ETSI Domain AI + Data Mgmt

Cross-domain marker for metaverse components combining artificial intelligence with data management capabilities including ML pipelines, intelligent data processing, analytics, and AI-driven data governance.

Bridge-To

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:hasPart ai:MLPipelineClassification))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:hasPart ai:IntelligentAnalyticsCategorization))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:hasPart ai:MachineLearningOperations))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:hasPart ai:FeatureStore))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:hasPart ai:DataGovernanceFramework))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:hasPart ai:DataLineageSystem))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:hasPart ai:MLBillOfMaterials))

Dependency Relationships

SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:requires ai:DataPipeline))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:requires ai:FeatureEngineering))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:requires ai:ModelRegistry))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:requires ai:DataLake))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:requires ai:StreamProcessing))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:requires ai:ConsistencyProtocols))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:requires ai:DistributedSystems))

Capability Relationships

SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:enables ai:MLPipelineClassification))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:enables ai:IntelligentAnalyticsCategorization))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:enables ai:DataAnalytics))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:enables ai:MachineLearningOperations))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:enables ai:FeatureStore))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:enables ai:DataLineage))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:enables ai:AIDriverDataGovernance))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:enables ai:RealTimeAnalytics))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:enables ai:AnomalyDetection))

Implementation Relationships

SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:implements ai:EUAIActDocumentationRequirements))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:implements ai:ETSIISGENIArchitecture))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:implements ai:ISOIEC42001AIManagementSystem))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:implements ai:MLBillOfMaterials))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:implements ai:DataGovernanceFramework))

Support Relationships

SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:supports ai:Metaverse))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:supports ai:DigitalTwin))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:supports ai:AutonomousSystems))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:supports ai:PersonalisationEngine))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:supports ai:ContentRecommendation))

Usage Relationships

SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:uses ai:DeepLearning))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:uses ai:MachineLearning))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:uses ai:DataAnalytics))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:uses ai:NaturalLanguageProcessing))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:uses ai:ReinforcementLearning))

Reduction Relationships

SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:reducesTo ai:ETSIDomainAI))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:reducesTo ai:ETSIDomainDataManagement))
SubClassOf(ai:ETSIDomainAIDataMgmt
  ObjectSomeValuesFrom(ai:reducesTo ai:CrossDomainStandardisationMarker))

About

ETSI Domain AI + Data Mgmt occupies the intersection of two foundational work streams within ETSI’s metaverse and telecommunications standards landscape: the ETSI Domain AI classification covering machine learning model deployment, AI system architecture, and intelligent service orchestration, and the ETSI Domain: Data Management classification covering data storage, synchronisation, lifecycle management, and distributed systems for virtual environments. The cross-domain intersection captured by this marker is not merely additive; it identifies standardisation requirements that emerge specifically from the combination — where AI systems actively participate in constructing, operating, governing, and optimising data pipelines rather than simply consuming data produced by independent data management systems.

The domain’s technical roots lie in the development of Machine Learning Operations (MLOps) as a discipline spanning data engineering, model development, deployment, and monitoring — and in ETSI’s ETSI ISG ENI (Experiential Networked Intelligence) system, which was defined in 2017 as an AI engine providing intelligent service management for network operations, maintenance, security, and orchestration. ENI’s architecture — comprising a context-aware policy management layer, input processing and normalisation pipeline, internal AI module, and data storage framework — constitutes the canonical ETSI technical reference for AI-data management integration, demonstrating how AI systems consume, transform, and act upon network management data in closed-loop control processes. The data engineering pipeline in ENI’s architecture maps directly to modern MLOps concepts: raw data flows through cleaning and feature processing in the input processing and normalisation block, enters the AI module for model inference or training, and produces outputs that are stored in the data storage framework for subsequent policy application. This architecture is directly applicable to metaverse data management contexts where AI systems must process continuous streams of virtual world state data, user behaviour telemetry, and environment sensor feeds to produce intelligent orchestration decisions.

The domain’s relevance to metaverse applications derives from the characteristic data management challenges of persistent, large-scale virtual environments at network scale. A functional metaverse generates and consumes data at volumes and velocities that cannot be managed through static, manually configured data pipelines: AI-driven Data Processing must dynamically partition and route data streams, adaptive Feature Engineering must extract relevant signals from heterogeneous environment telemetry, and intelligent Anomaly Detection must identify technical failures and policy violations within continuous data flows. The standardisation requirements that emerge from these challenges include: latency specifications for AI-driven data processing in real-time virtual environment state management (sub-millisecond for physics simulation synchronisation, tens of milliseconds for behavioural analytics); consistency protocols for distributed AI data pipelines operating across multi-region metaverse infrastructure; Feature Store specifications enabling efficient serving of pre-computed features to AI inference workloads at the ApplicationLayer; and ML Pipeline Classification frameworks that systematically categorise pipeline configurations (batch, stream, hybrid; online learning vs. offline batch; centralised vs. federated) to enable interoperability and conformity assessment.

