Disclaimer: This article is for informational purposes only and does not constitute legal advice. Organizations deploying AI systems in the EU, or targeting EU residents, should consult qualified legal counsel to assess their specific obligations under Regulation (EU) 2024/1689 and applicable national implementing measures.
Classification is the threshold question of EU AI Act compliance. According to Eurostat’s enterprise AI adoption series (ISOC_EB_AI), AI tool use is now documented across businesses of all sizes in EU member states — which means that for most organizations, the question is no longer whether to conduct a classification analysis, but how to do it rigorously. Regulation (EU) 2024/1689 enters its most consequential enforcement period on 2 August 2026, when obligations for high-risk AI systems listed in Annex III become fully applicable. For GRC teams and legal-adjacent engineering leaders, understanding where AI coding tools may tentatively fall within the four-tier framework is the prerequisite for every downstream compliance decision.
The EU AI Act’s Four-Tier Risk Framework
The EU AI Act establishes what the European Commission describes as “a clear, easy-to-understand approach, based on four different levels of risk.” The architecture is proportional: regulatory obligations intensify as potential harm increases. The four tiers are:
Prohibited practices — AI applications banned outright under Article 5
High-risk AI systems — the most regulated tier, governed primarily by Articles 6 through 27
Limited-risk AI systems — systems carrying specific transparency obligations
Minimal-risk AI systems — systems facing few or no regulatory requirements
The regulation defines “AI system” in Article 3 as “a machine-based system that is designed to operate with varying levels of autonomy” that may generate “outputs such as predictions, recommendations, decisions, or other content” influencing physical or virtual environments. Confirming that a tool meets this definition is the necessary first step before any risk-tier analysis proceeds.
The EU AI Office published Guidelines on the Definition of an AI System in February 2025 to support this threshold determination — clarifying which software products fall within scope, and providing a practical foundation for classification work even where edge cases remain unresolved.
Tier 1 — Prohibited Practices (Article 5)
Article 5 of the regulation establishes eight categories of AI applications that are prohibited outright. As generally understood, these represent use cases where potential harm to fundamental rights or human dignity is considered irremediable through technical or organizational controls, making any deployment incompatible with the regulation’s objectives.
— Systems using subliminal or purposefully manipulative techniques to distort behavior and impair informed decision-making
— Systems exploiting vulnerabilities of specific groups — by age, disability, or economic situation — in ways that materially alter behavior
— Social scoring systems classifying individuals based on behavior or characteristics in ways leading to disproportionate treatment
— Predictive criminal risk assessment based solely on profiling of natural persons
— Untargeted facial recognition database scraping from the internet or CCTV footage
— Emotion inference in workplace or educational settings (with narrow exceptions for medical or safety contexts)
— Biometric categorization to infer sensitive attributes including race, political opinions, union membership, religious beliefs, or sexual orientation
— Real-time remote biometric identification in public spaces by law enforcement (with narrow exceptions for serious crime)
For most software development organizations, standard AI coding tools fall clearly outside these categories. The EU AI Office published Guidelines on Prohibited AI Practices in February 2025 to support interpretation. Organizations building internal monitoring systems, workforce analytics platforms, or HR automation tools should seek independent legal assessment before deployment.
Tier 2 — High-Risk AI Systems (Article 6 and Annex III)
High-risk classification is where the EU AI Act’s compliance infrastructure concentrates. Article 6 establishes two pathways to high-risk status.
Pathway 1: An AI system qualifies as high-risk when it functions as a safety component of a product covered by EU harmonization legislation listed in Annex I — such as medical devices, machinery, or aviation systems — and that product requires third-party conformity assessment before market placement.
Pathway 2: AI systems listed in Annex III are automatically considered high-risk, regardless of whether they are embedded in a regulated physical product.
Annex III identifies eight high-risk sectors:
1. Biometrics — remote identification systems, emotion recognition, biometric categorization
2. Critical infrastructure — AI managing digital networks, utilities, or road traffic
3. Education and vocational training — access decisions, learning outcome evaluation, exam monitoring
4. Employment and worker management — recruitment, performance monitoring, task allocation, termination decisions
5. Essential services — credit scoring, health and life insurance risk assessment, public benefit eligibility
6. Law enforcement — risk assessment, evidence evaluation, recidivism prediction
7. Migration, asylum, and border control
8. Administration of justice and democratic processes
For software teams, Areas 4 and 3 merit the closest attention. Any AI system used to evaluate developer performance, automate task allocation, or inform promotion decisions falls within Area 4. Any AI tool integrated into a recruitment pipeline presents similar classification exposure. High-risk systems carry obligations including risk management (Article 9), data governance (Article 10), technical documentation (Article 11), automatic logging (Article 12), transparency (Article 13), human oversight (Article 14), and accuracy standards (Article 15) — examined in detail in the next article in this series.
The Article 6(3) Derogation — Not All Annex III Systems Are High-Risk
Article 6 provides a derogation. An AI system that falls within an Annex III category may nonetheless be excluded from high-risk classification if it “does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons.” As generally understood, the derogation applies where the system:
— Performs only a narrow procedural task
— Improves the result of a previously completed human activity
— Detects decision-making patterns without influencing or replacing human judgment
— Performs only preparatory tasks for a human assessment
One firm constraint applies: AI systems performing profiling of natural persons remain high-risk regardless of the derogation. Providers asserting non-high-risk status for an Annex III system must document their assessment before placing the system on the market (Article 49). The derogation is not self-executing — it must be affirmatively justified and recorded before market entry.
