EU AI Act Art.94: Procedural Rights for GPAI Economic Operators — Right to Be Heard, File Access, and AI Office Enforcement Protections (2026)

Post #804 in the sota.io EU Compliance Series

When the AI Office opens a formal investigation into a GPAI model provider, the enforcement process does not proceed in secret. Article 94 of the EU AI Act establishes a structured set of procedural rights that economic operators — GPAI model providers, downstream deployers using GPAI APIs, and distributors in the GPAI supply chain — hold against the Commission during enforcement proceedings. These rights are not optional courtesies. They are binding procedural requirements that must be satisfied before any adverse measure — whether an administrative fine under Art.101, an interim measure, a mandatory withdrawal order, or a finding of systemic risk — can validly be imposed.

Understanding Art.94 is practical preparation, not theoretical compliance. The AI Office began operating in February 2025. The first formal GPAI investigations are expected in the second half of 2026, as GPAI Code of Practice obligations crystallize after the August 2 GPAI deadline. GPAI providers building AI products on European infrastructure need to know what rights they hold and what documentation they must prepare to exercise those rights effectively.

Where Art.94 Sits in the GPAI Enforcement Architecture

The EU AI Act creates a bifurcated enforcement architecture. For AI systems deployed by operators — companies running AI applications in employment decisions, credit scoring, medical devices, or other regulated contexts — enforcement runs through 27 national competent authorities (NCAs) under Art.73–79. For GPAI models — foundation models, large language models, multi-modal systems classified under Art.51 — enforcement runs through a single EU-level body: the AI Office at the Commission, operating under Arts.80–94.

The GPAI enforcement chain within the AI Office framework proceeds as follows:

Art.90 — Information Requests: The AI Office may request GPAI providers to supply information necessary to assess compliance. Providers must respond accurately; providing misleading information itself constitutes a violation subject to separate fines under Art.101(3).

Art.91 — Evaluations: The AI Office may conduct evaluations to assess whether a GPAI model complies with its obligations, including evaluating systemic risk under Art.55. The AI Office may commission independent technical evaluations from qualified third parties.

Art.92 — Interim Measures: Where there is urgent risk of serious, irreparable harm linked to a GPAI model with systemic risk, the AI Office may order interim measures before a final decision — a fast-track power that can restrict a GPAI provider's operations within the EU while the full investigation proceeds.

Art.93 — Commitments: GPAI providers may offer commitments — voluntary compliance measures — that the AI Office can make binding by decision. Accepted commitments terminate ongoing proceedings and replace them with binding undertakings enforceable as Commission decisions.

Art.94 — Procedural Rights: Before any adverse decision — fine, interim measure, mandatory action, systemic risk designation — the AI Office must provide GPAI economic operators with the procedural protections defined in Art.94.

Art.101 — Administrative Fines: The Commission imposes GPAI-specific fines — up to €35 million or 7% of global turnover — for serious GPAI obligation failures, after Art.94 procedural rights have been respected.

Art.94 is the procedural gateway through which all adverse outcomes must pass.

What Article 94 Specifically Requires

The Right to Be Heard

Before the AI Office can adopt any decision imposing obligations, fines, or measures on a GPAI economic operator, Art.94 requires that the affected operator be given a genuine, substantive opportunity to present its case. This is not a notification requirement — it is a participation right.

In practice, this means the AI Office must:

1. Serve a Statement of Objections — a formal document setting out the AI Office's preliminary findings, the specific GPAI obligations allegedly violated, the evidence supporting those findings, and the proposed measures or sanctions under consideration.

2. Allow the GPAI operator a defined period to respond in writing — typically 4–8 weeks depending on the complexity of the case, though the AI Act does not specify exact timeframes (these will be established in AI Office procedural rules).

3. Consider the written response in its final decision. An AI Office decision that ignores substantive arguments raised in the written response is procedurally defective and subject to CJEU annulment.

For GPAI developers, the Statement of Objections is the pivot point of any enforcement proceeding. It is the document around which your defense is constructed. Receiving one means the AI Office has made preliminary findings against you — not final findings. The response period is your opportunity to challenge the factual basis, contest the legal interpretation, offer alternative technical explanations, and propose remedial measures.