The domain’s regulatory context is shaped by the EU AI Act’s documentation and transparency requirements for AI systems. Systems classified as high-risk under the Act must maintain technical documentation including descriptions of training data, data processing methodologies, and data governance measures — effectively mandating the adoption of Data Lineage tracking and ML Bill of Materials (ML-BOM) tooling. ETSI’s standards work in this domain therefore addresses not only the technical architecture of AI data pipelines but the documentation, audit trail, and transparency infrastructure required for regulatory compliance. ETSI TS 104 008, which introduces a framework for Continuous Auditing-Based Conformity Assessment aligned with the EU AI Act, is directly relevant to AI + Data Mgmt standardisation, providing the conformity assessment framework within which data governance and pipeline documentation standards operate.

Historical Development and Context

The emergence of ETSI Domain AI + Data Mgmt as a distinct cross-domain classification reflects the historical trajectory of intelligent network management standards at ETSI, which progressively incorporated data-driven and AI-powered approaches to network operations automation over the period 2014–2026. The foundational work was established by ETSI’s ENI (Experiential Networked Intelligence) ISG, created in 2017 with a mandate to define AI-driven network management architectures that could replace manual and rule-based network operations with closed-loop autonomous systems. ENI’s initial work identified the data management challenge as fundamental: building an AI engine capable of driving network policy management decisions required not merely an inference capability but a complete data pipeline ecosystem — ingestion, normalisation, feature extraction, model training, evaluation, deployment, and continuous monitoring — all operating in the demanding telecommunications network management context of high availability, low latency, and regulatory accountability.

The parallel development of the ETSI ISG MEC (Multi-Access Edge Computing) framework from 2014 onward established the compute substrate within which AI data pipelines in the domain operate. MEC Phase 1 and Phase 2 established the fundamental edge hosting environment and service APIs. Phase 3 (completed April 2024) addressed the complex heterogeneous cloud ecosystem that characterises modern network edge deployments, introducing federation, multi-domain, and multi-tenancy capabilities that enable AI services to operate across organisational and infrastructure boundaries. Phase 4 (current) explicitly incorporates AI-native edge operations as a priority, addressing the AI inference service exposure and AI data pipeline management requirements that the AI + Data Mgmt domain captures.

The metaverse context added new dimensions to the AI data management challenge beyond the telecommunications network management origins of ENI and MEC. Metaverse platforms generate heterogeneous data at volumes and velocities that significantly exceed typical telecommunications network telemetry: user interaction event streams measured in billions of events per day, spatial state data for persistent virtual environments spanning terabytes of continuously updated geometry and physics state, economic transaction data from virtual economies operating in near-real-time, and media asset metadata for content libraries comprising hundreds of millions of user-generated and platform-generated items. AI systems that operate within this data environment — performing real-time personalisation, content moderation, economic fraud detection, quality of experience optimisation — require data management infrastructure of a different character than telecommunications network management, combining the low-latency requirements of operational telecommunications systems with the scale requirements of consumer internet platforms and the compliance requirements of regulated data environments. The ETSI Domain AI + Data Mgmt cross-domain marker emerged to capture precisely this combination of requirements, identifying standards work items that address all three dimensions simultaneously.

Components and Architecture

The standardisation scope of ETSI Domain AI + Data Mgmt encompasses several interacting technical component families, each representing a distinct set of standards requirements:

• ML Pipeline Classification: a systematic framework for categorising machine learning pipeline configurations deployed in metaverse and telecommunications contexts. Classification dimensions include: pipeline topology (batch, stream, lambda, kappa); learning paradigm (offline/batch training, online learning, continual learning, federated learning); data modality (structured telemetry, unstructured media, multimodal); hardware target (cloud, edge, on-device); and lifecycle stage (data ingestion, feature engineering, training, evaluation, deployment, monitoring). ML Pipeline Classification enables interoperability between pipeline components from different vendors and provides the vocabulary for conformity assessment specifications under the EU AI Act. ETSI’s 3GPP alignment activities (AI/ML in 3GPP spanning 2017–2030) extend this classification into telecommunications network operations management contexts.