As the European Parliament Research Service has noted in its analysis of the regulation, the risk-based framework was intentionally designed to allow proportionate treatment — but only where a lower classification can be substantiated with documented evidence.
Tier 3 — Limited-Risk Systems and Transparency Requirements
Limited-risk AI systems face narrower obligations concentrated on transparency. As generally understood, the regulation requires disclosure when users interact with a chatbot or automated system (unless obvious from context), marking of AI-generated synthetic content as artificially generated where it concerns matters of public interest, and notification to subjects when emotion recognition or biometric categorization systems are in operation.
Many AI coding assistance tools that operate through conversational interfaces may trigger these transparency obligations even where high-risk classification does not apply. Organizations deploying conversational AI in internal developer workflows should consider whether user disclosure requirements are relevant to their specific deployment.
Tier 4 — Minimal-Risk AI Systems
The regulation imposes no mandatory obligations on AI systems classified at minimal risk. Most AI tools currently in commercial use — including recommendation algorithms, spam filters, and general productivity tools — are likely to fall into this tier. Organizations may voluntarily adopt codes of conduct encouraged under the regulation, but no binding compliance obligations attach at this level.
A Worked Example — Classifying a Typical Dev Team’s AI Tool Stack
To illustrate how classification analysis proceeds in practice, consider a software engineering team at a financial services organization using a range of AI tools. The following tentative classifications are offered for analytical purposes only. Actual classification depends on specific use case, deployment context, and independent legal assessment. Classification may change as additional Commission guidance is issued under the regulation.
AI-powered code completion (IDE plugin)
Tentative tier: Minimal-risk or limited-risk. A tool suggesting completions to a developer who reviews and accepts or rejects every suggestion performs a support function for human work. Absent any context bringing it within Annex III, this tool is unlikely to be classified as high-risk under the current framework. Limited-risk transparency obligations may apply if the interface could be confused for a human system.
AI-assisted pull request analysis (automated code review)
Tentative tier: Limited-risk; Article 6(3) derogation potentially applicable. A tool that surfaces risk signals for a human reviewer performs what Article 6(3) characterizes as a “preparatory task for human assessment.” Where human review authority is retained and the tool does not output binding determinations, the derogation pathway may apply. Classification would shift if the tool is used to gate developer access to production systems or inform performance-related decisions.
AI-driven developer performance scoring
Tentative tier: High-risk — Annex III, Area 4 (employment and worker management). Any AI system influencing performance assessments, task allocation, or promotion and termination decisions for employees falls within the Annex III Area 4 definition. This classification applies whether the system is vendor-supplied or internally developed. Full high-risk obligations would apply from 2 August 2026.
AI coding agent used for autonomous task execution
Tentative tier: GPAI provisions apply; high-risk possible depending on deployment. General-purpose AI models are governed under Title VIII of the regulation. Under Article 51, models with training compute exceeding 10²⁵ FLOPs are presumed to carry systemic risk and face additional obligations. Beyond model-level designation, the deployer of a GPAI model as an autonomous agent in a software pipeline must separately assess whether the specific deployment brings the system within an Annex III category.
AI used in security vulnerability triage for critical systems
Tentative tier: Case-by-case assessment required. A system analyzing vulnerabilities in software controlling utilities, financial infrastructure, or public communications networks may fall within Annex III Area 2 (critical infrastructure). Classification depends on what the underlying software governs, not merely on what the AI tool does in isolation.
Classification Is Contextual, Not Categorical
The EU AI Act does not operate as a product registry where a tool name yields a definitive classification. As Article 6 and the derogation provisions make clear, classification is a contextual judgment that depends on what the system does, how it is deployed, who is affected by its outputs, and whether meaningful human oversight is embedded in the workflow.
The NIST AI Risk Management Framework (AI RMF 1.0, January 2023) provides a complementary analytical lens for organizations approaching this judgment. Through its Govern, Map, Measure, and Manage functions, the framework encourages organizations to characterize AI risk systematically before committing to a classification — supporting the kind of documented risk analysis that Article 6(3) derogation claims require. While the NIST AI RMF is a US voluntary framework and does not itself constitute EU compliance evidence, its risk characterization methodology aligns closely with the documentation-intensive analysis the regulation demands.
Building the Governance Layer for Any Classification Outcome
Regardless of where a specific AI tool tentatively falls in the framework, the practical challenge for GRC teams is the same: maintaining structured, retrievable evidence of how AI systems actually operate in production.
re-entry.ai operates as an MCP Gateway layer across AI coding workflows — capturing structured records of every model call, code change, and automated decision passing through development pipelines. For organizations working through classification analysis, this creates a verifiable foundation: rather than reconstructing AI system behavior from fragmented logs or developer recollection, governance teams and legal counsel can work from a consistent audit record that reflects actual system operation.
This does not substitute for legal assessment or formal classification determination. Organizations remain responsible for determining whether their specific AI deployments fall within the EU AI Act’s scope. But the governance infrastructure that supports that determination — and that addresses the Article 12 logging requirements for systems that are ultimately classified as high-risk — can be operational before the 2 August 2026 deadline.
Next in this series: EU AI Act Compliance Requirements: A Practical Checklist for Dev Teams — a structured walkthrough of the Article 9–15 obligation set for organizations whose AI tools are classified as high-risk. Part of the EU AI Act mini series for GRC teams and engineering leaders at re-entry.ai.