2026-04-12·13 min read·sota.io team

EU AI Act Article 84: Annual Market Surveillance Reporting — Developer Guide (2026)

Article 84 of the EU AI Act creates the accountability layer that sits on top of the entire enforcement machinery. While Articles 79 through 83 govern how market surveillance authorities investigate individual AI systems and impose corrective measures, Article 84 forces those authorities to account for their work at Union level — to the Commission, to the European Parliament, to the Council, and to the public.

For developers and providers of high-risk AI systems, Article 84 matters for a reason that is not immediately obvious: the data your AI system generates during its operational life — incident reports, conformity records, post-market monitoring logs — flows directly into the Art.84 reporting pipeline. If your system triggers an MSA investigation, the outcome becomes part of the annual data set. If your documentation was incomplete when audited, that gap is counted. If a corrective measure was imposed, the type and outcome are aggregated and published.

Understanding what MSAs collect and how that data is processed is essential for two reasons. First, it tells you exactly what documentation you need to maintain proactively. Second, if that documentation is stored on US-operated cloud infrastructure, it is subject to CLOUD Act compellability — meaning US law enforcement could access your EU compliance records via your cloud provider before they reach the MSA.

What Article 84 Actually Requires

Article 84 creates two interlocking obligations.

Art.84(1) — MSA Annual Reports to the Commission. Each member state's market surveillance authority must report annually to the Commission on the results of its enforcement activities. The reporting obligation covers:

Results of market surveillance activities (inspections conducted, AI systems examined)
Results of post-market monitoring activities (serious incidents reported, safety events tracked)
Corrective measures ordered and their outcomes (withdrawals, modifications, prohibitions)
Cooperation activities with other MSAs and the EU AI Office
Cases where cross-border notification under Art.82 was triggered
Penalties imposed under Art.99 and their amounts

Art.84(2) — Commission Summary Publication. The Commission compiles from those national reports a Union-wide summary and makes it available to the European Parliament, the Council, and the public. The summary includes:

A general analysis of market surveillance outcomes across member states
Findings based on post-market monitoring activities of high-risk AI providers
Identified systemic issues and patterns across product categories or sectors
Aggregate enforcement statistics without identifying individual providers

The public summary creates a market intelligence signal that every developer should monitor. Sectors where systematic compliance failures appear in Art.84 summaries become targets for intensified surveillance in subsequent cycles.

What MSAs Actually Collect: Eight Data Categories

When an MSA conducts market surveillance activities that feed into its Art.84 annual report, it draws from eight distinct data categories. Each maps to documentation you should maintain.

Category 1 — EUAIDB Registration Completeness. MSAs cross-reference the EU AI database (established under Art.71) against AI systems they find in the market. Systems found without EUAIDB registration, or with incomplete registration data, are counted as formal non-compliance under Art.83. Gaps feed directly into Art.84 statistics.

Category 2 — Technical Documentation Audit Findings. When MSAs inspect high-risk AI systems, they check Annex IV technical documentation (Art.11). Missing, incomplete, or inadequate documentation is a finding. The Art.84 report captures how many systems had complete documentation versus how many required remediation.

Category 3 — Conformity Assessment Records. MSAs verify whether conformity assessments were conducted as required (Art.43-44) and whether notified body involvement was mandated. Providers who completed self-assessment when notified body review was required generate findings.

Category 4 — Post-Market Monitoring Data. Under Art.72, providers of high-risk AI systems must maintain post-market monitoring plans. MSAs verify these plans exist and that serious incidents were reported to authorities under Art.73. Each incident report — or failure to report — becomes a data point in the Art.84 cycle.

Category 5 — Corrective Measure Outcomes. The results of Art.79 (risk identification), Art.80 (Union safeguard), Art.81 (compliant-system risk), Art.82 (formal notification), and Art.83 (CE marking violations) investigations are captured. MSAs report how many investigations they opened, how many led to corrective measures, and what types of measures were imposed.

Category 6 — Cooperation and Referral Activity. MSAs track how many cases involved cross-border coordination with other MSAs or the EU AI Office. Cases where Art.82's cross-border notification was triggered — because a non-compliant system was already found in another member state — are separately counted.

Category 7 — Complaint Statistics. Under the EU AI Act, individuals and organizations can submit complaints about AI systems to MSAs. The number of complaints received, how many led to investigations, and how many were resolved without action are captured in Art.84 reports.

