sota.io
2026-04-12·12 min read·sota.io team

EU AI Act Article 7: How the Commission Expands Annex III Without a Parliamentary Vote (2026)

Your AI system passed the Article 6 classification test. It is not listed in Annex III. It does not function as a safety component in an Annex II product. You are not a high-risk provider — today.

Article 7 is the provision that can change that classification without a full legislative cycle. It grants the European Commission the authority to add new categories to Annex III — expanding the list of high-risk AI systems — through a delegated act procedure. The European Parliament and Council have a two-month objection window. No affirmative vote is required.

The practical implication: a system that is not high-risk today can become high-risk within months of a Commission decision, triggering the full Article 16 compliance stack — Arts 9–15, QMS, technical documentation, conformity assessment, EU database registration.

This guide explains the Art.7 mechanism, which categories are candidates for addition, what monitoring obligations arise, and how to design systems that can absorb a classification change without a complete rebuild.

What Article 7 Actually Says

Article 7(1) states that the Commission is empowered to adopt delegated acts in accordance with Article 97 to amend Annex III by adding new AI system categories to the high-risk list.

The criteria the Commission must apply when evaluating whether to add a category are set out in Art.7(2):

  1. Intended purpose in sensitive areas: Is the AI system intended to be used in areas involving decisions with significant consequences for individuals?
  2. Output use: Does the system produce outputs — recommendations, decisions, or content — that have a material effect on people?
  3. Harm potential: Is there a significant risk of harm to health, safety, or fundamental rights, considering both probability and severity of harm?
  4. Affected persons: Does the system affect a significant number of persons, particularly from vulnerable groups?
  5. Power asymmetry: Does use of the system create significant power imbalances between the deployer and the individuals subject to it?

The Commission is not free to add categories arbitrarily. Each addition must satisfy these five criteria and be supported by an impact assessment. But "impact assessment" is a Commission-internal process — it does not require external validation or parliamentary approval.

The Delegated Act Procedure: How It Works

Understanding the procedural mechanics helps you predict timelines:

Step 1: Commission Initiates Assessment

The AI Office or Commission services identify a candidate category. This can be triggered by market surveillance findings, incident reports under Art.73, or proactive monitoring under Art.97.

Step 2: AI Office Consultation

Under Art.7(2), the Commission must consult the AI Office before adopting delegated acts amending Annex III. The AI Office can prepare a technical opinion, but its opinion is advisory, not binding.

Step 3: Delegated Act Published

The Commission adopts the delegated act and publishes it in the Official Journal. The act specifies the new Annex III category with sufficient precision to allow classification determination.

Step 4: Two-Month Objection Window

The European Parliament and Council each have two months to object to the delegated act. If neither objects within the window, the act enters into force. The window can be extended once, but in practice this rarely occurs.

Step 5: Transitional Period

The delegated act typically includes a transitional period before the new category becomes binding. Existing systems may receive a grace period to come into compliance.

Total minimum timeline: Commission proposal to binding application can theoretically happen in under six months. In practice, the process from identification to application typically takes 12–24 months, but developers should not design for the maximum.

Candidate Categories: What Could Be Added to Annex III

The Commission has not published a formal shortlist, but public consultation documents, recital language, and AI Office working papers consistently identify several candidate categories:

1. Insurance Underwriting AI

AI systems that assess individual risk for life, health, or property insurance are conspicuously absent from current Annex III. Category 5(b) covers credit scoring for essential services, but insurance underwriting is treated separately in many national legal traditions.

Why this is a candidate:

Developer implication: Insurance underwriting AI built today should be designed with the Art.16 compliance stack in mind. Adding a QMS and technical documentation retroactively is significantly more expensive than building them in.

2. Autonomous HR Scoring Beyond CV Screening

Current Annex III Category 4 covers AI for CV screening, shortlisting candidates, and making or influencing employment decisions. A potential expansion: AI systems that perform ongoing performance scoring, attrition prediction, or career trajectory analysis for existing employees.

Why this is a candidate:

Developer implication: If your AI system produces any output that managers use to make decisions about existing employees — including dashboards, ranking systems, or performance predictions — monitor Art.7 developments closely.

