Skip to content

EU AI Act Compliance

The EU AI Act (Regulation 2024/1689) establishes the world's first comprehensive legal framework for artificial intelligence. For organizations deploying high-risk AI systems, context governance is no longer optional -- it is a legal obligation. Provena maps directly to six key articles of the regulation, providing the technical controls needed to demonstrate compliance.

Enforcement is active

The EU AI Act entered into force on 1 August 2024. Obligations for high-risk AI systems apply from 2 August 2026. Organizations that fail to implement adequate governance controls face fines of up to EUR 15 million or 3% of global annual turnover.

Article-by-Article Mapping

Article 9 -- Risk Management

What the article requires. Providers of high-risk AI systems must establish, implement, document, and maintain a risk management system. This system must identify and analyze known and reasonably foreseeable risks, estimate and evaluate those risks, and adopt appropriate and targeted risk management measures.

How Provena satisfies it. Provena's three-layer governance pipeline -- provenance validation, freshness checking, and hash-chained audit trails -- constitutes a systematic risk management system for context inputs. Every context entry is validated against configurable provenance requirements (ProvenanceValidator), checked for staleness (FreshnessChecker), and recorded in a tamper-evident chain (ChainHasher). Validation results are classified as VALID, INCOMPLETE, or MISSING for provenance, and FRESH, STALE, or UNKNOWN for freshness, enabling quantitative risk assessment.

trail = ContextTrail(
    required_fields=["source_url", "author", "created_at"],
    max_age_days=90,
    strict_mode=True,  # Governance failures raise exceptions
)

Article 10 -- Data Governance

What the article requires. Training, validation, and testing datasets must be subject to appropriate data governance and management practices. Data must be relevant, sufficiently representative, and free of errors. Datasets must take into account the specific geographical, contextual, behavioral, or functional setting of the system.

How Provena satisfies it. The ProvenanceValidator ensures that every context input carries source metadata -- origin URL, author, creation date, and version. Inputs without provenance are flagged as MISSING; inputs with incomplete metadata are flagged as INCOMPLETE. The FreshnessChecker verifies timestamps against a configurable threshold (max_age_days) and uses temporal pattern detection to identify date references in content text, catching stale data even when metadata timestamps are absent.

from provena.models import ProvenanceMetadata

provenance = ProvenanceMetadata(
    source_url="https://api.example.com/v2/data",
    author="data-engineering-team",
    created_at=datetime(2026, 6, 15, tzinfo=timezone.utc),
    version="2.1.0",
)

record = trail.log(content, source="retriever", provenance=provenance)
# record.provenance_result.status -> "VALID"
# record.freshness_result.status  -> "FRESH"

Article 12 -- Record-Keeping

What the article requires. High-risk AI systems must be designed and developed with capabilities enabling the automatic recording of events (logging) relevant to identifying situations that may result in risk. Logging must enable traceability of the system's operation throughout its lifecycle.

How Provena satisfies it. Provena implements a Merkle-style hash chain using SHA-256 (or HMAC-SHA256 when a signing key is provided). Each TrailRecord contains:

  • A SHA-256 content hash of the input
  • Provenance validation status and missing fields
  • Freshness check status with detected dates
  • A chain hash linking to the previous record
  • Timestamp, source type, and arbitrary metadata

The chain is tamper-evident: modifying any record invalidates all subsequent hashes. The verify_chain() method performs full-chain verification, returning a ChainVerdict that identifies the exact record where integrity was broken.

verdict = trail.verify_chain()
assert verdict.intact  # True if no tampering detected
# verdict.broken_at    -> record ID where chain broke (if any)
# verdict.total_records -> number of records verified

Retention requirements

Article 12 requires log retention for a period appropriate to the intended purpose of the high-risk AI system, which is at least six months unless provided otherwise in applicable Union or national law. Provena's SQLite backend provides persistent storage. For production deployments, enable the OpenTelemetry exporter to forward governance events to durable, centrally-managed storage systems.

