™The Article 3(1) definition, and what it excludes
Article 3(1) defines an AI system as a machine-based system designed to operate with varying levels of autonomy, that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions. The operative verb is infers. A traditional rule-based system that maps inputs to outputs through hand-coded logic is not an AI system under the Act. Recital 12 of Regulation 2024/1689 is explicit on this point.
The high-risk gate: Annex III, point 5(b)
Where an AI system is used to evaluate the creditworthiness of natural persons or establish their credit score, it is high-risk. This does not extend automatically to instrument-level analysis of a bond, which concerns the issuer as a legal person, not a natural person. The scope of point 5(b) is narrower than commonly assumed. Where a governance system informs a credit decision on a natural person, the high-risk regime applies in full: risk management (Art. 9), data governance (Art. 10), technical documentation (Art. 11), record-keeping (Art. 12), transparency (Art. 13), human oversight (Art. 14) and accuracy, robustness and cybersecurity (Art. 15).
Why architecture matters more than classification
Whether or not a system is legally classified as high-risk, supervised financial entities are expected to explain the basis of every decision they take. A deterministic pipeline does this by construction: the ruleset is versioned, the inputs are hash-chained to the output, and the same file, run twice, produces the same bytes. A large-language-model pipeline does not, unless it is wrapped in structural constraints (temperature 0, seeded, fully cached, output schema-validated), and even then the model weights are opaque.
Model use in extraction, not in decision
The AI Act's high-risk regime does not prohibit machine-learning models. It requires that where models operate in the decision path, they meet Articles 9 to 15. The engineering answer, adopted by BondGovernance, is to confine model use to structured extraction (turning free-text indenture provisions into a typed schema) and to keep the decision engine deterministic. Extraction outputs pass through validation and are only used as evidence, never as authority.
Timeline and enforcement
The AI Act phases in over 24 months from 1 August 2024. Prohibitions on unacceptable-risk practices applied from 2 February 2025. Governance and general-purpose AI provisions apply from 2 August 2025. The full high-risk regime for Annex III systems applies from 2 August 2026. Firms with model-based components in the decision path should treat this as the compliance horizon.