Access to the Investigation File

Article 94 entitles GPAI economic operators to access the AI Office's investigation file — the documents the AI Office relied on in reaching its preliminary findings. This right is fundamental to effective defense: you cannot challenge findings based on evidence you have not seen.

File access is not unlimited. Documents may be restricted where they contain:

• Business secrets of third parties — technical specifications, commercial contracts, or financial data of other market participants obtained during the investigation
• Internal AI Office deliberative documents — working papers, internal legal assessments, and draft documents forming part of the deliberative process
• Confidential information obtained from member states — national authority communications that the originating authority has marked confidential

What you are entitled to see is the evidentiary core: the technical reports, evaluation results, expert assessments, third-party complaints (in redacted form), and other factual materials the AI Office uses to support its findings.

For GPAI providers, file access is particularly important in systemic risk investigations. If the AI Office has commissioned an independent technical evaluation of your GPAI model under Art.91, you are entitled to see that evaluation before it is used as the basis for a systemic risk designation or fine. If the evaluation contains methodological errors, inappropriate benchmarks, or mischaracterizes your model's capabilities, you cannot contest those errors without access to the document.

The Right to an Oral Hearing

In addition to written submissions, Art.94 provides GPAI operators the right to request an oral hearing before the AI Office. The oral hearing provides an opportunity to present your position directly, clarify ambiguities in the written record, respond to questions from AI Office officials, and supplement your written defense with live technical explanation.

Oral hearings are particularly valuable in cases involving complex technical issues — fine-tuning procedures, systemic risk thresholds, capability evaluations, copyright compliance methodologies — where written submissions may not fully convey technical nuance. In GPAI enforcement, the difference between a model that presents systemic risk and one that does not may rest on technical details that benefit from direct explanation and follow-up questioning.

Third parties with sufficient interest — downstream deployers who have reported the GPAI model to the AI Office, competitors claiming competitive harm from the model's market behavior, civil society organizations with standing — may also have a right to be heard in the same proceedings under Art.85(5). Your oral hearing may therefore include other parties whose interests conflict with yours.

Confidentiality of Business Secrets

Any information marked as confidential by a GPAI operator — training data composition details, fine-tuning methodologies, safety evaluation procedures, commercial pricing structures, customer lists, technical architecture specifications — must be protected by the AI Office from disclosure to third parties.

Art.94 confidentiality protection applies to documents submitted in response to Art.90 information requests, documents provided as part of the written defense, and documents submitted as part of oral hearing proceedings.

The protection is substantive: the AI Office may not disclose your confidential business information to other parties (including complainants) or to the public in its published decisions. Decisions must be redacted to remove confidential content before publication. Violations of confidentiality by AI Office staff are subject to professional disciplinary procedures under EU Staff Regulations.

For GPAI providers, the confidentiality regime has a practical implication: you must proactively identify and mark confidential material. Submitting documents without confidentiality markings may waive protection. Your legal counsel should establish a document classification protocol at the outset of any Art.90 information request or formal investigation.

The CLOUD Act Problem in GPAI Enforcement Proceedings

The AI Office's evidence-gathering powers under Art.90 and Art.91 may extend to infrastructure and operational data held by GPAI providers. For GPAI providers operating on US-based cloud infrastructure — AWS, Google Cloud, Azure — this creates a structural conflict.

The CLOUD Act (Clarifying Lawful Overseas Use of Data Act) allows US law enforcement to compel US cloud providers to disclose data held in their infrastructure, regardless of where the data is physically located, under 18 U.S.C. § 2703. A GPAI provider operating its model training infrastructure on AWS us-east-1 and its inference infrastructure on Azure cannot guarantee to the AI Office that the confidential technical information provided in enforcement proceedings remains accessible only to EU authorities.

The specific risk vectors for GPAI providers in enforcement proceedings:

Training data audit trails on US infrastructure: The AI Office may request records of training data sources, filtering decisions, and copyright clearance procedures. Where these records are stored in AWS S3 or Google Cloud Storage, US law enforcement may compel their disclosure independently of the EU enforcement proceeding.

Model evaluation logs: Systemic risk evaluations under Art.91 may require access to model safety evaluation records, red-teaming outputs, and internal capability assessments. These records, if stored on US-controlled infrastructure, are accessible to US authorities under the CLOUD Act.