• Intelligent Analytics Categorization: systematic taxonomy and evaluation frameworks for AI-powered analytics systems deployed in metaverse and network management contexts. Covers descriptive analytics (real-time dashboards, anomaly reporting), diagnostic analytics (root cause identification using AI pattern recognition), predictive analytics (user behaviour forecasting, infrastructure failure prediction), and prescriptive analytics (AI-generated optimisation recommendations and automated action). In metaverse contexts, Intelligent Analytics Categorization encompasses player behaviour analytics, content consumption pattern analysis, virtual economy monitoring, and infrastructure telemetry intelligence. Standards requirements include: analytics accuracy thresholds for operational decision-making, latency specifications for time-sensitive analytics (network slice management, quality of experience optimisation), privacy-preserving analytics specifications for user behaviour data under GDPR and ePrivacy Regulation constraints, and Explainable AI requirements for analytics outputs used in automated policy decisions.

• Machine Learning Operations (MLOps): the set of practices, tools, and cultural norms that streamline and standardise the deployment, monitoring, and maintenance of machine learning models in production metaverse and network management environments. MLOps standards requirements include: model versioning and Model Registry specifications, CI/CD pipeline integration specifications for ML model deployment, model performance monitoring and drift detection protocols, and rollback and canary deployment specifications for AI model updates in production virtual environments. As of 2025, over 60% of enterprises prioritise integrated governance in MLOps-DevOps flows, and the ETSI work programme explicitly includes MLOps as a standardisation area in its 2024–2025 work programme. ETSI’s definition of MLOps encompasses the Machine Learning Bill of Materials (ML-BOM), a specialised component inventory for AI systems cataloguing models, datasets, parameters, and training configurations — directly analogous to the Software Bill of Materials (SBOM) concept in software supply chain security.

• Feature Store: a data management component that maintains pre-computed feature vectors for efficient serving to AI inference workloads, decoupling feature engineering computation from feature consumption. Feature stores in metaverse contexts must maintain features at multiple freshness levels (historical aggregates for batch analytics, near-real-time features for session-level personalisation, millisecond-fresh features for real-time game AI) and serve them with latency compatible with the target application (sub-millisecond for real-time game AI, hundreds of milliseconds for recommendation engines). Standards requirements include: feature metadata schemas (feature names, types, computation frequency, data lineage), feature serving API specifications, feature versioning and reproducibility specifications for model retraining, and privacy access control specifications for features computed from user behaviour data.

• Data Governance for AI systems: the policies, processes, technical controls, and organisational structures that ensure data quality, security, privacy, and regulatory compliance across AI data pipelines. In the ETSI Domain AI + Data Mgmt context, data governance specifically addresses: training data provenance tracking (what data was used to train which model, when, and under what licensing conditions); feature computation auditing; model output data retention and deletion; and the intersection with EU AI Act Article 17 requirements for quality management systems in high-risk AI systems. Modern data governance frameworks as of 2025 align with IEC 42001 (AI management systems), NIST AI RMF, or the EU AI Act’s quality management requirements, providing the standards alignment backbone for ETSI-specific work.

• Data Lineage tracking: automated recording of data provenance chains across AI pipeline stages, enabling retrospective analysis of model outputs, compliance audit trails, and debugging of data quality failures. Data Lineage tools track the origin, transformations, and movement of data from source systems through feature engineering, training, and inference stages. Standards requirements include: lineage metadata schemas, lineage API specifications for tool interoperability, lineage storage and query performance specifications, and lineage report formats for regulatory compliance documentation. Gartner projects that by 2026, 60% of large enterprises will have deployed data lineage tools for regulatory and operational risk — a trajectory that ETSI standardisation must accommodate with interoperability specifications enabling lineage tooling from different vendors to exchange lineage metadata.

• ETSI ISG ENI system architecture: the primary ETSI technical reference for AI-data management integration, defining the closed-loop AI architecture for intelligent network service management. ENI’s Context-Aware Policy Management system, input processing and normalisation pipeline, internal AI module, and data storage framework constitute the reference architecture for ETSI AI + Data Mgmt standards, extended from network management to metaverse environment management contexts in Phase 4 and subsequent ETSI work.