Category 8 — Penalty Data. The number and amount of fines imposed under Art.99 are aggregated. The Art.84 report provides transparency into how aggressively each member state is enforcing the Act — creating a compliance pressure signal that providers can track by jurisdiction.

Why Your Product May Be in the Data

Article 84 data is aggregate and anonymized in the public summary. But the underlying data is not. When an MSA conducts an inspection of your product, the investigation file is retained. If a corrective measure is imposed, the provider name, system name, violation type, and measure taken are all in MSA records. If a cross-border notification under Art.82 was triggered, those records exist in multiple MSA databases and in the EU AI Office's coordination files.

The Art.84 annual report does not name individual products in its public version. But the underlying data that feeds the report is held by public authorities and subject to:

Freedom of information requests — depending on member state law, interested parties may be able to request MSA investigation records
Art.70 confidentiality protections — MSAs must protect commercially sensitive information, but the protection has limits (public interest exceptions apply)
CLOUD Act compellability — see next section

The practical implication: your product's compliance record is not entirely private once an MSA interaction occurs. Designing your AI system to generate clean MSA outcomes — complete documentation, proper registration, timely incident reports — is not just about avoiding fines. It is about the data trail your product creates in the Art.84 reporting pipeline.

Art.84 × Art.70: Confidentiality Limits

Article 70 requires all parties involved in the EU AI Act implementation — including MSAs — to protect confidential information they obtain in the course of their duties. This includes commercial secrets, financial data, and personal data.

The confidentiality protection in Art.70 applies to individual investigation files and the underlying business documentation that providers submit. The Art.84 annual report is exempt from individual confidentiality because it uses aggregate data.

But the interaction between Art.70 and Art.84 creates a nuanced risk for providers:

During an investigation: Your technical documentation, conformity assessment records, and incident reports are submitted to the MSA under Art.70 confidentiality protection. The MSA cannot publish them directly.

After an investigation concludes: The outcomes — whether a corrective measure was imposed, what type, and what the resolution was — are part of the Art.84 data set. Aggregate statistics are public. The specific finding is in the MSA's files.

In cross-border cases: Under Art.82, when an MSA notifies the Commission and other MSAs, your product's non-compliance record is shared across 27 member states' authorities plus the EU AI Office. Art.70 applies to each recipient, but the circle of recipients is wide.

The Art.84 → Art.85 Regulatory Feedback Loop

Article 84 is not merely a reporting obligation — it is the evidence base for regulatory evolution under Art.85.

Article 85 requires the Commission to review the EU AI Act's implementation and submit a report to Parliament and Council by 2 August 2027 (three years after the Act entered into force). That review draws directly from Art.84 annual reports. If the Art.84 data shows:

Systematic enforcement gaps in certain product categories
Sectors with disproportionately high rates of formal non-compliance
Evidence that the Art.5 prohibited AI list needs expansion or amendment
Evidence that the Annex III high-risk AI list needs updating

Then Art.85 provides the mechanism to amend those provisions. The Art.84 → Art.85 loop is how the EU AI Act evolves — and developers who track Art.84 annual reports will have advance notice of which product categories are drawing regulatory attention.

Practical monitoring strategy: The Commission's Art.84 summary is a free, public, annual intelligence signal. When it is published (expected Q2-Q3 of each year starting 2027), developers should check:

Which Annex III categories had the most corrective measures
Which member states are most aggressively enforcing
Whether their sector appears in the systematic issues findings

CLOUD Act × Art.84: The Documentation Compellability Risk

The CLOUD Act (Clarifying Lawful Overseas Use of Data Act, 2018) allows US law enforcement to compel US companies to produce data held on servers anywhere in the world — including in EU data centers operated by US cloud providers.

The Art.84 risk is not primarily about the MSA reports themselves (which are held by EU public authorities, beyond CLOUD Act reach). The risk is about the source documentation that feeds into the Art.84 reporting pipeline.

When your AI system is subject to MSA inspection, the MSA requests your compliance documentation. If that documentation is stored on AWS, Azure, or GCP — even in EU data centers — a US government request under the CLOUD Act could compel the cloud provider to produce it before you have the opportunity to formally submit it to the MSA or assert Art.70 confidentiality protections.

