EU AI Act Art.91: AI Office Inspections of GPAI Providers — On-Site and Remote Access Developer Guide (2026)

EU AI Act Article 91 is the AI Office's most intrusive investigative power: the right to physically enter a GPAI model provider's premises, examine systems and documents, take copies, and seal records. If an Art.90 information request is the documentary phase of a GPAI investigation, Art.91 is the access phase. Where Art.90 asks the provider to send documents, Art.91 sends AI Office inspectors to the provider's location — or, in a remote variant, connects directly to the provider's systems.

Art.91 inspections are not routine audits. They are triggered when the AI Office has reason to believe that an Art.90 response was incomplete, when the provider is non-cooperative, when the Scientific Panel has raised a qualified alert requiring physical verification of model infrastructure, or when the AI Office determines that on-site access is necessary to assess systemic risk. The Commission decision that authorises an inspection under Art.91 is itself a formal enforcement act — challengeable before the Court of Justice of the European Union — and the penalties for obstructing an inspection under Art.99(4)(e) are equal in severity to penalties for refusing an Art.90 information request.

For GPAI model providers, Art.91 is the provision that makes data centre location, compute infrastructure sovereignty, and document-management practices directly material to regulatory risk. A provider whose model training infrastructure sits entirely within EU jurisdiction, with AI Office-accessible documentation, faces a fundamentally different inspection exposure than a provider whose model weights, training data, and benchmark results are stored on US cloud infrastructure subject to CLOUD Act jurisdiction.

Art.91 became applicable on 2 August 2025, concurrent with all GPAI obligations in Chapter V.

Art.91 in the GPAI Enforcement Chain

Art.91 sits between the documentary investigation (Art.90) and emergency enforcement measures (Art.93):

In practice, most Art.91 inspections are preceded by an Art.90 request. The information request establishes what documentation exists; the inspection verifies it. However, Art.91 does not legally require a prior Art.90 request — the Commission can authorise an inspection directly where circumstances demand immediate access.

Art.91(1): Commission Decision Authorising Inspection

Unlike the Art.90 simple request, which the AI Office can issue without Commission involvement, an Art.91 inspection always requires a Commission decision. The decision must specify:

1. Legal basis — Art.91 of Regulation (EU) 2024/1689
2. Subject of the inspection — the GPAI model provider and the specific model(s) under investigation
3. Purpose and scope — what the AI Office is seeking to verify
4. Commencement date — when the inspection begins
5. Obstruction consequences — reference to Art.99(4)(e) penalties

The Commission decision is a formal legal act. It can be challenged before the CJEU under Art.263 TFEU, although a challenge does not automatically suspend the inspection obligation. Providers who believe an Art.91 decision is procedurally defective — for instance, because the stated purpose is disproportionately broad — should engage EU regulatory counsel immediately and assess whether an application for interim suspension before the General Court is warranted.

The requirement for a Commission decision (rather than a mere AI Office determination) reflects the severity of on-site inspection as an investigative measure. It creates an institutional check: the AI Office must persuade the Commission that physical access is necessary and proportionate before inspectors can arrive at a provider's premises.

Art.91(2): On-Site Inspection Powers

When authorised by Commission decision, AI Office officials are empowered to:

• Enter any premises of the provider — offices, data centres, computing facilities, storage locations
• Examine books, records, and documents relating to the GPAI model — technical documentation, training data records, model architecture files, benchmark results, evaluation reports, contracts with downstream providers
• Take copies of any documents examined — physical or digital
• Seal premises, books, or records for the duration of the inspection, to the extent necessary and proportionate
• Request oral explanations from any staff member present

The sealing power is particularly significant for compute infrastructure. AI Office inspectors can seal server rooms, restrict access to model weights stored on-site, and prevent modification of training data repositories during the inspection period. Sealing is explicitly limited by proportionality — it cannot be used beyond what the inspection requires — but in practice, a sealed compute environment can halt model operations.

The right to enter "any premises" is not limited to the provider's registered office. It extends to any location where the provider conducts business related to the GPAI model: co-location data centres where the provider has leased rack space, third-party compute providers with which the provider has a contractual relationship for model training, and any facility where model-related documentation is stored.