Communications during enforcement proceedings: Legal advice communications between GPAI providers and their counsel, stored in Microsoft 365 or Google Workspace tenants, are reachable under CLOUD Act compulsion, potentially undermining legal professional privilege in EU proceedings.

GPAI providers building on EU-sovereign infrastructure — where data processing occurs on servers under EU jurisdiction with no US-parent company able to receive CLOUD Act orders — eliminate this exposure. Infrastructure jurisdiction is an active variable in enforcement risk, not a background detail.

What GPAI Developers Must Prepare Before Proceedings Begin

Art.94 rights are reactive — they activate once the AI Office takes action. Exercising them effectively requires preparation that happens before any formal investigation opens.

Technical Documentation as Procedural Defense Material

Art.53 requires GPAI providers to maintain technical documentation of their models. That same documentation becomes your primary defense material in Art.94 proceedings. Specifically:

Training data documentation: Records of training dataset composition, data sourcing decisions, filtering procedures, and copyright compliance assessments. If the AI Office alleges Art.53(1)(a) violations (inadequate training data documentation), your contemporaneous records are your defense.

Model capability assessments: Benchmarking records, capability evaluation results, and internal risk assessments. If the AI Office classifies your model as having systemic risk under Art.51, your own capability assessments — showing where your model performs below systemic risk thresholds — are the evidence you need to contest that classification.

Safety and accuracy testing records: Art.55 systemic risk obligations require adversarial testing and incident reporting. Your testing logs, red-teaming records, and incident reports demonstrate compliance with these obligations and pre-empt Art.101 fine exposure.

Copyright policy implementation evidence: Records of the text and data mining opt-out monitoring procedure, RLHF copyright-related training data exclusion decisions, and downstream provider copyright notification. If the AI Office alleges Art.53(1)(c) violations, documented implementation is your defense.

Communication Protocols for Information Requests

When an Art.90 information request arrives, you typically have a defined response window. The response you provide becomes part of the AI Office investigation file — which you are later entitled to access under Art.94. Inconsistencies between your Art.90 responses and your Art.94 defense will be noted.

Establish a single designated point of contact for AI Office communications. All information provided to the AI Office should be reviewed by legal counsel before submission, classified for confidentiality at document level, and archived in your compliance management system.

Legal Representation from Day One

Art.94 does not require GPAI providers to have legal representation, but the procedural complexity of Commission enforcement proceedings — which follow Competition DG procedural models developed over decades of merger control and antitrust enforcement — makes specialist EU regulatory counsel essential. GPAI enforcement is new territory, but the procedural framework it borrows from is not.

Engage EU regulatory counsel before receiving an Art.90 information request if possible. If you are a GPAI provider with models above 10^25 FLOPs (the provisional systemic risk threshold), formal proceedings are a foreseeable event, not a theoretical risk.

Enforcement Timeline: What to Expect

The AI Act does not codify precise enforcement timelines — these will emerge from AI Office procedural rules and practice. Based on analogous Commission enforcement procedures in competition law and digital regulation, GPAI enforcement proceedings are likely to follow approximately this structure:

From Art.90 information request to final Commission decision, a contested GPAI enforcement proceeding may take 18–36 months. Interim measures under Art.92 can be imposed much faster — within days of a proceeding opening — but are themselves subject to Art.94 procedural protections.

Systemic Risk Proceedings: Heightened Stakes for Art.94 Rights

GPAI models classified as having systemic risk under Art.51 face enhanced Art.55 obligations — adversarial testing, incident reporting, cybersecurity measures — and enhanced enforcement exposure. For systemic risk GPAI providers, Art.94 proceedings have higher stakes.

The systemic risk designation itself is an adverse finding that triggers Art.94 rights. Before the AI Office can formally designate your GPAI model as presenting systemic risk — whether based on the 10^25 FLOP computation threshold or on a qualitative capability assessment — you are entitled to respond to the preliminary finding.

