Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Regulation

EU proposes delaying the AI Act's high-risk rules by up to 16 months

On November 19, the European Commission published a digital package that includes a “digital omnibus” to simplify the EU’s rules on artificial intelligence, cybersecurity and data. The headline change for AI: obligations for high-risk AI systems under the AI Act would apply only once supporting standards and tools are in place, giving companies “up to 16 months to comply”, according to the Commission.

CIO Dive reported that this would move the deadline for high-risk rules from August 2026 to December 2027. The proposals still need the approval of the European Parliament and the Council.

/ 01What the Commission proposed

The AI changes are targeted rather than a rewrite. The Commission’s summary lists simpler rules, more room for real-world testing, and “centralised oversight of AI systems.” CIO Dive described the latter as centralised oversight, through the European AI Office, of AI systems built on general-purpose models, along with simplifications such as lighter technical documentation for small and medium-sized companies.

The wider package also touches data rules: consolidating EU data legislation under the Data Act, targeted amendments to the GDPR, a single entry point for cybersecurity incident reporting, and changes to cookie rules. The Commission estimates the simplification could save “up to €5 billion in administrative costs by 2029.”

Reaction split along familiar lines. CIO Dive quoted the advocacy group EDRi calling the plans a “sweeping deregulation agenda for EU digital policies,” while Henna Virkkunen, the Commission’s executive vice-president for tech sovereignty, said European startups and small businesses are “often held back by layers of rigid rules.” Eduardo Ustaran of Hogan Lovells cautioned businesses not to mistake simplification for “abandoning of the principles, rights and rules that are at the core of the European digital regulation framework.”

/ 02Why it matters

The delay is real, but it is a delay to a deadline, not a removal of the requirements. Under the proposal, the high-risk obligations stay in the law. What changes is the date by which organisations must be able to prove them, and the link between that date and the arrival of harmonised standards.

That link is telling. The Commission is acknowledging that compliance is hard to demonstrate without agreed methods for doing so. For organisations building or deploying high-risk systems, the open question has never been only “when”; it has been “how do we show an auditor what our system did and why.”

/ 03The traceability angle

Much of what any high-risk AI regime asks for comes down to evidence. Logs that let someone reconstruct a decision. Documentation of data sources and processing that matches what actually runs. Enough visibility into the system’s reasoning that a person can step in. Systems whose outputs cannot be traced to their inputs struggle with all of this, no matter how much time they are given.

Architectures that keep knowledge in an explicit, versioned structure have an easier path. When the facts an AI system relies on live in a knowledge graph with provenance attached, and the reasoning over those facts is deterministic, the audit trail is a by-product of normal operation rather than a reconstruction after the fact.

There is a practical reason to start early regardless. Evidence trails are much cheaper to design in than to bolt on. A system that already records which data it used and which steps it took can produce documentation in whatever format a standard eventually requires. A system that never recorded those things has to be instrumented, re-tested and often partly rebuilt. The extra months are most valuable to organisations that use them for the second kind of work before they are forced to.

Smaller providers stand to gain the most from the proposed relief on documentation, but they also have the least capacity to rebuild later. For them, choosing architectures that make evidence cheap is a strategic decision, not a compliance chore.

A later deadline is more time to build the evidence trail, not a reason to stop building it.

/ 04What to do now

  • Do not pause compliance programmes. The proposal still has to pass Parliament and Council, and its final form may differ.
  • Use the time on the hard parts. Data lineage, decision logging and documentation that matches production are slow to retrofit.
  • Watch the standards. If the effective date is tied to harmonised standards, their content will define what “compliant” looks like in practice.
How it unfolded
NOV 19, 2025Commission publishes digitalomnibus[1]NOV 21, 2025This postAUG 2026Original high-risk rulesdeadline[2]UPCOMINGDEC 2027Proposed high-risk rulesdeadline[2]UPCOMINGNOV 19, 2025Commission publishes digital omnibus [1]NOV 21, 2025This postAUG 2026 · UPCOMINGOriginal high-risk rules deadline [2]DEC 2027 · UPCOMINGProposed high-risk rules deadline [2]
Aktiver’s take — Financial services

Sixteen more months buys time to build evidence, not a reason to postpone it. The firms that will struggle when the high-risk rules land are the ones whose systems cannot say, for one decision, which data and which logic produced it.

Take credit scoring, one of the Act’s high-risk categories. In a graph, each decision links to the applicant attributes it read, the policy version in force that day, the source tables those attributes came from, and the rule that fired. Because reasoning over the graph is deterministic, rerunning a contested decision returns the same outcome with the same evidence chain, which is what a supervisor or a complaint will ask for. When harmonised standards finally fix the documentation format, the content already exists as queryable lineage, and producing the file is an export rather than a forensic project. Ingestion Hub builds that model from the SQL tables, CSVs and policy PDFs a lender already has, with subject-matter experts refining it.

How we’d model it

Credit decisionApplicant attributeSource tablePolicy versionHigh-risk AI systemTechnicaldocumentationusedAttributesourcedFromgovernedBydocumentedByCreditdecisionApplicantattributeSource tablePolicyversionHigh-risk AIsystemTechnicaldocumentationusedAttributesourcedFromgovernedBydocumentedBy
  • Credit decision —usedAttribute→ Applicant attribute
  • Applicant attribute —sourcedFrom→ Source table
  • Credit decision —governedBy→ Policy version
  • High-risk AI system —documentedBy→ Technical documentation

See where this applies →

Sources — read the originals
  1. Simpler digital rules to help EU businesses grow — European Commission commission.europa.eu ↗
  2. EU plans to pull back on digital regulations — CIO Dive ciodive.com ↗
  3. Simpler EU digital rules and new digital wallets to save billions for businesses and boost innovation — European Commission digital-strategy.ec.europa.eu ↗
Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading