# The AI Act keeps moving. What happens to the decisions already made?

Canonical: https://www.meshqu.com/blog/ai-act-still-moving-decisions-stand-still

The EU AI Act has been delayed again and its timelines are still being argued over. The decisions already made under it cannot move with them.

Published: 2026-04-29T00:00:00.000Z

<Callout variant="emphasis" title="TL;DR" surface>

- The EU AI Act has been delayed again: legislators failed to agree on whether parts of the regulation should be pushed back.
- The framework is not a fixed rulebook. Timelines, standards, and guidance continue to evolve in motion.
- Most governance programmes assume stable policy. When the rules shift, reconstructing a past decision becomes interpretation, not evidence.
- A decision captured at execution, with its input, policy, context, outcome, version and timestamp, does not need to be reinterpreted when the rules change.

</Callout>

The EU's AI Act has been delayed again. Legislators in Brussels failed to agree on whether parts of the regulation should be pushed back, and timelines, obligations, and scope are still being debated.

That is not the Act going away. It is the Act still taking shape. The rules a decision is made under this year may not be the rules it is later judged against.

## The AI Act did not arrive as a single switch-on moment

It took shape gradually: proposed, reworked, politically agreed, brought into force, and now still being implemented in phases.

<Timeline
  eyebrow="EU AI Act timeline"
  items={[
    {
      marker: '2021',
      title: 'Proposal tabled',
      body: 'The European Commission tables the first draft of a horizontal regulation for AI.',
    },
    {
      marker: '2022',
      title: 'Amendments negotiated',
      body: 'Council and Parliament negotiate amendments through the year.',
    },
    {
      marker: '2023',
      title: 'Political agreement',
      body: 'Trilogue lands a provisional deal between the Council, Parliament, and Commission.',
    },
    {
      marker: '2024',
      title: 'Law enters force',
      body: 'The text becomes law, with obligations switched on in stages over the years that follow.',
    },
    {
      marker: '2026',
      title: 'Delay debated',
      body: 'Legislators in Brussels fail to agree on whether parts of the regulation should be pushed back. Timelines and obligations are still being contested.',
    },
    {
      marker: 'Ongoing',
      title: 'Phased implementation',
      body: 'Requirements land over time. Standards are still being defined. Guidance is still being written.',
    },
  ]}
  footer="A system designed to take shape gradually, not a fixed rulebook handed down on day one."
/>

Recent discussions in Brussels show the same pattern. The disagreement is not about whether the AI Act applies. It is about **how** and **when** it can realistically be enforced. The latest round produced requests to extend compliance timelines and proposals to delay high-risk obligations, not because the regulation has gone away but because the machinery around it is not yet ready to operate. Authorities are still being designated. Conformity assessment bodies are not yet in place. Harmonised standards are still emerging.

As one legal analysis put it, the current changes "introduce a degree of uncertainty, whilst at the same time giving the prospect of additional time."

The framework exists. The machinery does not.

## What is compliant today may not be compliant next year

Most governance programmes are built on a simple idea: the rules stabilise. You define policy, implement controls, align systems, and then operate.

The AI Act is not a fixed rulebook. Interpretations evolve, guidance is layered in over time, and standards determine how compliance is measured. What passes review next year may not match the conditions that existed when the original decision was made.

A system produces an outcome. A loan is declined. A transaction is flagged. A customer is scored.

At that moment, a specific policy exists. A specific model version. A specific threshold. The decision is the consequence of all three meeting at one specific time.

_Illustrative example: a schematic decision, not a real one._

<MomentCard
  eyebrow="At execution"
  title="What is in force, right now"
  fields={[
    { label: 'Policy', value: 'A specific version exists', tone: 'accent' },
    { label: 'Model', value: 'A specific version exists' },
    { label: 'Threshold', value: 'A specific value exists' },
  ]}
/>

## The reasoning is reconstructed, not observed

Six months later, the regulation has shifted, the guidance has been reinterpreted, the policy has been updated and the model has been retrained. The context in which the decision was made no longer exists in the same form. Then a regulator asks why the decision happened, not under today's rules, but under the rules that existed then.

The organisation looks back. It finds logs, events, and fragments spread across systems. The inputs can be recovered and the outcomes located. The policy version in force at execution, the model version that ran, the threshold it applied and any guarantee that the record is unchanged are not among them, so the reasoning is assembled from what remains rather than observed as it was.