Documentation Type	Art.84 Relevance	CLOUD Act Risk Level
Annex IV Technical Documentation	MSA audit trigger	HIGH — stored in cloud systems
EU Declaration of Conformity	Art.83 formal compliance	HIGH — typically PDF in cloud storage
EUAIDB Registration Data	Art.84 statistics source	LOW — EU AI Office sovereign infrastructure
Serious Incident Reports (Art.73)	Art.84 post-market data	HIGH — often logged in cloud systems
Post-Market Monitoring Plan (Art.72)	Art.84 data category	MEDIUM — depends on storage location
Conformity Assessment Records	Art.43/44 compliance	HIGH — notified body correspondence in cloud

The two-jurisdiction problem: your compliance documentation may simultaneously be:

Subject to EU MSA inspection request (Art.79) requiring disclosure to EU authorities
Subject to US CLOUD Act request requiring disclosure to US law enforcement

EU-native infrastructure — providers incorporated as EU entities with no US parent — eliminates the US vector from this dual-compellability exposure. The data remains under a single legal order.

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


class CloudActRisk(Enum):
    HIGH = "HIGH"
    MEDIUM = "MEDIUM"
    LOW = "LOW"
    NONE = "NONE"


class DocumentationType(Enum):
    ANNEX_IV_TECHNICAL_DOCS = "annex_iv_technical_docs"
    EU_DECLARATION_OF_CONFORMITY = "eu_doc"
    EUAIDB_REGISTRATION = "euaidb_registration"
    SERIOUS_INCIDENT_REPORT = "serious_incident_report"
    POST_MARKET_MONITORING_PLAN = "post_market_monitoring_plan"
    CONFORMITY_ASSESSMENT_RECORDS = "conformity_assessment_records"


CLOUD_ACT_RISK_MATRIX: dict[DocumentationType, dict[str, CloudActRisk]] = {
    DocumentationType.ANNEX_IV_TECHNICAL_DOCS: {
        "aws_eu": CloudActRisk.HIGH,
        "azure_eu": CloudActRisk.HIGH,
        "gcp_eu": CloudActRisk.HIGH,
        "sota_io": CloudActRisk.NONE,
        "hetzner": CloudActRisk.NONE,
        "scaleway": CloudActRisk.NONE,
        "ovhcloud": CloudActRisk.NONE,
    },
    DocumentationType.EU_DECLARATION_OF_CONFORMITY: {
        "aws_eu": CloudActRisk.HIGH,
        "azure_eu": CloudActRisk.HIGH,
        "gcp_eu": CloudActRisk.HIGH,
        "sota_io": CloudActRisk.NONE,
        "hetzner": CloudActRisk.NONE,
    },
    DocumentationType.EUAIDB_REGISTRATION: {
        # EUAIDB is EU AI Office infrastructure — always EU sovereign
        "aws_eu": CloudActRisk.LOW,
        "azure_eu": CloudActRisk.LOW,
        "gcp_eu": CloudActRisk.LOW,
        "sota_io": CloudActRisk.NONE,
    },
    DocumentationType.SERIOUS_INCIDENT_REPORT: {
        "aws_eu": CloudActRisk.HIGH,
        "azure_eu": CloudActRisk.HIGH,
        "gcp_eu": CloudActRisk.HIGH,
        "sota_io": CloudActRisk.NONE,
    },
    DocumentationType.POST_MARKET_MONITORING_PLAN: {
        "aws_eu": CloudActRisk.MEDIUM,
        "azure_eu": CloudActRisk.MEDIUM,
        "gcp_eu": CloudActRisk.MEDIUM,
        "sota_io": CloudActRisk.NONE,
    },
    DocumentationType.CONFORMITY_ASSESSMENT_RECORDS: {
        "aws_eu": CloudActRisk.HIGH,
        "azure_eu": CloudActRisk.HIGH,
        "gcp_eu": CloudActRisk.HIGH,
        "sota_io": CloudActRisk.NONE,
    },
}


@dataclass
class Art84ReportingCalendar:
    """
    Track Art.84 MSA reporting cycles and Commission summary publication.
    