3. Healthcare Triage AI

Annex III Category 5(a) covers AI used by essential services providers for public benefit assessments. But clinical triage AI — systems that determine the urgency and priority of patient care — sits in an ambiguous space.

Why this is a candidate:

Developer implication: Healthcare AI developers should track both MDR and EU AI Act Art.7 developments. A system that is a Class I medical device today may become a high-risk AI system under Annex III via a delegated act — and the compliance requirements of the two frameworks do not perfectly overlap.

4. Political Content and Micro-Targeting AI

AI systems used to generate or target political content — campaign messages, voter mobilization, issue framing — are referenced in recitals as potential future additions. Current Annex III does not include political AI, but the AI Office has flagged this as a priority monitoring area.

Why this is a candidate:

Developer implication: Election AI is an area of active EU regulatory development. Systems built for political parties, campaigns, or issue advocacy organizations should be designed with potential future high-risk classification in mind.

5. Credit Scoring Expansion

Annex III Category 5(b) already covers AI for creditworthiness assessment. A potential expansion: AI systems used to set dynamic pricing, loyalty program terms, or service access conditions based on predicted lifetime value or behavioral scoring.

Why this is a candidate:

What Art.7 Means for Your Classification Strategy

Obligation 1: Establish a Classification Review Process

Article 7 creates an implicit monitoring obligation for any AI system that operates near Annex III boundaries. Specifically:

Obligation 2: Design for Classification Resilience

A system designed with the Art.16 compliance stack as a possible future state is significantly cheaper to bring into compliance if classification changes. Specifically:

Obligation 3: Include Art.7 Risk in Vendor Contracts

If you deploy AI systems from third-party vendors, your contracts should address Art.7 reclassification risk:

These questions are not hypothetical. They should be in every AI vendor contract for systems that operate in candidate-category domains.

CLOUD Act Angle: Classification Documentation Storage

Art.7 creates classification documentation — analysis records establishing that a system is not high-risk. This documentation is produced by legal and engineering teams, often stored in shared drives, legal databases, or project management tools.

If your classification documentation is stored on U.S.-provider infrastructure (AWS, Azure, GCP, Microsoft 365, Salesforce), it is subject to the Clarifying Lawful Overseas Use of Data (CLOUD) Act. U.S. law enforcement can compel production of documents stored on U.S.-provider systems regardless of where the data physically resides.

Scenario: Your AI system is later reclassified as high-risk after a delegated act. A national market surveillance authority investigates your compliance. Your classification analysis — which may have concluded the system was not high-risk — is now evidence in a regulatory proceeding. If that analysis is stored on U.S. infrastructure, it could be simultaneously discoverable by U.S. authorities under CLOUD Act.

Mitigation: Store classification analysis, legal opinions, and risk assessments on EU-native infrastructure outside CLOUD Act reach. sota.io provides EU-native deployment with data residency guarantees for exactly this class of sensitive compliance documentation.

Python Tooling: AnnexIII Classification Monitor

from dataclasses import dataclass, field
from enum import Enum
from datetime import date, datetime
from typing import Optional

class CategoryStatus(Enum):
    NOT_HIGH_RISK = "not_high_risk"
    HIGH_RISK_CURRENT = "high_risk_current"
    BORDERLINE = "borderline"
    MONITORING_REQUIRED = "monitoring_required"

class CandidateCategory(Enum):
    INSURANCE_UNDERWRITING = "insurance_underwriting"
    HR_EMPLOYEE_SCORING = "hr_employee_scoring"
    HEALTHCARE_TRIAGE = "healthcare_triage"
    POLITICAL_CONTENT = "political_content"
    BEHAVIORAL_PRICING = "behavioral_pricing"