This is not a failure of one system. It is structural. Across AI governance, the data already exists: risks are catalogued, incidents are recorded, frameworks are published. But they exist side by side, not as a single coherent artefact. Even recent work to map the AI risk landscape underlines this. Datasets can be brought together into a shared interface, but the connections between them still have to be drawn by the reader.

The information is there. The decision is not.

Where the policy version, the model version, the threshold and the integrity guarantee were never recorded, reconstruction cannot supply them. What it produces is an account read through the rules and the systems that exist now: **interpretation rather than evidence.**

## The decision is captured as it is made

> A system built for stable policy breaks here. A system built for *moving* policy assumes this from the start.

Policies will change, standards will evolve, interpretations will shift. So the decision is captured at the moment it is made, rather than referenced later. The input, the policy, the context, the outcome, the version, and the timestamp are preserved together as they were.

When the rules change, that record does not have to be reinterpreted. It can be checked against the policy version recorded with it, rather than against the policy in force today.

That is a narrower claim than being right. The record does not establish that the policy was well drafted or that the outcome was correct. It establishes what was decided, on what inputs, under which version: the part that reconstruction cannot supply.

_Illustrative example: a generated receipt, not a measured retrieval._

<ResolvedReceipt
  question="Which rules were in force when this decision was made on 14 March?"
  receiptId="RCP-AIACT-00481"
  resolvedIn="Illustrative retrieval"
/>

MeshQu does not replace the systems that make these decisions. It captures what those systems do not retain: the decision as it stood at execution.

The AI Act will not arrive as a single, stable regime. It will continue to evolve through standards, guidance, and enforcement. The rules will keep moving; a decision already made will not. If it was not recorded as it was made, no delay to the timetable recovers it.

<FinalCtaBanner size="compact" ctaLabel="Inspect a receipt" ctaHref="/receipts">
  A record made at execution can still be read under the rules that applied then.
</FinalCtaBanner>

## Sources & context

<Sources
  items={[
    {
      source: 'POLITICO',
      title: 'EU legislators fail to clinch deal to delay AI law',
      href: 'https://www.politico.eu/article/eu-legislators-fail-to-clinch-deal-to-delay-ai-law/',
    },
    {
      source: 'EUROPEAN COMMISSION',
      title: 'AI Act regulatory framework and timeline',
      href: 'https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai',
    },
    {
      source: 'BRUEGEL',
      title: 'AI Act implementation pressure',
      href: 'https://www.bruegel.org/policy-brief/efficiency-and-distribution-european-unions-digital-deregulation-push',
    },
    {
      source: 'CEPS',
      title: 'How the AI Act evolved through negotiation',
      href: 'https://www.ceps.eu/eu-policymakers-youve-done-a-mostly-good-job-on-the-ai-act-now-finish-it-already/',
    },
    {
      source: 'DLA PIPER',
      title: 'Phased implementation and compliance preparation',
      href: 'https://knowledge.dlapiper.com/dlapiperknowledge/globalemploymentlatestdevelopments/2024/eu-ai-act-to-enter-into-force-implications-for-employers',
    },
    {
      source: 'A&O SHEARMAN',
      title: 'Why obligations are still shifting',
      href: 'https://www.aoshearman.com/en/insights/ao-shearman-on-tech/generative-ai-and-the-eu-ai-act-a-closer-look',
    },
    {
      source: 'CLIFFORD CHANCE',
      title: 'Timeline uncertainty and the Digital Omnibus',
      href: 'https://www.cliffordchance.com/insights/resources/blogs/talking-tech/en/articles/2023/04/the-eu-ai-act--concerns-and-criticism.html',
    },
    {
      source: 'EURONEWS',
      title: 'Implementation delays and enforcement gaps',
      href: 'https://www.euronews.com/my-europe/2025/11/19/european-commission-delays-full-implementation-of-ai-act-to-2027',
    },
    {
      source: 'OECD',
      title: 'Regulatory uncertainty in AI governance',
      href: 'https://www.oecd.org/en/publications/progress-in-implementing-the-european-union-coordinated-plan-on-artificial-intelligence-volume-2_3ac96d41-en.html',
    },
    {
      source: 'MIT AI RISK INITIATIVE',
      title: 'Introducing the AI Risk Navigator',
      href: 'https://airisk.mit.edu/',
    },
  ]}
/>