Why EU AI Act compliance starts with an AIBOM
The AI Act's documentation duties assume you know which AI is in your software. Most teams don't yet.
TL;DR: The EU AI Act sets binding obligations for AI systems placed on the EU market. The Digital Omnibus on AI, in force since 27 July 2026, moved most high-risk requirements to 2 December 2027, while transparency duties under Article 50 have applied since 2 August 2026. In practice, compliance with Article 11, Article 50 and Article 53 depends on being able to inventory and document the AI/ML components embedded inside the software being shipped. Most enterprises cannot yet do this at code level.
The EU AI Act, formally Regulation (EU) 2024/1689, is the first comprehensive legal framework for artificial intelligence. It entered into force on 1 August 2024 and applies in phases, based on a risk classification of the AI systems concerned. Prohibited practices have been illegal since 2 February 2025. General-purpose AI model obligations applied from 2 August 2025.
The Regulation applies to providers and deployers of AI systems placed on or used in the EU market, regardless of where the organisation is established.
When does the EU AI Act actually apply?
On 7 May 2026, the Council presidency and the European Parliament reached a provisional agreement on the Digital Omnibus package. It was adopted as Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force from 27 July 2026. Annex III high-risk AI systems (use-case-based classifications covering areas such as employment, education, essential services and law enforcement) have their obligations deferred from 2 August 2026 to 2 December 2027, a delay of 16 months. High-risk AI embedded in regulated products under Annex I, including medical devices, machinery and toys, is deferred from 2 August 2027 to 2 August 2028. The grace period for AI-generated content transparency solutions was reduced from six months to three, with watermarking compliance due on 2 December 2026.
What did not change: the risk-based classification architecture, the four-tier system, the general-purpose AI obligations under Article 53, the conformity assessment regime, the AI Office's oversight role, and the penalty structure under Article 99. The Omnibus moved timelines. It did not move the underlying framework.
Penalties under Article 99 reach up to €35 million or 7% of global annual turnover for prohibited-practice violations, €15 million or 3% for high-risk system non-compliance, and €7.5 million or 1% for supplying incorrect information to authorities.
How does the EU AI Act affect software that contains AI components?
For high-risk AI systems, providers must establish a risk management system, govern training and testing data, maintain technical documentation, log automated operations, ensure human oversight, and meet accuracy, robustness and cybersecurity standards. The technical documentation under Article 11 includes a description of the AI system, its components, the data used in development, and the methods and tools applied. Article 50 introduces transparency obligations for AI systems that generate or manipulate content, including the labelling of synthetic output. Article 53 obligates providers of general-purpose AI models to publish a sufficiently detailed summary of the content used to train the model.
AI/ML capability is now embedded across mainstream software products, often without being declared at the package boundary. A typical enterprise application may include API calls to commercial large language model providers, open source ML libraries such as PyTorch or TensorFlow, model files distributed alongside containers, fine-tuned model weights stored as static assets, or orchestration code from frameworks such as LangChain. These components are AI systems or components of AI systems under the Regulation, but they are rarely visible to the procurement, compliance and legal functions that own the AI Act response.
The Regulation does not separate AI built in-house from AI integrated from third parties. If an AI system is placed on the EU market, the obligations follow the system. Article 11 technical documentation must describe the system's components. Article 50 transparency rules apply to systems that generate or manipulate content, regardless of whether the underlying model is in-house or hosted by a third party. Article 53 obligations for general-purpose AI providers may flow to deployers that materially modify a model and present it as their own.
In practice, this means an AI Act response begins with an accurate, code-level inventory of every AI/ML component in scope. An AI Bill of Materials is that structured inventory of AI and ML components present in a software system. A complete AIBOM should document: AI/ML libraries and SDKs by component, version, licence and origin; model files and weights referenced or embedded in the codebase; API credentials and call chains that reach external AI providers, with identification of which model and provider each credential addresses; and package metadata sufficient to establish provenance and assess downstream licence obligations. CycloneDX and the SPDX AI profile both support the format. What neither format can compensate for is incomplete detection at the source: an AIBOM generated from manifest files alone will miss vendored components, inlined SDK calls, and model files that are stored as static assets rather than declared as dependencies.
What should organisations do now?
Earnie builds that inventory from the code. Detection operates at snippet level rather than at the package manifest, which is necessary because AI components (SDK calls, model file references, machine learning library imports) are frequently inlined, vendored or copied in ways that manifest-only tools do not surface. Model files are matched against known models with their licence and lineage, and the result is exported as a CycloneDX AIBOM. In practice, this provides the inventory layer that Article 11 technical documentation, Article 50 transparency obligations and Article 53 general-purpose AI training-content summaries depend on but do not generate.
The next quarter's work is the same whichever deadline applies. Inventory the AI systems placed on the EU market and the AI development tools used to build them. Classify each system against the Annex III categories. Identify a single point of accountability in writing. Begin generating the Article 11 technical-file evidence that any future assessor will request, including the AIBOM coverage that documents AI-generated code.
Treating the Omnibus delay as permission to pause is not advisable. Evidence accumulated steadily over the coming months sits in a stronger position than evidence assembled under deadline pressure.
