/2026-08-02-agents-as-employees-regulatory-labor-framing]], which left both halves explicitly unverified because every source in that pass was written before publication. Filed HIGH rather than the default Medium: the answer determines whether the parent brief's compliance dates are correct.

Answer, up front — both halves resolve clean:

Question Answer Evidence tier
(a) Published in the OJ before 2 Aug 2026? Yes — 24 July 2026, nine days ahead of the deadline Primary (EUR-Lex)
(b) Annex III date preserved at 2 Dec 2027? Yes — unchanged from the political agreement Primary (recital 40, OJ text)
(c) OJ citation OJ L, 2026/1744, 24.7.2026 Primary (EUR-Lex)

What we already know (from the vault)

What the web says

Caveat on evidence depth, stated plainly. Two attempts to read the enacting article that amends Article 113 (the full omnibus text, then the consolidated AI Act) both truncated before reaching it — these are long documents and the fetch returned only the earlier portion each time. So the 2 December 2027 date is confirmed from the published primary text at recital level, corroborated by the consolidation metadata and by four independent law firms, but the operative Article 113(3) wording was not read character-for-character in this pass. That is a real limit. It is a narrow one — recital 40 is part of the same OJ-published instrument and states the date explicitly — but a residual conditional mechanism inside the operative article cannot be excluded to a 100% standard. Follow-up #1 closes it.

Convergences and contradictions

Superseded dates still circulating — correction list

Flagged so the parent brief and any future pass can route around them:

  1. 2 August 2026 as the Annex III high-risk date — superseded for standalone Annex III systems. Still recited by artificialintelligenceact.eu's staffing page (caught in the parent pass). Note this date is not wholly dead: Article 50 transparency obligations did go live 2 August 2026 unchanged, so the date remains correct for transparency and wrong for high-risk.
  2. "Expected before 2 August 2026" / "pending OJ publication" — Gibson Dunn, 27 May 2026. Correct when written, stale now. Anything carrying this framing predates 24 July.
  3. "Proposed deferral" / "Regulation Proposal" framing — DLA Piper GENIE and the Commission's own digital-strategy library page. The instrument is adopted law, not a proposal.
  4. "Conditional on harmonised standards" — an inference gloss appearing in search summarization, not supported by recital 40. The standards gap is the stated reason for the date, not a condition attached to it.

Corrections owed to the parent brief

[[research/2026-08-02-agents-as-employees-regulatory-labor-framing]] needs two edits, both marked "unverified" in the original and now resolvable:

Synthesis for RDCO

The headline is that a speculative compliance clock became a citable one, and RDCO's positioning read survives intact. The parent brief's recommendation — keep the "AI workforce" framing, drop the employee/payroll register, treat Annex III exposure as a product-scope question — rested on what Annex III governs, not on when it bites. Publication does not disturb any of that. What it does disturb is the confidence tier. Thirteen days ago the correct RDCO sentence was "the heavy regime is expected to slide to late 2027, pending publication." Today it is "standalone high-risk obligations under Annex III apply from 2 December 2027 under Regulation (EU) 2026/1744, published 24 July 2026." That is the difference between hedged commentary and a fact with a citation, and in enterprise procurement conversations the second one is worth considerably more than the first.

The more interesting finding is the lag, because the lag is the opportunity. A regulation entered into force on 27 July 2026 and, three weeks later, the Commission's own library page still frames it as a proposal, a major law firm's employment note still says "proposed deferral," and the most-trafficked free AI Act reference still recites the superseded date on at least one page. The parent brief spotted one instance of this and treated it as a footnote. It is not a footnote, it is the pattern: the public compliance layer for European Union (EU) AI regulation is running weeks behind the law, and most people forming views right now are reading stale pages. That is precisely the "small, real piece of expertise that most people in the category get wrong" the parent brief identified as cheap credibility, except it is cheaper and more concrete than the parent brief realized. Knowing that Annex III moved is mildly useful. Knowing the OJ citation, the in-force date, and which widely-cited sources are still wrong is the kind of specificity that reads as someone who actually checked.

Two operational consequences, both small and both real. First, Article 50 transparency is now a live obligation rather than a forthcoming one, with the Article 50(2) watermarking grace running to 2 December 2026. For a firm whose pitch is that agents do the work, the never-let-an-agent-pass-as-human rule is now current-tense in any EU-facing interaction, and the four-month watermarking runway is the only near date on the calendar. Second, the 2 December 2027 Annex III date is far enough out that the parent brief's scope warning stays prospective: if an RDCO-built agent ever screens candidates, allocates work to humans by traits or behavior, or monitors human performance for a client with EU workers, that deployment lands in Annex III, and there are now sixteen months of runway rather than an expired deadline. As a US solo consultancy with no known EU-worker-facing deployments, RDCO's own compliance burden today remains zero. This matters as client-facing content and as a procurement answer, which is exactly where the parent brief placed it.

The methodological note worth keeping. This brief cost one primary source read and resolved two flags the parent pass had to leave open. Both halves confirmed, in the direction the parent guessed but declined to assert. The cheap, time-decaying factual check on a prior brief's explicit unknown is a high-yield shape for the research queue, and the reason it worked is that the parent brief wrote its unknowns down as specific, checkable questions rather than burying them in hedged prose.

Why this is in the vault

Discharges the top open follow-up of [[research/2026-08-02-agents-as-employees-regulatory-labor-framing]] with a primary EUR-Lex citation, and supplies the exact regulation number, OJ citation, and in-force date that any Sanity Check piece on AI regulation or any enterprise-procurement answer about EU exposure will need to cite rather than hedge. It also carries the specific edit list needed to correct the parent brief's two unverified markers.

Open follow-ups

Related

Sources

Vault:

Web — PRIMARY (fetched and read):

Web — SECONDARY (corroboration only, surfaced in search, not fetched):

Web — SUPERSEDED OR STALE (do not cite for dates):

Paywalls encountered: none.

Research budget used: QMD 3 queries (1 failed on the hyphen-negation parse, 2 successful), WebSearch 2, WebFetch 3 (cap reached — this is the binding constraint on follow-up #1), vault docs read 1 full + 2 via search snippets.