  Use Cases and Major Families

  The cross-domain marker applies to several distinct deployment families within metaverse and telecommunications infrastructure:

  Virtual world state analytics and optimisation: AI-driven analytics systems process continuous telemetry streams from persistent virtual environments — player positions, interaction events, economic transactions, environmental state changes, infrastructure load metrics — to detect anomalies, optimise resource allocation, and generate operational intelligence for platform operators. Standards requirements include: telemetry data schema specifications for inter-platform telemetry portability, stream processing latency specifications for real-time anomaly detection, and privacy-preserving analytics specifications to enable aggregate intelligence without exposing individual user behaviour. Intelligent Analytics Categorization provides the classification framework for systematically defining the capabilities and limitations of AI analytics systems deployed in this use case.

  ML model lifecycle management for metaverse AI services: AI services within metaverse platforms — NPC behaviour models, content recommendation engines, quality of experience prediction models, anti-cheat detection systems — require lifecycle management infrastructure to develop, deploy, monitor, update, and retire models in production environments. Machine Learning Operations standards define the interface specifications, quality gates, and audit trail requirements for ML model lifecycle management in metaverse platform engineering contexts. Key standards requirements include: model deployment rollback specifications to protect against degraded model performance in live environments, model performance monitoring thresholds triggering automated retraining or alerting, and ML Bill of Materials documentation for AI systems deployed in metaverse contexts requiring EU AI Act conformity assessment.

  Federated and privacy-preserving AI training on metaverse data: The geographically distributed nature of metaverse infrastructure — with users and compute assets across multiple jurisdictions with different data protection regimes — creates strong incentives for federated learning approaches that train AI models on distributed data without centralising raw data. Standards requirements include: federated learning coordination protocol specifications (aggregation server and client protocols), differential privacy noise calibration specifications for federated training, cross-jurisdiction data governance specifications for federated metaverse analytics, and performance specifications for federated model convergence relative to centralised training baselines.

  AI-driven network slice management for metaverse applications: 3GPP’s AI/ML standardisation activities include Management Data Analytics Services that use AI to analyse network operational data and recommend or automate network management decisions — directly applicable to managing the network slices required for metaverse application delivery (guaranteed low-latency slices for real-time physics synchronisation, best-effort slices for background asset streaming, high-bandwidth slices for spatial audio delivery). ETSI Domain AI + Data Mgmt standardisation provides the data management layer within which these AI-driven network management systems operate, specifying the data pipeline architectures that enable real-time AI inference on network telemetry.

  Intelligent data quality management for AI training pipelines: Metaverse platforms accumulate large volumes of user-generated data that are used as training material for NPC behaviour models, recommendation systems, and content moderation classifiers. Maintaining the quality, representativeness, and compliance of this training data requires AI-driven data quality monitoring, automated detection of data distribution shifts, and intelligent deduplication and filtering pipelines. Standards requirements include: data quality metric specifications for AI training datasets, automated data quality monitoring API specifications, distribution shift detection algorithm specifications, and data curation documentation requirements for EU AI Act-compliant AI system technical files.

  Digital twin data intelligence: Digital twins of physical infrastructure — smart buildings, urban infrastructure, industrial plants — integrated with metaverse representations generate rich operational telemetry that requires AI-driven analysis to produce actionable intelligence. ETSI Domain AI + Data Mgmt covers the standardisation of AI analytics and ML pipeline components within digital twin data architectures, including the AI inference workloads that process Digital Twin sensor data, the Feature Store components that serve pre-computed features to digital twin analytics applications, and the Data Governance frameworks that manage the sensitive operational data flowing through digital twin systems.

  Cross-Domain Relationships and Taxonomy Navigation

  Understanding ETSI Domain AI + Data Mgmt requires situating it precisely within the ETSI domain taxonomy and clarifying the standardisation boundaries with adjacent domains. Like all ETSI cross-domain markers, it is an intersection point rather than an independent domain — ensuring that standards addressing the AI-data management convergence are discoverable from both parent domain branches and from related cross-domain intersection points.

  ETSI Domain: Data Management (no AI qualifier): the parent Data Management domain covering data storage, synchronisation, lifecycle management, database systems, consistency protocols, and caching infrastructure for distributed virtual environments without an AI-specific focus. Standards addressing traditional data management infrastructure (relational database specifications, distributed consistency protocols, CDN caching specifications) fall here but not in AI + Data Mgmt. Standards addressing AI-enhanced versions of these functions — intelligent query optimisation using ML, AI-driven cache eviction policies, predictive data pre-fetching using user behaviour models — fall at the intersection and carry both the Data Management domain tag and the AI + Data Mgmt cross-domain marker.