    First cycle: 2027 (covering 2026 enforcement activities after
    the Act reached full application on 2 August 2026).
    """

    reporting_year: int

    def msa_submission_deadline(self) -> date:
        """
        MSAs submit annual reports to Commission.
        EU AI Act does not specify exact date; Q1 following year is industry expectation.
        """
        return date(self.reporting_year + 1, 3, 31)

    def commission_summary_expected(self) -> date:
        """Commission compiles Union-wide summary. Typically Q2 of following year."""
        return date(self.reporting_year + 1, 6, 30)

    def public_release_expected(self) -> date:
        """Commission makes summary available to Parliament, Council, and public."""
        return date(self.reporting_year + 1, 9, 30)

    def art85_review_incorporates(self) -> bool:
        """Whether this reporting year's data feeds the Art.85 Commission review (due 2027-08-02)."""
        return self.reporting_year == 2026


@dataclass
class Art84CompliancePreparation:
    """
    Prepare AI system documentation for Art.84 MSA reporting cycle audits.
    
    MSAs collect 8 categories of data. Each maps to documentation
    you should maintain proactively to generate clean audit outcomes.
    """

    provider_name: str
    system_name: str
    infrastructure_provider: str  # e.g., "aws_eu", "sota_io", "hetzner"
    annex_iii_category: Optional[str] = None

    def assess_cloud_act_exposure(self) -> dict[DocumentationType, CloudActRisk]:
        """
        Return CLOUD Act risk for each documentation type given storage infrastructure.
        
        Art.84 risk: your compliance docs feed MSA reports. US CLOUD Act can
        compel these docs from US cloud providers before MSA Art.70 protections apply.
        """
        return {
            doc_type: risks.get(self.infrastructure_provider, CloudActRisk.LOW)
            for doc_type, risks in CLOUD_ACT_RISK_MATRIX.items()
        }

    def high_risk_documents(self) -> list[DocumentationType]:
        """Return documentation types with HIGH CLOUD Act risk on current infrastructure."""
        exposure = self.assess_cloud_act_exposure()
        return [doc for doc, risk in exposure.items() if risk == CloudActRisk.HIGH]

    def art84_readiness_score(self) -> dict:
        """
        Assess readiness for Art.84 MSA inspection based on infrastructure and category.
        """
        exposure = self.assess_cloud_act_exposure()
        high_risk_count = sum(
            1 for risk in exposure.values() if risk == CloudActRisk.HIGH
        )
        total_docs = len(exposure)
        
        return {
            "provider": self.provider_name,
            "system": self.system_name,
            "infrastructure": self.infrastructure_provider,
            "documents_assessed": total_docs,
            "high_cloud_act_risk_docs": high_risk_count,
            "cloud_act_clean": high_risk_count == 0,
            "recommendation": (
                "No CLOUD Act dual-compellability risk. Documentation under single EU legal order."
                if high_risk_count == 0
                else f"{high_risk_count}/{total_docs} documentation types have HIGH CLOUD Act risk. "
                "Consider EU-native infrastructure to eliminate dual-compellability exposure."
            ),
        }


def generate_art84_audit_checklist(system_name: str, infra: str) -> list[dict]:
    """Generate Art.84-readiness checklist for a given AI system and infrastructure."""
    prep = Art84CompliancePreparation(
        provider_name="provider",
        system_name=system_name,
        infrastructure_provider=infra,
    )
    high_risk = prep.high_risk_documents()

    return [
        {
            "category": "EUAIDB Registration",
            "check": "EUAIDB registration is complete with all mandatory fields",
            "art84_relevance": "MSAs cross-reference EUAIDB against surveilled systems",
            "priority": "P0",
        },
        {
            "category": "Technical Documentation",
            "check": "Annex IV documentation exists and is current",
            "art84_relevance": "Documentation completeness is a tracked MSA metric",
            "priority": "P0",
        },
        {
            "category": "Cloud Act Risk",
            "check": f"High-risk documentation types: {[d.value for d in high_risk]}",
            "art84_relevance": "Dual-compellability before MSA Art.70 protections activate",
            "priority": "HIGH" if high_risk else "NONE",
        },
    ]

The Art.84 Compliance Checklist: 40 Items

Section 1: EUAIDB Registration Readiness (10 Items)

Provider name and contact are registered in EUAIDB with current details
System name and version match deployed system exactly
Intended purpose description in EUAIDB matches Annex IV documentation
Annex III category is correctly identified in registration
Conformity assessment type (self-assessment vs. notified body) is recorded
Notified body number is included if third-party assessment was performed
EU Declaration of Conformity reference is linked in registration
CE marking affixing date is recorded in EUAIDB
Substantial modification notifications have been submitted if system was updated
Registration has been updated after every major version release

