Deepwraite® · EU rule-making
Our contributions to the EU AI Act guidelines
We contribute to how the AI Act is interpreted – from the perspective usually missing in these consultations: small and medium-sized enterprises, skilled trades and municipalities. Concrete cases from advisory practice instead of theory.
Listed in the Stakeholders’ Catalogue of the Apply AI Alliance of the European Commission.
Article 6 – Classification of high-risk AI systems
What it’s about
The guidelines determine when an AI system counts as “high-risk” – and thereby triggers extensive obligations. For small businesses, a single question of interpretation often decides whether they are affected or not. In practice, classification is rarely a question of interpretation but of application.
What we contributed
- The Article 6(3) filter, translated into SME reality. The dividing line between a “narrow procedural task” and a substantive judgement is explained using school and visa examples – not transferable to invoice classification, email triage or quote review. We contributed in/out examples from the skilled trades and SMEs.
- Agentic and multi-agent systems. When does a chained agent pipeline lose the filter? An operational test is missing – we provide a fully worked municipal example.
- Proportionality of registration. Even systems correctly classified as not high-risk must be documented and registered (Art. 6(4)). Counter-intuitive and little known for micro-enterprises – we ask for an SME-proportionate path.
- Critical infrastructure / municipal utilities. When does a predictive-maintenance model tip into a “safety component”? What role does an independent, non-AI safety layer play? In/out examples from water and district-heating networks.
- Employment / SME HR. Recruiting filters (4a), the materiality threshold for shift planning and task allocation (4b), and a due-diligence checklist for using third-party applicant-tracking systems.
Article 50 – Transparency obligations for AI systems
What it’s about
Article 50 governs labelling and disclosure obligations: for AI interaction, AI-generated content, deepfakes and certain texts. In practice, it is often unclear to SMEs and municipalities which obligation applies to whom – and how it can be met technically.
What we contributed
- Role allocation: provider vs. deployer. Anyone deploying a commercial SaaS or LLM product under their own name remains the deployer – the upstream supplier is the provider. We address the gap when the provider cannot technically deliver the marking.
- Watermarking vs. provenance. The difference between embedded, model-side watermarks and after-the-fact provenance (C2PA) – and the GPAI downstream problem when marking is only “encouraged” rather than mandatory.
- Transparency in interaction (50(1)). Who discloses in multi-agent cascades? Retention for voice bots; consolidated disclosure across the DSA, UCPD and GDPR.
- Biometrics and emotion (50(3)). A negative list for standard municipal tools – e.g. plain speaker attribution when taking minutes, which is not categorisation.
- Deepfakes and texts (50(4)). Workable “editorial responsibility” for small municipalities, a feasible audience assessment, and role allocation in advertising-agency constellations.
- SME proportionality. The SME fine cap (Art. 99(6)) and the asymmetry between Code-of-Practice signatories and small businesses.
Both contributions are practitioner submissions from ongoing advisory mandates – not legal advice. Case examples anonymised.
Deepwraite® is the AI-governance practice of L&L Kommunikationskonzepte GmbH. We enable SMEs, skilled trades and municipalities to adopt AI quickly, measurably and compliantly – in line with the EU AI Act and ISO/IEC 42001. Artificial intelligence that works. Governance that holds.