  ETSI Domain AI (parent, without Data Mgmt qualifier): the parent AI domain covering all AI/ML capabilities in the metaverse taxonomy. Standards addressing general AI inference infrastructure fall in the AI domain but are not specifically tagged AI + Data Mgmt unless they address data pipeline or data management requirements that are specific to AI systems and differ from general-purpose data management. Standards addressing ML pipeline data ingestion, feature store specifications, model training data governance, or online learning data stream management fall at the intersection and carry the AI + Data Mgmt cross-domain marker.

  ETSI Domain AI + Creative Media (sibling cross-domain marker): the parallel AI-Creative Media intersection. Creative AI systems have specific data management requirements — large creative training dataset management, provenance chain management for AI-generated artefacts, feature stores for creative style and content features — that overlap with AI + Data Mgmt scope. Standards addressing training data management for generative creative AI models may carry both AI + Creative Media and AI + Data Mgmt markers.

  ETSI Domain AI + Governance (sibling cross-domain marker): the AI-Governance intersection. AI data management governance — data lineage for EU AI Act compliance, bias assessment for training data, ML-BOM documentation requirements — overlaps with AI + Governance scope. Standards specifically addressing the governance mechanisms implemented within AI data pipelines (audit trail requirements for EU AI Act compliance, bias monitoring in feature engineering pipelines, access control specifications for training data) may carry both AI + Data Mgmt and AI + Governance markers.

  ETSI Domain: Data Management + AI: a distinct cross-domain marker in the ETSI taxonomy representing the same intersection from the data management domain’s perspective rather than the AI domain’s perspective. The two markers — AI + Data Mgmt and Data Management + AI — are complementary navigational entry points to the same standards scope; standards carrying one marker will typically also carry the other, but the marker choice affects discoverability from different taxonomy entry points. The distinction is analogous to whether a standard is filed under “AI systems that manage data” or “data management systems that use AI” — both descriptions apply to the same underlying technical specifications.

  The ETSI ISG ENI architecture and the ETSI ISG MEC Phase 4 AI inference service specifications are the primary technical vehicles carrying the AI + Data Mgmt cross-domain marker, as they are the ETSI standards most directly concerned with the integration of AI and data management capabilities in telecommunications and metaverse infrastructure contexts. Future 3GPP AI/ML Release 19 and ETSI Phase 5 MEC specifications are expected to carry the marker as the standards elaboration of AI-native network data management continues.

  Evaluation Frameworks and Quality Benchmarks

  Establishing quality standards for AI data management systems requires formal evaluation frameworks that span both the technical pipeline performance dimensions (latency, throughput, reliability) and the AI-specific quality dimensions (model performance, data quality, governance compliance). ETSI Domain AI + Data Mgmt standardisation must engage with both sets of evaluation requirements, producing specifications that enable objective comparison and conformity assessment of competing AI data management implementations.

  For ML pipeline performance, the primary evaluation dimensions are: data ingestion throughput (records per second), feature computation latency (median and 99th percentile), model training time (wall clock hours per training epoch on reference hardware), model serving latency (median and 99th percentile inference time), and pipeline availability (uptime percentage measured against defined service level objectives). These dimensions map directly to telecommunications network management quality metrics — throughput, latency, reliability — that ETSI has long-established experience in standardising through its ENI and MEC work programmes, providing a natural bridge from telecommunications quality engineering to AI pipeline quality engineering.

  For data quality within AI pipelines, evaluation dimensions include: completeness (percentage of non-null values for required fields), consistency (absence of contradictory values across records and tables), accuracy (conformance to source-of-truth references where available), timeliness (latency between event occurrence and data availability in the pipeline), and representativeness (statistical coverage of the intended population of use cases). Data quality metrics for AI training data additionally include bias assessments: demographic representation analysis for training datasets that underpin AI systems making decisions about people (content moderation classifiers, personalisation systems), distribution coverage analysis for training datasets that represent rare events or edge cases, and documentation completeness assessments for datasets used in EU AI Act high-risk AI system technical files.

  For federated learning pipeline evaluation, specific quality benchmarks address convergence: the round-to-accuracy curve (model accuracy as a function of federated communication rounds), communication efficiency (model accuracy per unit of communication bandwidth consumed), and heterogeneity tolerance (accuracy degradation as a function of statistical and computational heterogeneity across federated clients). These benchmarks are important for ETSI standards specifying federated learning pipelines for cross-platform metaverse analytics, where client heterogeneity — different platform architectures, dataset sizes, and data distributions across metaverse providers — is an inherent architectural characteristic.