The qualitative systemic risk assessment under Art.51(2) is particularly significant. Where the AI Office determines that a GPAI model below the 10^25 FLOP threshold nonetheless presents systemic risk based on its capabilities, the assessment rests on technical judgments about model behavior, downstream reach, and potential impact. These judgments are contestable. GPAI providers whose models are being evaluated for qualitative systemic risk classification have a direct financial interest in exercising their Art.94 right to access the technical evaluation and respond substantively.

SME and Startup Considerations

Article 94 procedural rights apply regardless of the GPAI provider's size. SMEs and startups do not receive reduced procedural protection — nor do they face reduced procedural burden. A startup with a GPAI model above systemic risk thresholds faces the same enforcement proceedings as a large model provider.

However, the EU AI Act's proportionality principle (Art.55(5) and recital 149) requires the AI Office to consider the economic capacity of GPAI providers when setting fine amounts under Art.101. The Art.94 written defense is the mechanism through which SMEs and startups argue for proportionality in fine-setting. A well-documented proportionality argument — demonstrating the company's financial position, the good-faith compliance effort made, and the absence of consumer harm — can be decisive in reducing a nominal fine to a livable sanction.

Python Implementation: Art94ProceedingsTracker

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

class ProceedingStatus(Enum):
    PRE_INVESTIGATION = "pre_investigation"
    ART90_REQUEST = "art90_information_request"
    ART91_EVALUATION = "art91_technical_evaluation"
    STATEMENT_OF_OBJECTIONS = "statement_of_objections_received"
    DEFENSE_PERIOD = "written_defense_period"
    ORAL_HEARING = "oral_hearing_scheduled"
    COMMITMENTS_TRACK = "art93_commitments_track"
    FINAL_DECISION = "final_decision_issued"
    CJEU_APPEAL = "cjeu_appeal"

@dataclass
class Art94Right:
    right_name: str
    description: str
    deadline: Optional[datetime]
    exercised: bool = False
    counsel_notified: bool = False
    notes: str = ""

@dataclass
class Art94ProceedingsTracker:
    proceeding_id: str = field(default_factory=lambda: str(uuid.uuid4()))
    gpai_model_name: str = ""
    status: ProceedingStatus = ProceedingStatus.PRE_INVESTIGATION
    opened_date: Optional[datetime] = None
    counsel_firm: str = ""
    confidentiality_officer: str = ""
    rights: list = field(default_factory=list)

    def initialize_rights(self):
        self.rights = [
            Art94Right(
                right_name="right_to_be_heard",
                description="Submit written defense in response to Statement of Objections",
                deadline=None,
            ),
            Art94Right(
                right_name="file_access",
                description="Request access to AI Office investigation file",
                deadline=None,
            ),
            Art94Right(
                right_name="oral_hearing",
                description="Request oral hearing before AI Office",
                deadline=None,
            ),
            Art94Right(
                right_name="confidentiality_marking",
                description="Mark all submitted documents for confidentiality protection",
                deadline=None,
            ),
            Art94Right(
                right_name="cjeu_appeal",
                description="Lodge Art.263 TFEU annulment action against final Commission decision",
                deadline=None,
            ),
        ]

    def advance_to(self, new_status: ProceedingStatus, event_date: datetime):
        self.status = new_status
        self.opened_date = self.opened_date or event_date
        for right in self.rights:
            if not right.exercised and right.deadline is None:
                right.deadline = event_date  # Set baseline for deadline calculation

    def get_pending_rights(self) -> list:
        return [r for r in self.rights if not r.exercised]

    def compliance_summary(self) -> dict:
        pending = self.get_pending_rights()
        return {
            "proceeding_id": self.proceeding_id,
            "model": self.gpai_model_name,
            "status": self.status.value,
            "total_rights": len(self.rights),
            "rights_exercised": len(self.rights) - len(pending),
            "pending_rights": [r.right_name for r in pending],
            "counsel_engaged": bool(self.counsel_firm),
            "confidentiality_officer_assigned": bool(self.confidentiality_officer),
        }


# Usage example
tracker = Art94ProceedingsTracker(gpai_model_name="MyFoundationModel-v2")
tracker.initialize_rights()
tracker.counsel_firm = "EU Regulatory Counsel GmbH"
tracker.confidentiality_officer = "Chief Legal Officer"
tracker.advance_to(ProceedingStatus.ART90_REQUEST, datetime.now())
print(tracker.compliance_summary())

EU-Sovereign Infrastructure and Art.94 Enforcement Positioning

GPAI providers operating on EU-sovereign infrastructure hold a stronger evidential position in Art.94 proceedings for two reasons.

