One deadline moved. The other one is already here.

Watch what a deferred deadline does to a programme plan. The governance workstream was never the part anyone wanted to fund — it produced no feature, demoed badly, and survived reviews only because a date in August made it non-optional. Move the date sixteen months and the workstream does not get cancelled. It gets re-baselined into the next budget cycle, which in a large organisation is a slower way of saying the same thing.

That re-baselining has been happening across Europe for about three weeks now, and we think most of it is reading the regulation wrong.

On 28 July, Hunton Andrews Kurth reported that Regulation (EU) 2026/1744 — the Digital Omnibus on AI — was published in the Official Journal on 24 July and entered into force on 27 July, six days before the AI Act's original high-risk deadline. Stand-alone Annex III high-risk systems now apply from 2 December 2027. High-risk AI embedded in regulated products moves to 2 August 2028. Gibson Dunn's 27 May account of the political agreement carries the detail most summaries drop: the obligation on member states to stand up regulatory sandboxes slid as well, to 2 August 2027. The mechanism meant to help firms prepare moved with the thing it was meant to prepare them for.

Our read is that the omnibus deferred the audit and left the exposure where it was — and that the one obligation which did land on 2 August is the one that covers what enterprises actually put into production this year.

The article nobody re-baselined

Article 50 was not deferred. Since 2 August it has required that a person interacting directly with an AI system be told so, unless that is obvious to a reasonably informed person; that synthetic audio, image, video and text be marked in a machine-readable format and be detectable as artificially generated or manipulated; and that deployers of emotion-recognition or biometric-categorisation systems inform the people exposed to them. The disclosure has to arrive, in the regulation's own phrasing, in a clear and distinguishable manner at the latest at the time of the first interaction.

Now hold that against the inventory of what a 2026 enterprise AI programme genuinely shipped. Not the Annex III credit model — that is still in design review, which is roughly why the deadline moved. What shipped was a voice agent on the service line. A drafting assistant that produces member correspondence. A summariser whose output ends up in a letter that a real person opens. Those are the live systems, and they sit squarely inside the article that stayed.

Which produces an uncomfortable inversion. The obligations that got deferred attach mostly to systems still on the roadmap. The obligation now in force attaches to the ones already answering calls.

The obvious objection

The reasonable pushback is that sixteen months is a long runway, the harmonised standards are not finished, and building compliance machinery against an unfinished target wastes money that could ship product. Wait for the standards, then build once, correctly.

That argument depends on the standards arriving on a schedule, and the co-legislators specifically declined to make that true. Pinsent Masons reported on 7 May that the agreed deferral uses fixed dates rather than backstop dates tied to the availability of guidance — a change from the Commission's original proposal. Read that carefully. The application date no longer moves with the readiness of the tooling in either direction. If the standards land late, the date does not follow them. The organisation that decided to wait has committed to a plan whose first dependency has no delivery date, and whose second dependency is a fixed one.

The deeper problem is that a chunk of the work is not policy work, and cannot be compressed. You can write a governance policy in a quarter. You cannot write a year of logs in a quarter. Technical documentation and record-keeping obligations are questions about what a system did — which inputs, which version, who reviewed the output, what it said on a Tuesday in March. A system that recorded none of that between August 2026 and mid-2027 cannot produce it retroactively, and the honest answer to the auditor is that the period is dark. That is not a compliance gap that a consultant closes in the final quarter. It is missing data.

There is a smaller, sharper version of the same point in the same reporting: providers must register high-risk systems in the EU database even where they believe an exemption applies. The exemption still costs you the filing, and the filing presumes you know which systems you have. Most enterprises we talk to could not produce that inventory today, and the inventory is not a 2027 artefact — it is the input to every decision about which of these dates even apply to you.

The list is the same list

Strip the regulation off entirely and look at what high-risk compliance actually asks for: a named human who exercises oversight at a specific step, logging sufficient to reconstruct a decision after the fact, documentation of what the system is for and where it is known to fail, and a way to demonstrate that the thing running in production is the thing that was reviewed.

That list does not change when you delete the word Brussels from it. It is what you need to operate an agent anywhere near a member, an applicant or an employee, and to answer for the answer it gave when someone contests it. We have argued this in other contexts — the record the workflow leaves behind is not a compliance deliverable bolted on at the end, it is the thing that makes the workflow operable at all — and the omnibus is a good test of whether an organisation believed that or was simply complying.

The tell is what happens to the workstream now. If it was scoped as evidence-production for an auditor, it dies in the re-baseline and gets rebuilt from nothing in 2027. If it was scoped as instrumentation for an operation, the date moving is irrelevant, because the logs were never for Brussels.

What we would ask this quarter

Three questions, and none of them wait for a standard.

• Which systems that are live today talk directly to a person, or generate content a person will read as though a person wrote it? That set is under an obligation now, not in 2027.
• If a determination made by one of those systems is contested in eighteen months, what can we replay — the output, or the run? An output without the run is a claim, not a record.
• Who is the named human at the oversight step, and is that person's sign-off enforced by the workflow or by a policy document? A gate that software can route around has already been routed around.

The deferral bought sixteen months. It is worth noticing what it bought sixteen months of. Not silence — engineering time, on a system that will be asked, on 2 December 2027, to account for how it behaved during them.