  For MLOps pipeline governance quality, evaluation dimensions include: model documentation completeness (percentage of required ML-BOM fields populated), data lineage coverage (percentage of training data with full upstream provenance chains recorded), model performance monitoring coverage (percentage of production models with active performance monitoring enabled), and mean time to detect model degradation (MTTD for drift and performance alerts). These governance quality metrics support continuous auditing under ETSI TS 104 008 and EU AI Act conformity assessment, enabling quantitative benchmarking of AI data management governance maturity across organisations and implementations.

  Academic Context

  The academic foundations of ETSI Domain AI + Data Mgmt draw from machine learning systems research, data engineering, and distributed systems communities. The ENI architecture’s closed-loop AI design mirrors control theory principles applied to network management, with roots in autonomic computing research (Kephart and Chess, 2003). The development of MLOps as a formal discipline was catalysed by Sculley et al.’s landmark paper “Hidden Technical Debt in Machine Learning Systems” (NIPS 2015), which identified the systemic engineering challenges of maintaining ML systems in production — data dependencies, model staleness, monitoring gaps — that MLOps practices and the standards covering them are designed to address.

  The arXiv paper “Governing Cloud Data Pipelines with Agentic AI” (2024, arXiv:2512.23737) proposes frameworks for AI-driven governance of data pipelines, directly relevant to the AI + Data Mgmt domain’s focus on intelligent data pipeline management. The extensive MLOps research literature covers specific components: feature store design (Zaharia et al., 2018, on Feature Store for ML), model monitoring and drift detection (Gama et al., 2014; Bayram et al., 2022), and federated learning (McMahan et al., 2017) — all contributing to the academic foundations of standards work in this domain.

  The ETSI ENI architecture has generated academic publication through its working group outputs, documented in IEEE ComSoc Technology Blog coverage of ENI Release 2 (2022) and the OECD AI catalogue entry for ETSI GR ENI 009 (Data Processing Mechanisms). The 3GPP AI/ML research activities, coordinated between ETSI and 3GPP, have produced academic output through the “Next Decade of Telecommunications AI” survey (arXiv:2101.09163), providing the telecommunications systems context within which ETSI AI + Data Mgmt standards operate.

  UK academic contributions come from several leading institutions. The University of Manchester’s School of Computer Science hosts significant ML systems and data engineering research, with the newly opened £120 million AI research hub providing infrastructure for large-scale ML pipeline research directly relevant to this domain. The Alan Turing Institute’s data science programmes, with industrial partners including the Office for National Statistics, GCHQ, and major financial institutions, address data governance and ML pipeline standards challenges that feed into ETSI UK delegation positions. UCL’s Centre for Artificial Intelligence, which co-leads the UKRI generative AI hub, contributes data infrastructure research with partners including IBM and BT. The University of Edinburgh’s Informatics faculty has historically contributed to data management and distributed systems research with direct relevance to ML pipeline architecture standards. Newcastle University’s Digital Institute and the PETRAS National Centre of Excellence for IoT Systems Cybersecurity contribute AI data governance research relevant to edge-deployed AI systems in the telecommunications and metaverse context.

  Benchmark Datasets and Reference Implementations

  Standards work in AI + Data Mgmt requires reference implementations and benchmark datasets enabling objective comparison of AI data pipeline components. Key resources include:

  ML pipeline benchmarking:

• MLPerf Training: industry benchmark suite from MLCommons measuring ML training throughput on standard hardware configurations; provides reference performance baselines for training pipeline efficiency specifications

• MLPerf Inference: benchmark measuring inference latency, throughput, and efficiency for deployed ML models on various hardware targets; directly relevant to edge inference deployment specifications within ETSI ISG MEC

• TPC-DS: the Transaction Processing Performance Council’s decision support benchmark extended to include ML workloads; used for data warehouse and analytics pipeline performance specification

  Data governance and lineage reference implementations:

• Apache Atlas: open source metadata management and governance platform; reference implementation for data lineage tracking in Hadoop/Spark-based AI data pipelines

• OpenLineage: open standard specification (IETF format) for tracking data and job lineage across data pipeline executions; the most widely adopted open lineage specification for AI pipeline documentation

• Marquez (LinkedIn Open Source): metadata service for tracking job runs, dataset lineage, and data quality metrics; reference implementation for OpenLineage integration

  Feature store reference implementations:

• Feast: open source feature store supporting both online (Redis/DynamoDB) and offline (BigQuery/Snowflake) feature serving; primary open source reference implementation for feature store architecture standards

