What is the EU AI Act?
The EU AI Act is the world’s first comprehensive legal framework for artificial intelligence. Enacted in 2024 and now progressively entering into force, it classifies AI systems by risk level and assigns obligations accordingly – to developers, deployers, and the suppliers whose work feeds into AI systems.
At its core, the Act is designed to protect three things: people’s fundamental rights, including privacy, dignity, and protection from discrimination. The integrity of decisions that affect people’s lives, such as access to employment, credit, healthcare, or education. And public trust in AI as a technology. It is not designed to slow AI down. It is designed to ensure AI earns trust – and in doing so, create the conditions for sustainable, responsible innovation across Europe. It holds the full supply chain accountable for meeting them.
If you are building or deploying AI in a regulated context, that accountability does not begin at the model. It begins with the data.
Compliance is not something you add at the end
There is a common assumption in how organisations approach the EU AI Act: that compliance is something you address once a model is built. You complete a conformity assessment, assemble your technical documentation, register the system, and tick the boxes. The data that trained the model is, by then, a historical artefact.
That assumption is wrong – and in high-risk AI, it is an expensive one.
Article 10 of the Act places data governance at the centre of high-risk AI compliance. Training, validation, and testing datasets must be relevant, representative, and sufficiently free from errors. Potential biases must be identified and addressed. The entire process must be documented well enough to demonstrate all of this to a regulator. These are not vague ambitions. They are legal requirements – and the decisions that determine whether you can meet them are made in the earliest stages of development, not at the end.
Before deployment, high-risk AI systems must undergo a conformity assessment. You bear the burden of proof: you must demonstrate that your system is trustworthy and meets all applicable requirements. If an incident occurs or a regulator investigates, that same obligation applies. You do not know when you will need your documentation to hold up – at assessment stage, following a complaint, or during a supervisory review. The only way to be certain it does is to build it from the very first annotation decision.
Retrofitting compliance onto a model that was not built with it in mind is rarely straightforward and sometimes impossible. You cannot reconstruct an audit trail that was never created. You cannot remove bias from a model whose training data encoded it from the start. You cannot make a system traceable and explainable if the annotation process that underpins it was never documented. This is the compliance risk that is hardest to recover from: not a gap you discover during assessment, but one that was created on day one.
What high-risk actually means in practice
High-risk AI under the Act covers a significant portion of commercially relevant AI: hiring tools, credit and insurance scoring, medical devices, access to education, law enforcement applications. In each of these contexts, the Act imposes the full compliance burden: conformity assessment, human oversight, post-market monitoring, fundamental rights impact assessments.
The sectors where AI accuracy matters most are precisely the sectors where the stakes of getting the data foundation right are highest. An edge case missed in labelling is not a minor data quality issue. It is a failure mode waiting to surface in production, in a context where errors have real consequences.
If you operate in one of these sectors, you need annotation partners who understand this. Not suppliers who provide labelled data and walk away, but partners whose working practices are designed from the ground up for the compliance environment you operate in.
The compliance gap no one talks about
Most organisations commissioning AI models focus their compliance attention on the model itself: its outputs, its documentation, its human oversight mechanisms. Fewer scrutinise the upstream data process with the same rigour – and that is where the risk accumulates.
Bias does not typically appear in algorithms. It appears in datasets – in what gets labelled, how consistently, by whom, and according to which guidelines. A model trained on inconsistently labelled data will produce inconsistent outputs. A model trained on biased data will systematically disadvantage certain groups. Neither can be fully corrected after the fact.
The same is true of explainability. The Act requires AI decisions to be traceable and understandable – to regulators and to the people affected by them. That is often framed as a technical challenge. In practice, it is a documentation challenge – and annotation is where that documentation either exists or does not. A well-annotated dataset is not just training data. It is evidence. When a model’s decision is questioned, the annotation record is part of what makes it defensible.
If your audit trail is incomplete, if labelling decisions were not recorded, if inter-annotator agreement was never measured – you cannot reconstruct it. You are exposed on bias and on explainability at the same time. This is why, in regulated sectors and for high-risk models, high-quality annotation is a compliance decision – not merely a technical one.
Compliance as a supply chain question
The EU AI Act creates a clear chain of liability: from AI developer to deployer to data supplier. Annotation services sit in that chain. Where annotated data contributes to a high-risk AI system, its quality becomes legally material to your compliance position.
This reframes the procurement question. Choosing an annotation partner is not purely a question of cost per label or turnaround time. It is a question of whether that partner’s quality framework, documentation practices, and audit trails will hold up when your AI system faces regulatory scrutiny.
The organisations that are building this thinking into their procurement now – rather than discovering the gap at conformity assessment stage – are the ones that will have the cleaner compliance stories to tell, and the better-performing models to show for it.
auticon provides annotation services for high-risk AI applications across aviation, agriculture, healthcare, and other sectors where accuracy is non-negotiable. We build our quality framework to support Article 10 compliance from the first labelling decision to the final audit trail.