Art.91(3): Remote Inspection

Art.91 also provides for remote inspection as an alternative to or supplement for on-site access. Remote inspection may be used where:

• The information needed can be obtained without physical presence
• On-site inspection would be disproportionately disruptive
• The provider is located in a jurisdiction where logistics of on-site access are complex
• A targeted verification of specific digital assets can be conducted via secure connection

Remote inspection involves AI Office officials connecting to the provider's systems — model repositories, documentation databases, evaluation environments — via a secure channel under conditions set by the Commission decision. Providers must grant access and cannot selectively filter what is accessible during the remote session.

For providers with EU-based infrastructure, remote inspection is operationally manageable: establish a dedicated secure access environment, ensure documentation is organised and retrievable, and have technical staff available to answer AI Office inspector questions. For providers whose infrastructure is outside the EU, remote inspection creates jurisdictional complexity — the question of which law governs a remote access session that crosses international boundaries is not resolved by Art.91 and requires legal analysis.

Art.91(4): NCA Assistance Obligation

When the AI Office conducts an inspection, National Competent Authorities (NCAs) must provide active assistance to AI Office officials. This assistance obligation has several practical dimensions:

• Facilitation of entry — NCAs can exercise their domestic inspection powers under national law to support AI Office access where the provider attempts to obstruct entry
• Joint inspection teams — NCA officials can accompany AI Office inspectors and participate in the inspection
• Expertise provision — NCAs with sector-specific knowledge (for instance, a financial services NCA where the GPAI model is used in banking applications) can provide technical assistance
• Post-inspection enforcement — where an inspection reveals non-compliance that falls within NCA jurisdiction (for example, high-risk AI system obligations under Art.73), NCAs can initiate parallel enforcement action

The NCA assistance obligation reflects the dual-authority structure of GPAI enforcement: the AI Office has primary jurisdiction over GPAI model providers under Chapter V, but NCAs retain jurisdiction over downstream deployers and high-risk AI system applications. An Art.91 inspection of a GPAI provider may simultaneously reveal information relevant to NCA enforcement actions against downstream deployers — and the NCA assistance structure creates a pathway for that information to flow appropriately.

Art.91(5): Obstruction Penalties

Refusing or obstructing an Art.91 inspection triggers penalties under Art.99(4)(e):

Art.101 adds periodic penalty payments of up to 1.5% of average daily global turnover per day for continued non-compliance. These penalties are cumulative with penalties for underlying substantive violations of GPAI obligations under Art.99(3) (EUR 15,000,000 or 3%).

The obstruction penalties apply to any conduct that impedes the inspection — not only outright refusal to open premises. Selectively withholding access to server rooms, providing inspectors with an employee who lacks knowledge of the relevant systems, or "losing" documentation between an Art.90 response and an Art.91 inspection visit can all constitute obstruction.

Art.90 vs Art.91: The Escalation Matrix

For providers facing an Art.90 information request, the Art.90 response is therefore a critical threshold: a complete, well-organised response that satisfies the AI Office's underlying concerns prevents the investigation from escalating to Art.91. An Art.90 response that appears evasive — even if technically complete — increases the probability of an Art.91 follow-up.

CLOUD Act Intersection: Physical Infrastructure

Art.91 has significant CLOUD Act implications when a GPAI provider's compute infrastructure is hosted on US-based cloud services.

Providers with hybrid infrastructure — model weights in EU data centres, training data in US cloud — should map which components are subject to Art.91 access and whether those components are also reachable by US government requests. The intersection of EU inspection rights and US CLOUD Act authority over the same data is a multi-party legal problem that requires US and EU counsel coordination.

EU-sovereign compute infrastructure eliminates the CLOUD Act conflict entirely for Art.91 purposes: if model weights, training data, and evaluation environments are hosted in EU-jurisdiction infrastructure outside US cloud provider contractual reach, Art.91 inspection access proceeds without jurisdictional complications.

Inspection Preparation Strategy

An Art.91 inspection has three phases for the provider: notice period (between Commission decision and inspection start), inspection execution (during which inspectors are on-site or remotely connected), and post-inspection (documentation of what was reviewed and follow-up actions).