Article 13 -- Transparency

What the article requires. High-risk AI systems must be designed and developed in such a way as to ensure that their operation is sufficiently transparent to enable deployers to interpret the system's output and use it appropriately.

How Provena satisfies it. Every governance event recorded by Provena includes an explanation of what was validated and why. Provenance results include the list of missing fields. Freshness results include the detected date, the content age in days, and the threshold that was applied. Source attribution (ContextSource enum: RETRIEVER, TOOL, AGENT, MEMORY, MCP, CUSTOM) identifies the origin type for every input. The trail.summary() method produces aggregate breakdowns by provenance status, freshness status, and source type.

summary = trail.summary()
# {
#   "total": 142,
#   "provenance": {"VALID": 118, "INCOMPLETE": 19, "MISSING": 5},
#   "freshness": {"FRESH": 130, "STALE": 8, "UNKNOWN": 4},
#   "sources": {"retriever": 95, "tool": 30, "agent": 17},
#   "signed": True
# }

Article 14 -- Human Oversight

What the article requires. High-risk AI systems must be designed and developed in such a way as to be effectively overseen by natural persons during their period of use. Human oversight measures must enable the individuals to whom oversight is assigned to properly understand the relevant capacities and limitations of the system and to intervene or interrupt the system as necessary.

How Provena satisfies it. The audit trail records both accepted and flagged context inputs, with detailed reasons for each governance decision. The trail.annotate() method allows human reviewers to attach oversight notes to any record, creating a documented chain of human judgment:

# Reviewer examines a flagged record
flagged = trail.query(provenance_status="INCOMPLETE", limit=10)

for record in flagged:
    # Review and annotate
    trail.annotate(
        record_id=record["id"],
        note="Approved: source verified via internal registry",
        reviewer="compliance-officer@example.com",
    )

The query() method supports filtering by provenance status, freshness status, source type, and time range, enabling compliance officers to focus on records that require human review.

Article 26 -- Deployer Obligations

What the article requires. Deployers of high-risk AI systems must monitor the operation of the system on the basis of the instructions of use. They must keep logs automatically generated by the system, to the extent such logs are under their control, for a period appropriate to the intended purpose, and for at least six months.

How Provena satisfies it. Provena's SQLite backend (storage_path parameter) writes all governance events to persistent local storage. For centralized log management and long-term retention, the OpenTelemetry exporter forwards governance spans to any OTel-compatible backend:

trail = ContextTrail(
    storage_path="/var/lib/provena/audit.db",
    otel_enabled=True,
    otel_service_name="production-rag-pipeline",
)

Six-month minimum retention

The six-month retention requirement under Article 26 is a legal floor. Configure your OTel collector and storage backend to retain governance events for at least this period. Organizations subject to sector-specific regulations (e.g., financial services, healthcare) may face longer retention requirements.

Compliance Mapping

The following table maps specific EU AI Act requirements to Provena features:

EU AI Act Requirement Article Provena Feature
Systematic risk identification and mitigation Art. 9 Three-layer validation pipeline (provenance, freshness, chain integrity)
Data governance and management practices Art. 10 ProvenanceValidator with configurable required fields
Data quality and relevance verification Art. 10 FreshnessChecker with temporal pattern detection
Automatic event logging Art. 12 SHA-256 hash-chained TrailRecord for every context input
Tamper-evident record-keeping Art. 12 Merkle-style chain with verify_chain() integrity verification
Forensic reconstruction capability Art. 12 Full chain replay via query() and export() (JSON/CSV)
Operational transparency Art. 13 Governance event details: validation status, missing fields, detected dates
Source attribution Art. 13 ContextSource enum and source metadata on every record
Human oversight enablement Art. 14 trail.annotate() for reviewer decisions on flagged records
Flagged input visibility Art. 14 query(provenance_status="INCOMPLETE") filtering
Log retention (6-month minimum) Art. 26 SQLite persistent backend + OTel export to durable storage
Operational monitoring Art. 26 trail.health() and trail.summary() for system status

Penalty Structure

The EU AI Act establishes a tiered penalty framework. Understanding which tier applies to your obligations is essential for risk assessment.