• Hopsworks: managed and open source feature store with integrated model registry; provides reference implementation for feature group versioning and point-in-time correct feature retrieval

• Tecton: managed feature platform; reference implementation for enterprise feature computation, storage, and serving specifications

  Federated learning frameworks:

• TensorFlow Federated (TFF): Google’s open source framework for federated learning research and production; primary reference implementation for federated aggregation protocol specifications

• PySyft (OpenMined): privacy-preserving ML framework supporting differential privacy, secure multi-party computation, and federated learning; reference implementation for privacy-preserving AI data pipeline specifications

• FATE (Federated AI Technology Enabler, WeBank): enterprise-grade federated learning framework with explicit focus on regulatory compliance; widely deployed in financial services federated learning contexts relevant to metaverse financial applications

  Model registry and MLOps reference implementations:

• MLflow: open source platform for the complete ML lifecycle including experiment tracking, model registry, and deployment; the most widely deployed open source reference implementation for MLOps specifications

• Weights & Biases (wandb): experiment tracking and model registry platform; widely used in research and production ML pipeline management

• Kubeflow: Kubernetes-native ML workflow platform providing reference architecture for cloud-native ML pipeline deployment specifications

  Current Landscape (2026)

  By mid-2026, the ETSI Domain AI + Data Mgmt standardisation landscape reflects rapid maturation of MLOps as a discipline and increasing regulatory pressure from the EU AI Act’s documentation requirements. ETSI EN 304 223 (2025) provides baseline security requirements for AI systems, including data security in training and processing pipelines — directly applicable to the AI data management domain. ETSI TS 104 008 introduces a Continuous Auditing-Based Conformity Assessment framework aligned with the EU AI Act, providing the auditing infrastructure within which data pipeline documentation and model provenance tracking must operate. The ETSI Work Programme 2024–2025 explicitly includes AI/ML data management standardisation as a priority area, reflecting demand from ETSI member organisations in the telecommunications industry who are deploying ML pipeline infrastructure for network automation.

  ETSI’s OpenSlice 2025Q4 release represents a major step toward AI-native autonomous network operations, delivering integrated AI/ML capabilities within the network slice management infrastructure — a concrete deployment of the AI + Data Mgmt domain’s principles in production telecommunications infrastructure. OpenSlice’s AI-driven service orchestration integrates ML model serving, telemetry analytics, and automated policy management within a single operational framework, demonstrating the feasibility of the integrated AI-data management architecture that ETSI Domain AI + Data Mgmt standards must support.

  The broader MLOps market reflects increasing enterprise investment in AI data pipeline governance: over 57% of surveyed organisations have implemented bias detection, and 45% use drift monitoring tools integrated into MLOps pipelines (State of AI Governance survey, 2025). Major MLOps platforms — MLflow, Weights & Biases, Vertex AI, Azure ML, SageMaker — have converged on common workflow patterns that reflect emerging community consensus on ML pipeline architecture, providing the industrial context within which ETSI’s normative standards must operate to achieve adoption. ISO/IEC 42001:2023 (AI Management Systems) and the NIST AI RMF provide the international governance framework alignment for ETSI AI + Data Mgmt standards, with the EU AI Act’s implementing acts expected to specify the harmonised standards (including ETSI outputs) applicable to AI system technical documentation requirements through 2026–2028.

  UK Context

  The UK’s position in ETSI Domain AI + Data Mgmt standardisation reflects its strengths in AI infrastructure research, financial services AI deployment, and telecommunications innovation. The BSI (British Standards Institution) routes UK national positions into ETSI technical committees, and UK contributions in this domain are shaped by the distinctive UK policy context: the ICO’s AI auditing framework, the FCA’s model risk management guidance for financial services AI, and the broader UK AI Safety Institute’s technical safety research. The UK’s National AI Strategy and the 2025 AI Opportunities Action Plan committed £2 billion for compute infrastructure under the Compute Roadmap, establishing the AI Growth Zones framework that provides the industrial deployment context for AI data pipeline infrastructure standards.