First, jurisdictional clarity for the AI Office. When investigating a GPAI provider whose entire infrastructure stack is within EU jurisdiction, the AI Office has clear, enforceable access to evidence through EU legal channels. There is no CLOUD Act conflict, no cross-jurisdictional discovery dispute, and no ambiguity about whether US law enforcement can independently access the same evidence the AI Office is reviewing. The proceeding operates entirely within a legal order the AI Office controls.

Second, technical compliance documentation authenticity. Infrastructure logs, training data audit trails, model evaluation records, and safety testing results produced and stored on EU-sovereign infrastructure cannot be alleged to have been accessed or modified by US authorities outside the AI Office's knowledge. The evidentiary chain is clean. For GPAI providers making good-faith compliance arguments in Art.94 proceedings, an unimpeachable audit trail — stored under EU jurisdiction, with verifiable access controls — is a material advantage.

EU-native managed platforms like sota.io — incorporated under German law, operating on EU infrastructure with no US-parent company — allow GPAI model API infrastructure and compliance documentation systems to be deployed under a single EU jurisdiction. Every training audit log, evaluation report, and compliance record remains within the legal order that governs your Art.94 proceedings.

25-Item GPAI Developer Procedural Rights Checklist

Pre-Investigation Preparation

• Designated Art.90 response coordinator identified and trained
• Legal counsel with EU regulatory enforcement experience retained
• Confidentiality officer assigned for document marking protocol
• Art.53 technical documentation current and version-controlled
• Training data audit trail stored on EU-sovereign infrastructure
• Model capability evaluation records archived and accessible
• Safety testing and red-teaming logs maintained with timestamps
• Copyright policy implementation evidence documented
• Art.90 response review protocol established (legal sign-off required)
• CLOUD Act exposure assessment completed for all infrastructure

During Art.90 Information Request

• Art.90 request acknowledged in writing within 5 working days
• All response documents classified and marked for confidentiality
• Counsel review completed before any document submission
• Response archive maintained in compliance management system
• Inconsistencies with prior regulatory submissions identified and addressed

During Formal Investigation

• Art.91 evaluation report access requested immediately upon proceeding opening
• Preliminary findings in Statement of Objections reviewed by counsel within 48h
• Written defense period deadline calendared with review milestones
• Oral hearing requested within the prescribed period
• Systemic risk threshold arguments prepared (technical + proportionality)

During Decision and Appeal

• Final Commission decision reviewed for Art.94 compliance (was defense considered?)
• Art.263 TFEU appeal deadline calendared (2 months from notification)
• Interim measure appeal (Art.278 TFEU) assessed if measure causes immediate harm
• Art.93 commitment offer evaluated as alternative to continued proceedings
• CJEU appeal brief prepared in parallel with commitment negotiations

Conclusion

Article 94 is the procedural shield that stands between a GPAI provider and an AI Office adverse decision. The right to be heard, access to the investigation file, oral hearing rights, and confidentiality protection are not formalities — they are enforceable legal entitlements that, if violated by the AI Office, make the resulting Commission decision voidable before the CJEU.

For GPAI developers, the practical message is clear: procedural rights are only as valuable as your preparation to exercise them. Technical documentation built to Art.53 standards serves double duty — it satisfies compliance obligations and provides the evidentiary foundation for Art.94 defense. Infrastructure deployed under EU jurisdiction eliminates the CLOUD Act variables that complicate enforcement proceedings for US-infrastructure providers. Legal counsel engaged before proceedings begin, not after a Statement of Objections arrives, makes the difference between an effective defense and a reactive scramble.

The AI Office is operational. GPAI enforcement will begin in earnest after August 2026. Art.94 proceedings are not a distant theoretical risk — they are the mechanism through which the GPAI regulatory framework becomes enforceable in practice.

sota.io runs on EU-sovereign infrastructure under German law with no US-parent company — no CLOUD Act exposure for your GPAI compliance documentation. Try it free.