Notice Period

The Commission decision will specify the inspection start date. The notice period — the gap between decision receipt and inspection start — is the preparation window. Use it to:

1. Engage EU regulatory counsel — immediately, before taking any other action
2. Assess CJEU challenge viability — if the decision appears procedurally defective, the challenge must be filed quickly
3. Map scope to available documentation — which documentation does the AI Office's stated purpose suggest they will seek?
4. Organise document repositories — inspectors move faster through well-organised documentation; disorganisation is not obstruction, but it extends inspection duration
5. Identify LPP-protected materials — prepare a privilege log for communications that are legally protected
6. Brief technical staff — engineers and data scientists who may be asked oral questions need to understand what they can and cannot say
7. Assess CLOUD Act exposure — for any requested materials in US-jurisdiction infrastructure, assess whether EU mirrors are available

Inspection Execution

During the inspection:

• Assign an inspection coordinator — a senior member of the legal or compliance team who accompanies inspectors throughout
• Do not volunteer information beyond the stated scope — answer questions precisely, do not expand beyond what is asked
• Do not obstruct — any attempt to limit inspector access beyond legally defensible grounds (e.g., LPP assertion, proportionality objection) should be coordinated with counsel
• Document everything — maintain a contemporaneous log of what inspectors accessed, what copies they took, which staff they questioned, and what was sealed
• Challenge sealing actions immediately if disproportionate — sealing is limited by proportionality; if inspectors attempt to seal business-critical infrastructure beyond inspection necessity, raise a proportionality objection in writing

Post-Inspection

After inspectors depart:

• Review the inspection record — what did they take, what did they question, what did they seal and unseal?
• Assess what may follow — Art.89 right-to-be-heard proceedings typically follow if the AI Office moves toward corrective measures
• Remediate any gaps identified — if the inspection revealed documentation gaps or systemic risk indicators, begin remediation before the AI Office's next step
• Preserve all materials reviewed — do not modify any documentation that was reviewed or copied during the inspection

Python: Art91InspectionResponse Implementation

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


class InspectionMode(Enum):
    ON_SITE = "on_site"
    REMOTE = "remote"
    HYBRID = "hybrid"


class InspectionPhase(Enum):
    NOTICE_PERIOD = "notice_period"
    ACTIVE_INSPECTION = "active_inspection"
    POST_INSPECTION = "post_inspection"
    CLOSED = "closed"


class CloudActRisk(Enum):
    NONE = "none"          # All infrastructure in EU jurisdiction
    LOW = "low"            # Some US-jurisdiction components; EU mirrors available
    HIGH = "high"          # Critical components in US jurisdiction; no EU mirror
    CRITICAL = "critical"  # All model infrastructure in US jurisdiction


@dataclass
class SealedItem:
    item_description: str
    sealed_date: date
    seal_lifted: bool = False
    seal_lifted_date: Optional[date] = None
    proportionality_objection_raised: bool = False


@dataclass
class Art91InspectionResponse:
    commission_decision_reference: str
    decision_received_date: date
    inspection_start_date: date
    inspection_mode: InspectionMode
    gpai_model_name: str
    provider_name: str
    inspection_phase: InspectionPhase = InspectionPhase.NOTICE_PERIOD
    cloud_act_risk: CloudActRisk = CloudActRisk.NONE
    eu_compliance_mirror_available: bool = False
    lpp_privilege_log_prepared: bool = False
    technical_staff_briefed: bool = False
    inspection_coordinator_assigned: bool = False
    documentation_organised: bool = False
    cjeu_challenge_assessed: bool = False
    sealed_items: list[SealedItem] = field(default_factory=list)
    documents_copied: list[str] = field(default_factory=list)
    oral_questions_log: list[str] = field(default_factory=list)

    def notice_days_remaining(self) -> int:
        today = date.today()
        if today >= self.inspection_start_date:
            return 0
        return (self.inspection_start_date - today).days

