Regulation

EU AI Act readiness for enterprises: what the December 2027 deferral did and did not change

Almost every EU AI Act page still ranking today was written against a deadline that no longer exists. Here is the timetable as it stands after Regulation (EU) 2026/1744, which obligations are live right now, and the 90 days of work the deferral did not remove.

By Brandon Da Costa, Founder, Digiton Dynamics

What actually changed on 27 July 2026

Regulation (EU) 2026/1744, the Digital Omnibus on AI, amended the AI Act. It was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026, six days before the deadline it postponed. The sequence, for anyone who has to write this into a board paper:

The Cloud Security Alliance titled its note on the change "deferred, not cancelled", and that is the whole reading in three words. The trilogue also added two new prohibited-practice categories covering non-consensual intimate imagery and AI-generated child sexual abuse material, with a grace period to 2 December 2026 for the technical safeguards.

The compliance calendar as it stands today

Penalty bands did not change. Up to EUR 35 million or 7 percent of global annual turnover for prohibited practices, and up to EUR 15 million or 3 percent for the remaining obligations, whichever is higher in each case.

What binds you in September 2026

Three things, and none of them are the ones the trade press has been writing about.

Article 5, since February 2025

The prohibitions have been enforceable for over eighteen months. Social scoring by public authorities, exploitation of vulnerability, untargeted facial-image scraping to build recognition databases, emotion inference in workplaces and schools. Two categories joined the list this summer. If any of these describe a system you run, the deferral is irrelevant to you.

General-purpose AI obligations, since August 2025

If your organisation places a general-purpose model on the EU market, Articles 51 to 56 have applied for a year. Technical documentation, a copyright policy, a public summary of training content, and systemic-risk obligations above the compute threshold. Most enterprises are deployers rather than providers here, but a fine-tuned model distributed to customers can put you in scope, and that determination is worth writing down rather than assuming.

Article 50 transparency, since 2 August 2026

This is the one that reaches the widest set of ordinary systems, and it went live on schedule five weeks ago. Three duties. Disclose to a person that they are interacting with an AI system, unless it is obvious from context. Mark synthetic audio, image, video and text in a machine-readable format that downstream systems can detect. Label deepfakes and AI-generated text published to inform the public on matters of public interest.

The chat assistant somebody added to a marketing site last spring is in scope. So is the summarisation feature inside a customer portal. Systems placed on the market before 2 August 2026 have until 2 December 2026 to meet the content-marking part, which is the deadline most organisations should currently be working to.

Why the deferral is not eighteen free months

Two arguments, one regulatory and one commercial.

The regulatory one is about ordering. Classification is the prerequisite for every later obligation. You cannot write Article 9 risk-management documentation, or Article 11 technical documentation, or design a human-oversight mechanism, until you know which of your systems are Annex III and which are not. That exercise is slow, it needs legal and engineering in the same room, and most organisations have not started. An appliedAI analysis cited in the Cloud Security Alliance research note of 13 March 2026 found that 40 percent of 106 enterprise AI systems examined could not be clearly classified under the Act's risk tiers. The same note reported that more than half of organisations have not established systematic inventories of the AI systems they operate. You cannot classify what you have not inventoried.

The commercial argument is simpler and it moves faster than the regulation. Enterprise procurement already asks about model provenance, data residency and human oversight, this year, in the first pass rather than at contract stage. A vendor who cannot answer loses the deal regardless of what any deadline says. Independent analyses summarised in the Cloud Security Alliance note put initial compliance cost at USD 8 to 15 million for large enterprises and USD 2 to 5 million for mid-size ones, which is a reason to start the inventory now rather than an argument for waiting.

There is also a supply constraint nobody talks about. Notified bodies for conformity assessment are thin on the ground, and the harmonised standards are still landing. Fifteen months out looks comfortable. Fifteen months out with a queue in front of the assessor does not.