Section 2: Technical Documentation Completeness (10 Items)

Annex IV Section 1 — System description and intended purpose is complete
Annex IV Section 2 — Design specifications and development decisions are documented
Annex IV Section 3 — Training data governance records are maintained (Art.10)
Annex IV Section 4 — Testing, validation, and verification procedures are documented
Annex IV Section 5 — Risk management documentation (Art.9) is current
Annex IV Section 6 — Post-market monitoring plan (Art.72) is in place
Annex IV Section 7 — Human oversight measures (Art.14) are documented
Annex IV Section 8 — Accuracy, robustness, cybersecurity measures (Art.15) are covered
Technical documentation was last reviewed within the past 12 months
Substantial modifications have triggered documentation updates as required

Section 3: Post-Market Monitoring and Incident Records (10 Items)

Post-market monitoring plan is operational and generates regular reports
Serious incident reporting process is defined per Art.73 requirements
Serious incidents have been reported to the relevant MSA within required timeframes
Near-miss data is collected to identify emerging risks before they become incidents
Incident classification procedures distinguish Art.73 serious incidents from minor events
User feedback channels are active and integrated into monitoring processes
Monitoring logs are retained for minimum periods required under Art.12
Corrective action tracking records any voluntary self-corrections and their outcomes
Post-market data is reviewed quarterly at minimum for trend identification
Art.73 incident reports are stored in EU-sovereign infrastructure (CLOUD Act risk check)

Section 4: Art.84 Audit Trail Preparation (10 Items)

Compliance documentation is stored in EU-sovereign infrastructure (single legal order)
CLOUD Act risk assessment has been conducted for all documentation storage locations
Art.84 annual report calendar is tracked (first cycle: 2026 activities → 2027 publication)
MSA contact information for relevant member states is known and current
Art.70 confidentiality claims are prepared for commercially sensitive documentation
Legal counsel is identified for potential MSA investigation response
Art.84 summary reports will be monitored annually for sector-level findings
Art.85 review timeline is tracked (Commission review due 2027-08-02)
Compliance documentation versions are tagged and auditable (no undated revisions)
EU AI Office contact information and escalation path is known

Preparing for the First Art.84 Cycle

The first Art.84 annual reports will cover enforcement activities from 2 August 2026 — when the Act reaches full application for high-risk AI systems — through the end of 2026. MSA reports to the Commission are expected by Q1 2027. The Commission summary is expected for publication in H2 2027.

This creates a concrete preparation timeline:

Milestone	Date	Developer Action
EU AI Act full application	2026-08-02	Compliance documentation complete
First MSA inspection window opens	2026-08 onwards	EUAIDB registration verified
MSA reports due to Commission	Q1 2027	Your compliance record is in the data
Commission summary publication	H2 2027	Sector-level findings available
Art.85 review report due	2027-08-02	Regulatory evolution signal available

For developers whose products are under active MSA surveillance or have received corrective measure notices, the Art.84 cycle means your product's compliance history becomes part of the permanent Union-level enforcement record. Proactive compliance — complete documentation, timely registrations, proper incident reporting — generates clean data points in the Art.84 pipeline rather than findings.

Infrastructure Jurisdiction and Art.84

The Art.84 → Art.85 feedback loop has a structural bias: systems with poor compliance documentation generate more MSA findings, more corrective measures, and therefore more Art.84 data points. Systems whose documentation is clean, current, and accessible generate fewer interventions.

Infrastructure jurisdiction is not the only factor, but it is a structural one. When compliance documentation is stored on EU-native infrastructure — German GmbH cloud providers, French SAS operators, no US parent companies — there is no CLOUD Act vector for pre-disclosure access. Your Art.73 incident reports, Annex IV documentation, and conformity assessment records are under a single legal order: EU law.

When that documentation is on AWS Frankfurt, Azure Germany West Central, or GCP Europe West — operated by US parent companies — there is a dual-compellability risk. Not a guaranteed risk, not a certainty, but a structural exposure that EU-native infrastructure eliminates by default.

For high-risk AI systems where Art.84 reporting cycles will include your product, that structural question is worth answering at the infrastructure selection stage rather than after an investigation notice arrives.

Deploy on sota.io — EU-native PaaS, German entity, no CLOUD Act exposure.