Provisional is not law

On 19 November 2025 the European Commission published the Digital Omnibus on AI, proposing among other things to move the compliance deadline for high-risk systems under Annex III from 2 August 2026 to 2 December 2027. A great many organisations read the headline, concluded they had gained sixteen months, and reallocated the budget.

Six months later, here is the actual state of that proposal. The second trilogue between Parliament, Council and Commission on 28 April 2026 ended without agreement. A provisional agreement was reached on 7 May 2026. It still requires formal adoption by both institutions before it changes anything.

A provisional agreement is a statement of intent between negotiating institutions. It is not a legal instrument, it is not in the Official Journal, and it is not a defence. Until adoption completes, the operative date in law remains 2 August 2026.

What is not being deferred at all

The deferral debate concerns Annex III high-risk obligations. It has never covered everything, and the parts that were never in question are the parts arriving first.

The transparency duties under Article 50 still apply from 2 August 2026. Disclosing to a person that they are interacting with an AI system. Marking synthetic audio, image, video and text in a machine-readable way. Labelling deepfakes. Obligations on general-purpose AI models are also on their own track, with the Commission’s enforcement powers arriving on the same date.

These are cheap to implement and conspicuous to omit. A missing chatbot disclosure is not a subtle finding buried in a technical file— it is visible to any regulator, journalist or competitor who opens your product in a browser. Organisations that deferred their whole AI Act programme on the strength of a headline about high-risk systems have, in many cases, also deferred the one set of obligations that was never delayed and takes a fortnight to satisfy.

Read the reason for the delay, not just the date

The Commission was reasonably candid about why it proposed to move the deadline. The harmonised standards that high-risk providers are supposed to conform to were not ready. National competent authorities had not all been designated. Notified bodies— the organisations that perform conformity assessments— were not in place in sufficient number. The proposal even contemplated linking the timeline to the availability of standards rather than to a fixed calendar date.

That is not a regulator concluding the obligations were unreasonable. It is a regulator conceding that the machinery for demonstrating compliance does not yet exist. The obligations are not being softened. The queue is being rescheduled.

Which produces a consequence worth thinking through carefully. When conformity assessment capacity finally exists and a fixed date does arrive, every deferred organisation in Europe reaches for it in the same quarter. Notified body capacity is finite, slow to build, and cannot be bought at short notice. Under those conditions your position in the queue is a commercial variable, not an administrative detail. Firms that prepared during the delay will be assessed. Firms that waited will be queuing behind them, explaining the wait to customers.

You may be in scope without selling into Europe

A point worth making because it is routinely missed in US and UK boardrooms: the AI Act reaches providers and deployers established outside the Union where the output produced by the system is used within it. Not where the company is incorporated, not where the model is hosted, not where the contract is signed— where the output lands.

That catches a lot of firms who have concluded the file is not theirs. A model scoring applicants, some of whom are in the EU. A system whose output reaches an EU subsidiary, an EU customer, or an EU employee. A product sold to a multinational that deploys it across its European operations. In each case the obligations can attach even though nothing about the business feels European.

The practical failure here is not defiance, it is scoping. Organisations run a careful assessment of their EU-facing product line and never examine the internal HR tool, the vendor-supplied fraud engine, or the analytics platform quietly processing European data. Those are the systems most likely to fall into Annex III categories, and they are the ones nobody has inventoried because nobody thinks of them as AI products at all.

The asymmetry that settles the decision

Strip away the legislative detail and the choice facing an executive is simple, because the outcomes are not symmetric.

The work the deferral postpones— system inventory and classification, risk management processes, technical documentation, data governance records, human oversight design, post-market monitoring, logging built to survive an audit— is realistically twelve to eighteen months of effort for an organisation of any size. That is longer than the reprieve being debated.

If the deferral is adopted and you started anyway, you spent the time building capability you were always going to need, and you are early. If the deferral is adopted and you waited, you have consumed most of the extension before beginning. If the deferral is not adopted, or is adopted in narrower form than the headline suggested, and you waited— you are non-compliant on a date that was in law the entire time, and your defence is that you were relying on a press release.

There is no branch of that tree where starting now is the wrong decision. That is unusual in regulatory planning and it is worth naming plainly, because it means this is not a judgement call about legislative forecasting. It is a resourcing decision that can be made without knowing the outcome.

The deadline your customers are enforcing

There is a further reason the statutory date is the wrong thing to plan around: for most firms it is no longer the binding constraint.

Procurement questionnaires from large regulated buyers already ask for AI system inventories, risk classifications, and human oversight descriptions. Insurers underwriting AI liability ask for validation boundaries and evidence of testing before they will quote. Enterprise security reviews now routinely include model provenance and data-handling questions that did not exist eighteen months ago. None of those parties adjusted their timelines when Brussels debated its own.

That is the practical shape of it. Your effective compliance date was set by your largest customer’s vendor assessment cycle, and it has almost certainly already passed.

What to do that survives either outcome

Start with inventory, because every other obligation depends on it and it is delay-proof. You cannot classify, document, or monitor systems you cannot enumerate, and most organisations genuinely do not know how many AI systems they operate once you count models embedded in purchased software. This work is worth doing even if the AI Act vanished tomorrow.

Then classify honestly against Annex III rather than optimistically. The categories that catch people are not the obvious ones: creditworthiness assessment, employment screening, access to essential services, and safety components in critical infrastructure sweep in a great deal of ordinary enterprise software that nobody thinks of as high-risk AI.

Then generate documentation as a by-product of building rather than as a retrospective exercise. Technical documentation assembled eighteen months after the fact is archaeology, it is expensive, and it is frequently wrong. The same applies to logging: post-market monitoring obligations assume you retained records you are probably discarding today, and no amount of later diligence recovers data you did not keep.

None of that requires knowing what Brussels decides. It requires accepting that the deadline was never really the point— the capability was.