@dataclass
class ClassificationRecord:
    """Immutable record of a classification decision at a point in time."""
    system_id: str
    system_name: str
    classification_date: date
    analyst: str
    status: CategoryStatus
    annex_iii_categories_evaluated: list[str]
    rationale: str
    candidate_categories_assessed: list[CandidateCategory] = field(default_factory=list)
    next_review_date: Optional[date] = None
    delegated_act_ref: Optional[str] = None  # OJ reference if triggered by delegated act

    def is_review_overdue(self, as_of: date = None) -> bool:
        as_of = as_of or date.today()
        if self.next_review_date is None:
            return False
        return as_of > self.next_review_date

    def requires_art7_monitoring(self) -> bool:
        return (
            self.status in (CategoryStatus.BORDERLINE, CategoryStatus.MONITORING_REQUIRED)
            or len(self.candidate_categories_assessed) > 0
        )


class AnnexIIIMonitor:
    """
    Tracks classification status and monitors for Art.7 delegated act triggers.
    
    Use: instantiate per AI system. Call assess() when system launches,
    updates materially, or when a delegated act amends Annex III.
    """
    
    def __init__(self, system_id: str, system_name: str):
        self.system_id = system_id
        self.system_name = system_name
        self.classification_history: list[ClassificationRecord] = []
    
    def assess(
        self,
        analyst: str,
        status: CategoryStatus,
        categories_evaluated: list[str],
        rationale: str,
        candidate_categories: list[CandidateCategory] = None,
        review_months: int = 12,
    ) -> ClassificationRecord:
        """Record a classification assessment."""
        today = date.today()
        record = ClassificationRecord(
            system_id=self.system_id,
            system_name=self.system_name,
            classification_date=today,
            analyst=analyst,
            status=status,
            annex_iii_categories_evaluated=categories_evaluated,
            rationale=rationale,
            candidate_categories_assessed=candidate_categories or [],
            next_review_date=date(today.year, today.month + review_months % 12 + 1, today.day)
            if review_months < 12
            else date(today.year + review_months // 12, today.month, today.day),
        )
        self.classification_history.append(record)
        return record
    
    def trigger_delegated_act_review(
        self,
        delegated_act_ref: str,
        new_category_description: str,
        analyst: str,
    ) -> dict:
        """
        Called when a new Annex III delegated act is published.
        Returns assessment of whether the new category applies.
        """
        latest = self.current_record()
        if latest is None:
            return {"error": "No classification record found. Run initial assess() first."}
        
        return {
            "system_id": self.system_id,
            "system_name": self.system_name,
            "delegated_act": delegated_act_ref,
            "new_category": new_category_description,
            "analyst": analyst,
            "current_status": latest.status.value,
            "action_required": (
                "IMMEDIATE RECLASSIFICATION REVIEW"
                if latest.status in (CategoryStatus.BORDERLINE, CategoryStatus.MONITORING_REQUIRED)
                else "STANDARD REVIEW"
            ),
            "compliance_deadline_estimate": "6 months from delegated act entry into force (check specific transitional period)",
            "timestamp": datetime.utcnow().isoformat(),
        }
    
    def current_record(self) -> Optional[ClassificationRecord]:
        if not self.classification_history:
            return None
        return max(self.classification_history, key=lambda r: r.classification_date)
    
    def compliance_report(self) -> dict:
        current = self.current_record()
        if current is None:
            return {"status": "NO_CLASSIFICATION_RECORD", "action": "Run initial classification assessment"}
        
        return {
            "system_id": self.system_id,
            "system_name": self.system_name,
            "current_status": current.status.value,
            "classification_date": current.classification_date.isoformat(),
            "next_review_due": current.next_review_date.isoformat() if current.next_review_date else None,
            "review_overdue": current.is_review_overdue(),
            "art7_monitoring_required": current.requires_art7_monitoring(),
            "candidate_categories_at_risk": [c.value for c in current.candidate_categories_assessed],
            "classification_history_count": len(self.classification_history),
        }