Tier Maximum Fine Scope Examples
Tier 1 EUR 35 million or 7% of global annual turnover Prohibited AI practices Social scoring, real-time biometric identification in public spaces
Tier 2 EUR 15 million or 3% of global annual turnover High-risk AI system obligations Inadequate risk management, missing audit trails, insufficient data governance
Tier 3 EUR 7.5 million or 1% of global annual turnover Misleading information Providing incorrect or incomplete information to notified bodies or authorities

Provena's territory is Tier 2

Most obligations addressed by Provena fall under Tier 2. Failure to maintain adequate data governance (Article 10), record-keeping (Article 12), or deployer log retention (Article 26) exposes organizations to fines of up to EUR 15 million or 3% of global annual turnover, whichever is higher.

Enforcement Timeline

Date Milestone
1 August 2024 Regulation enters into force
2 February 2025 Prohibited AI practices apply
2 August 2025 Obligations for general-purpose AI models apply
2 August 2026 High-risk AI system obligations apply (Provena's primary scope)
2 August 2027 Full enforcement for all AI systems, including those in Annex I

Illustrative Scenarios

Hypothetical scenarios, not documented enforcement actions

The two scenarios below are illustrative constructs, not reports of specific, verified regulatory actions. We built them from the penalty tiers and obligations set out in the Act itself and from the failure patterns commonly reported in RAG/agent post-mortems (stale sources, missing provenance, unreconstructable decision trails). No named company, fine amount, or regulator decision described here should be treated as an actual case record. They exist to show concretely what an Article 10/12/13/14/26 gap looks like in practice and how Provena's mechanisms map onto it.

Scenario 1: Wealth Management Firm (Hypothetical)

Consider a Frankfurt-based wealth management firm facing a hypothetical fine in the EUR 1-5 million range from a national financial supervisory authority for violations of EU AI Act Article 10 and Article 12 obligations.

What happened. The firm deployed a RAG-based advisory system that retrieved financial research reports and regulatory guidance to generate client recommendations. An internal audit revealed that the retrieval pipeline was serving research reports up to 14 months old without any staleness detection. The system cited a withdrawn ECB guideline in three client portfolios.

What was missing.

  • No provenance validation on retrieved documents -- the system could not distinguish between current and superseded regulatory guidance
  • No freshness checking -- stale research reports were treated identically to current ones
  • No audit trail -- when regulators requested evidence of what context inputs informed specific client recommendations, the firm could not reconstruct the decision chain

How Provena would prevent it. The FreshnessChecker would flag 14-month-old reports as STALE. The ProvenanceValidator would require source URLs and version identifiers, making it possible to detect when a cited guideline was withdrawn. The hash-chained audit trail would provide the forensic reconstruction capability regulators would demand.

Scenario 2: Insurance Claims Processor (Hypothetical)

Consider a French insurance company whose AI-assisted claims processing system is hypothetically halted by the ACPR (Autorite de Controle Prudentiel et de Resolution, France's insurance and banking supervisor) for violations of Article 12 and Article 26.

What happened. The company deployed an agentic AI system that retrieved policy documents, claims history, and medical guidelines to assist claims adjusters. The system cited a withdrawn clinical guideline in a denial decision, which was challenged by the policyholder. During the resulting investigation, regulators discovered that the company's log retention policy was 30 days -- well below the six-month minimum required by Article 26.

What was missing.

  • No version tracking on retrieved clinical guidelines -- the system could not detect that the guideline had been superseded
  • 30-day log retention violated the six-month minimum under Article 26
  • No tamper-evident logging -- the company could not demonstrate that logs had not been altered after the complaint was filed

How Provena would prevent it. The ProvenanceMetadata.version field and freshness checking would flag the withdrawn guideline. The SQLite backend with OTel export provides persistent, tamper-evident storage that meets the six-month retention requirement. The verify_chain() method provides cryptographic proof that logs have not been modified.

Next Steps