Who is doing this work in Europe

Three different categories get conflated on almost every page that ranks for this query, and conflating them is how buyers end up hiring the wrong one. A certification body audits you and cannot also build your system. A Big Four practice can run the programme and will subcontract the engineering. A delivery firm builds the thing but may have no assurance standing at all. Work out which you are buying before the first call.

Disclosure, because it belongs next to the table rather than in a footnote: Digiton Dynamics wrote this page and appears in it. Every other entry is a competitor or a body we would send a client to, and for a formal ISO/IEC 42001 certification the certification bodies above are the right call and we are not.

Where ISO/IEC 42001 fits

Buyers keep asking whether certification satisfies the Act. It does not, and it is not meant to. ISO/IEC 42001 is an AI management system standard: it certifies that you run a governance process, with the policies, roles, risk assessments and internal audits that implies. The AI Act asks about specific systems and specific obligations. The overlap is real and useful, because the documentation an AI management system produces is most of what an Article 11 technical file needs, and procurement teams increasingly ask for the certificate as a shortcut. Treat it as scaffolding for compliance rather than as compliance.

A 90-day plan for the window

Fifteen months sounds like room. Split into the work that has to happen sequentially, it is not.

1. Weeks 1 to 3, inventory. Every AI system in the organisation, including the ones bought on a departmental card and the features that arrived inside a SaaS product you already licensed. Owner, purpose, data in, decision influenced, users affected. This is the step most organisations skip and the one everything else depends on.
2. Weeks 3 to 6, Annex III classification. Legal and engineering together, per system, written down with reasoning. Expect a meaningful fraction to sit in the ambiguous band. Record the ambiguity rather than resolving it optimistically.
3. Weeks 4 to 8, Article 50 audit. This one is live now, so run it in parallel. Every system a person can talk to, every piece of synthetic media you publish. Disclosure present, marking machine-readable, deepfake labelling where relevant. Fix before 2 December 2026 for anything placed on the market before 2 August 2026.
4. Weeks 6 to 11, documentation. Article 9 risk management and Article 11 technical documentation drafted for anything classified high-risk. Drafted, not perfected. A draft you can improve beats a blank file with a 2027 date on it.
5. Weeks 8 to 12, the procurement answer sheet. One document answering model provenance, hosting region, training-data exposure, log access and retention. Your customers are asking. Your sales team is currently improvising.
6. Weeks 10 to 12, the December 2026 safeguards. Technical measures for the two new Article 5 categories, if any system you run could produce that output.

Across 8 countries, Digiton's average time from a signed scope to a production system is 45 days, which is roughly the length of one Annex III classification exercise. That comparison is the point: the classification work is not large. It is only slow to start.

Where the engineering actually bites

Compliance documents describe outcomes. Systems have to produce them, and two places break more builds than the rest combined.

The first is erasure. A GDPR deletion request has to reach every derived copy of a record: the vector store, the cached retrieval context, the agent's conversation memory, and any evaluation set assembled from production traffic. Most retrieval builds delete the source row and leave the embedding in place, which means the system can still recite the deleted record. The mechanism that fixes it, source-linked chunk identifiers and a re-index path, has to be designed in rather than added.

The second is citation adequacy. Article 50 and the transparency obligations sit alongside an expectation that an organisation can show where an answer came from. A retrieval system that returns a fluent paragraph with no traceable source is not auditable, whatever the model scored on a benchmark. Both are covered in detail on GDPR-compliant RAG.

The governance layer around an agent is a third: scoped identity, least privilege, human approval on irreversible actions, an immutable audit log and a kill switch. That is on AI agent governance.

What this page will get wrong, and when

This document is accurate as of 1 September 2026 and it has a shelf life. The Commission still has delegated acts to adopt, harmonised standards are landing through 2027, and the Omnibus process itself proved that a deadline in the Official Journal is not the same as a deadline that holds. Check the European Parliament legislative train and the Cloud Security Alliance research feed before quoting any date here in a board paper.

Frequently asked questions

Related