# Example: Insurance AI system near Annex III boundary
monitor = AnnexIIIMonitor("INS-001", "Premium Calculation AI v2.1")

record = monitor.assess(
    analyst="compliance-team@insurer.eu",
    status=CategoryStatus.BORDERLINE,
    categories_evaluated=["Cat.5(b) creditworthiness assessment", "Cat.5(a) essential services"],
    rationale="System calculates premium adjustments based on behavioral data. Not a creditworthiness "
               "assessment under current Annex III definition. Cat.5(b) requires 'access to essential "
               "services' — insurance is arguably essential but not explicitly covered. BORDERLINE.",
    candidate_categories=[CandidateCategory.INSURANCE_UNDERWRITING],
    review_months=6,  # review every 6 months given borderline status
)

print(monitor.compliance_report())
# {'current_status': 'borderline', 'art7_monitoring_required': True, ...}

# When Commission publishes delegated act:
result = monitor.trigger_delegated_act_review(
    delegated_act_ref="OJ L 2026/XXX",
    new_category_description="AI systems used to assess individual risk for life, health, or property insurance",
    analyst="compliance-team@insurer.eu",
)
print(result)
# {'action_required': 'IMMEDIATE RECLASSIFICATION REVIEW', ...}

30-Item Art.7 Future-Proofing Checklist

Classification Documentation (5 items)

Monitoring Triggers (5 items)

Candidate Category Risk Assessment (5 items)

Architecture for Classification Resilience (5 items)

Contract and Procurement (5 items)

Response Planning (5 items)

5 Common Developer Mistakes on Art.7 Risk

1. "We're not in Annex III, so we don't need to track Art.7"

The entire point of Art.7 is that Annex III can change without a full legislative cycle. Systems that are not currently high-risk are exactly the systems that need Art.7 monitoring — providers of currently high-risk systems are already in compliance or working toward it. The Article 7 risk falls disproportionately on systems in the grey zone.

2. Treating Classification as a One-Time Event

Classification analysis performed at system launch decays in relevance as the regulatory landscape evolves. The Art.7 delegated act procedure creates a continuing obligation to reassess. A system that was definitively not high-risk in 2024 may be high-risk under a 2026 delegated act — and the provider will not receive individual notice.

3. Conflating Art.7 with Full Legislative Revision

Art.7 allows the Commission to add categories to Annex III. It does not allow the Commission to change the fundamental high-risk threshold, modify the compliance obligations in Arts 9–15, or override the Annex III exclusions in Art.6(3). Delegated acts expanding Annex III do not change what compliance means — they change who is subject to it.

4. Assuming Long Transitional Periods

While the AI Act's original Annex III categories had multi-year implementation periods, delegated acts adopted under Art.7 can specify shorter transitional periods for new categories. The Commission's impact assessment may conclude that candidate-category systems should have a shorter runway given market awareness of the risk. Don't design compliance timelines around the original 2026 deadline.

5. Neglecting Vendor and Customer Contracts

If your AI system is reclassified as high-risk, your obligations change — but so do your vendors' obligations toward you, and your customers' expectations. Contracts that don't address reclassification risk will require renegotiation at the worst possible time: when you are simultaneously trying to achieve compliance under a new regulatory framework with a defined deadline.

Connecting Art.7 to the Broader Compliance Architecture

Art.7 is not an isolated procedural provision — it sits at the center of several compliance interdependencies:

Summary: Art.7 in Practice

Article 7 is the EU AI Act's adaptability mechanism. It acknowledges that the Commission cannot anticipate every AI use case that poses significant risk to fundamental rights in 2024, and gives the regulatory framework the ability to evolve faster than the full legislative process permits.

For developers and providers of AI systems in candidate-category domains, the practical response is:

  1. Document your current classification analysis — in a version-controlled, EU-native repository
  2. Assess your Art.7 exposure — which candidate categories apply, how far from the threshold are you?
  3. Design for compliance resilience — build the technical foundations of Arts 9–15 into your architecture now
  4. Monitor the Official Journal — delegated acts are published there, not in press releases
  5. Update your contracts — with vendors, customers, and insurers to address reclassification risk

The systems most harmed by an Art.7 delegated act are those whose developers assume their current classification is permanent.

Covered in this series: Art.5 Prohibited Practices · Art.6 High-Risk Classification · Art.7 Annex III Delegated Acts · Art.16 Provider Obligations · Art.99 Penalties

See Also