    def preparation_readiness_score(self) -> dict:
        checks = {
            "EU regulatory counsel engaged": self.lpp_privilege_log_prepared,
            "CJEU challenge viability assessed": self.cjeu_challenge_assessed,
            "Documentation organised for inspection": self.documentation_organised,
            "LPP privilege log prepared": self.lpp_privilege_log_prepared,
            "Technical staff briefed": self.technical_staff_briefed,
            "Inspection coordinator assigned": self.inspection_coordinator_assigned,
            "EU compliance mirror status confirmed": (
                self.cloud_act_risk == CloudActRisk.NONE
                or self.eu_compliance_mirror_available
            ),
        }
        score = sum(15 for v in checks.values() if v)
        return {
            "score": score,
            "checks": checks,
            "notice_days_remaining": self.notice_days_remaining(),
            "preparation_status": "READY" if score >= 90 else "AT_RISK" if score >= 60 else "INADEQUATE",
        }

    def cloud_act_risk_assessment(self) -> str:
        if self.cloud_act_risk == CloudActRisk.NONE:
            return "LOW — All infrastructure in EU jurisdiction. No CLOUD Act conflict."
        if self.cloud_act_risk == CloudActRisk.LOW and self.eu_compliance_mirror_available:
            return "MANAGED — EU mirror available. Supply inspection materials from EU-jurisdiction source."
        if self.cloud_act_risk == CloudActRisk.HIGH:
            return "HIGH — Critical model components in US jurisdiction. Engage US and EU counsel before inspection starts."
        return "CRITICAL — All infrastructure in US jurisdiction. Simultaneous CLOUD Act and Art.91 access creates multi-party conflict. Immediate counsel engagement required."

    def add_sealed_item(self, description: str) -> SealedItem:
        item = SealedItem(item_description=description, sealed_date=date.today())
        self.sealed_items.append(item)
        return item

    def inspection_summary(self) -> dict:
        return {
            "commission_decision": self.commission_decision_reference,
            "provider": self.provider_name,
            "model": self.gpai_model_name,
            "mode": self.inspection_mode.value,
            "phase": self.inspection_phase.value,
            "sealed_items_count": len(self.sealed_items),
            "documents_copied_count": len(self.documents_copied),
            "oral_questions_count": len(self.oral_questions_log),
            "cloud_act_assessment": self.cloud_act_risk_assessment(),
        }

    def preparation_priority(self) -> str:
        days = self.notice_days_remaining()
        if days == 0:
            return "ACTIVE — Inspection in progress. Follow execution protocol."
        elif days <= 3:
            return "CRITICAL — Finalise all preparation immediately."
        elif days <= 7:
            return "HIGH — Complete documentation organisation and staff briefing."
        elif days <= 14:
            return "MEDIUM — Preparation underway; CJEU challenge window closing."
        else:
            return "NORMAL — Systematic preparation phase."

15-Item Art.91 Inspection Preparation Checklist

Art.91 Series Context

Art.91 is part of the AI Office's GPAI investigation toolkit under Chapter V:

Art.91 represents the point at which a GPAI investigation transitions from documentary to physical verification. For providers who have managed their Art.90 response well and have documented compliance with GPAI obligations under Arts.51-55, an Art.91 inspection is a manageable, if disruptive, process. For providers with undocumented model development practices, training data of uncertain provenance, or benchmark results that diverge from the Art.53 technical documentation, an Art.91 inspection can rapidly escalate to Art.93 interim measures or Art.89 corrective proceedings.

See Also

• Art.90: AI Office Information Requests to GPAI Providers — Developer Response Guide — Art.90 is the documentary stage that precedes Art.91 inspections; an adequate Art.90 response prevents escalation to physical access.
• Art.89: Right to Be Heard Before Enforcement Measures — Developer Response Guide — The Art.89 procedural right applies after an Art.91 inspection leads to proposed corrective measures or penalties.
• Art.93: AI Office Interim Measures for Systemic Risk — Art.91 inspection findings can trigger Art.93 emergency measures where systemic risk is confirmed.
• Art.55: AI Office Evaluation of GPAI Models and Scientific Panel Powers — Scientific Panel qualified alerts are a primary trigger for both Art.90 requests and Art.91 inspections.
• Art.92: AI Office Monitoring of GPAI Market Developments — Reporting, Public Disclosure, and International Standards Developer Guide — Art.92 market monitoring intelligence is what identifies which providers need Art.91 inspections; inspection findings in turn feed back into Art.92 monitoring reports.