For most of its life, the EU AI Act has been something enterprises read about rather than something they had to answer to. That changed through the first half of 2026. National market surveillance authorities across the EU have been standing up their supervisory functions, guidance on conformity assessment has moved from draft to published, and the compliance deadlines for high-risk AI systems are no longer a future milestone on a roadmap slide. They are now. This is a useful moment to be precise about what "high-risk enforcement" actually means in practice, who it reaches, and the gap that still exists between where most enterprise AI governance programmes are and where the Act requires them to be.
From principle to process
The Act's early phase was dominated by prohibited-practice provisions and general-purpose model transparency obligations, which most enterprises could address with policy statements and vendor questionnaires. The high-risk regime is a different order of obligation entirely. It applies to AI systems used in specifically listed contexts, including employment and worker management, access to essential services such as credit scoring and insurance pricing, critical infrastructure, education and vocational assessment, and certain law-enforcement-adjacent and migration functions. If a system you provide or deploy falls into one of these categories, the Act no longer asks you to describe your intentions. It asks you to demonstrate, with evidence, that a defined set of technical and organisational controls are in place and operating.
National authorities have used the first part of 2026 to publish practical guidance clarifying how conformity assessment procedures apply to common enterprise patterns: AI embedded in HR platforms, AI used within credit decisioning pipelines, and AI components sourced from third-party vendors and integrated into a regulated product. The direction of that guidance is consistent. Responsibility does not stop at the vendor. Deployers who put a high-risk system into a real operational context inherit obligations of their own, even when they did not build the model.
What "high-risk" enforcement actually requires
Strip away the legal drafting and the obligations for a high-risk system reduce to a short list of concrete deliverables, each of which needs to exist as evidence, not intention. Providers need a conformity assessment appropriate to the system, whether through internal control or, for certain categories, a notified body, resulting in a declaration of conformity and CE marking. They need technical documentation detailed enough that a regulator unfamiliar with the system could understand its design, training data, performance characteristics and known limitations. They need risk management that runs across the system's lifecycle rather than a one-off assessment at launch, data governance controls over training and testing data, human oversight mechanisms that give a real person the practical ability to intervene, logging that is automatic and tamper-evident, and post-market monitoring that feeds real-world performance back into the risk file.
Deployers face a lighter but still substantial set of duties: using the system in line with its instructions, assigning human oversight to people with the competence and authority to act on it, monitoring for foreseeable risks, and, for public bodies and certain financial services and insurance use cases, conducting a fundamental rights impact assessment before deployment. Enforcement in 2026 has made clear that authorities intend to test these obligations against the underlying documentation, not a policy statement that they are met.
Extraterritorial reach: who is caught outside the EU
One of the most consequential enforcement themes this year has been the confirmation, in guidance and in practice, that the Act's reach is not confined to companies headquartered in the EU. Any provider that places a high-risk AI system on the EU market, or whose system's output is used within the EU, falls in scope regardless of where the provider is established. The same applies to deployers located outside the EU if the output of their high-risk system is used within the Union. For a UK, US or wider international enterprise, this means selling an AI-enabled HR platform, credit assessment tool or safety-critical component into the EU market brings the full high-risk regime with it, including the requirement to appoint an authorised representative established in the EU where the provider itself is not.
This is the point most non-EU enterprises still underestimate. Legal teams often assume that AI regulation is a jurisdictional question resolved by where the company is registered. Regulators have made clear through this year's guidance that it is resolved by where the system's effects land. Any organisation with EU customers, EU employees affected by an AI-driven HR tool, or EU users of a product with an embedded high-risk AI component should now assume the Act applies to them directly.
The gap between governance programmes and the Act's requirements
Most enterprises that have invested in AI governance over the past two years built programmes oriented around responsible AI principles: fairness reviews, model cards, ethics committees. Those are valuable, but they are not the same thing as what the Act's high-risk regime demands, and the gap is now showing up in practice. Few organisations have technical documentation in the specific form conformity assessment requires. Fewer still have logging built to the standard of automatic, tamper-evident record-keeping across the full lifecycle of the system, rather than application logs assembled after the fact. Human oversight is frequently described in policy but not engineered into the product, meaning the person nominally responsible for oversight has no practical mechanism to intervene before a decision takes effect. And post-market monitoring, the requirement to keep watching a system after deployment and feed findings back into its risk file, is usually the newest and thinnest part of any existing programme.
Closing this gap means treating the Act's requirements as an operating model with named owners, defined evidence and a cadence, the way a mature organisation treats financial controls or information security, rather than writing more policy.
A compliance checklist to stand up now
Enterprises that are behind on this do not need to solve every requirement simultaneously. The following gives a realistic starting sequence for providers and deployers of high-risk systems.
- Inventory every AI system in use or under development and classify each against the Act's high-risk categories, documenting the reasoning behind each classification.
- For systems classified high-risk, commission or complete the appropriate conformity assessment and prepare the declaration of conformity before the system goes into, or remains in, production.
- Assemble technical documentation to the standard a regulator would expect, covering design, training data provenance, performance evaluation and known limitations, and keep it current as the system changes.
- Engineer human oversight into the product itself, giving the designated reviewer a real, timely mechanism to intervene, not just a policy naming them as responsible.
- Implement automatic, tamper-evident logging across the system's operational lifecycle, and confirm it can be produced on request.
- Stand up post-market monitoring that routes real-world performance and incident data back into the system's risk file on a defined schedule.
- If your organisation sells into or is used within the EU from outside it, confirm whether an EU-based authorised representative is required and appoint one before, not after, a regulator asks.
Building the operating model, not just the paperwork
The organisations coping best have not treated this as a legal exercise bolted onto an existing AI programme. They have assigned a named accountable owner for AI Act compliance, built the evidence requirements into model development and procurement so documentation is produced as systems are built rather than reconstructed under pressure, and set a recurring review cadence that treats classification, documentation and monitoring as living obligations. That discipline is what regulators test for when they ask for evidence, and it is also what keeps the compliance burden manageable as the AI estate grows.
The enforcement phase of the EU AI Act rewards organisations that treat compliance as engineering rather than paperwork, and it is unforgiving of those that do not, wherever in the world they are based. If you need help assessing your exposure, closing the documentation and oversight gap, or standing up the operating model this regime requires, email sales@halfteck.com.