  Northern English industrial contributions to this domain are grounded in the smart manufacturing, digital infrastructure, and financial technology sectors. The University of Manchester’s Alliance Manchester Business School AI and Data Science research group addresses AI data governance challenges in industrial deployment, with the £120 million Manchester AI research hub (opened 2024) providing large-scale ML pipeline infrastructure. Leeds’s financial technology cluster — including several major banks’ analytics and ML operations centres — deploys AI data pipeline infrastructure at scale, providing industrial intelligence relevant to ETSI standards development. Sheffield’s Advanced Manufacturing Research Centre (AMRC) and the Henry Royce Institute contribute AI-driven data management research for manufacturing and materials contexts, with direct relevance to digital twin data management standards. The Newcastle University Digital Institute addresses data governance and intelligent analytics in public sector and smart infrastructure contexts, while the Sunderland Software City cluster contributes practical MLOps deployment experience from the north east’s growing technology sector.

  London and southern UK academic and industrial contributions are particularly strong. The Alan Turing Institute, co-located with the British Library and connected to a network of thirteen university partners, coordinates cross-institutional AI data governance research and actively contributes to BSI/ETSI standards alignment through its data science and AI safety programmes. UCL’s Centre for Artificial Intelligence, with BT as an industrial partner, addresses AI data management challenges in telecommunications contexts directly relevant to ETSI standardisation. Imperial College London’s Data Science Institute contributes federated learning and privacy-preserving analytics research applicable to distributed metaverse data pipeline architectures. King’s College London’s informatics group contributes healthcare AI data governance research that models best practice for AI data management in regulated high-stakes environments — a template for analogous standards requirements in metaverse contexts involving sensitive user data. Cambridge’s Computer Laboratory and its industrial collaborators (including Arm and British Telecom) contribute hardware-software co-design research for efficient ML inference at the network edge, relevant to the ETSI ISG MEC deployment layer that underlies AI + Data Mgmt edge inference standards.

  Scottish contributions include: the University of Edinburgh Informatics faculty’s machine learning systems research (AIAI — Artificial Intelligence Applications Institute — and the Bayes Centre), which addresses ML pipeline architecture and intelligent data management challenges; the University of Strathclyde’s Centre for Intelligent Dynamic Communications (CIDCOM), which contributes telecommunications AI and data management research relevant to the ETSI ENI and MEC Phase 4 work; and Digital Scotland’s investment in AI infrastructure for public sector data management, providing government deployment context for intelligent analytics standardisation.

  Future Directions (2026–2030)

  Several structural trajectories are expected to shape ETSI Domain AI + Data Mgmt standardisation through 2030:

  (1) EU AI Act implementing acts and documentation standards: The EU AI Act’s implementing acts, expected through 2026–2028, will specify harmonised standards for AI system technical documentation — creating a substantial pipeline of ETSI standards work covering:

  • ML Pipeline Classification schema specifications for Annex IV technical documentation

  • ML Bill of Materials schema standards for cataloguing AI system components

  • Data Lineage tracking specifications enabling retrospective compliance audit

  • Training data documentation standards for EU AI Act high-risk system technical files

  • Continuous monitoring specifications for deployed AI data management systems under TS 104 008

    (2) ISO/IEC 42001 alignment and management system standards: Alignment of ETSI AI + Data Mgmt standards with ISO/IEC 42001 (AI Management Systems) will enable organisations seeking AI management system certification to use ETSI specifications as implementation guidance, creating commercial incentives for ETSI adoption beyond regulatory mandate. The 42001 framework’s emphasis on risk-based data governance provides natural alignment with ETSI’s domain classification approach.

    (3) Federated and privacy-preserving ML standards: Federated learning will become increasingly prominent as metaverse platforms scale across jurisdictional boundaries. ETSI standardisation will address:

  • Federated learning coordination protocol specifications (aggregation server and client APIs)

  • Differential privacy noise calibration specifications for federated training data pipelines

  • Cross-jurisdiction data governance interoperability specifications (GDPR vs. UK GDPR vs. data localisation requirements)

  • Federated model convergence quality specifications relative to centralised training baselines

  • Secure aggregation protocol specifications for federated learning in adversarial network environments

    (4) Data Mesh and Data Fabric architectural standards: The emergence of Data Mesh (domain-centric data ownership with federated governance) and Data Fabric (AI-driven unified data access layer) patterns in large-scale metaverse data platforms will require ETSI standards for the AI components operating within these architectures:

  • Intelligent data catalogue specifications (AI-driven metadata management and semantic search)

  • Automated data quality monitoring API specifications for distributed data product architectures

  • AI-driven data lineage propagation specifications across mesh domain boundaries

  • Federated governance policy specification schemas for cross-domain data access control

    (5) Digital twin AI data integration standards: The convergence of AI data management with Digital Twin infrastructure will create new standards requirements as digital twin platforms generate rich operational telemetry processed by real-time AI analytics. ETSI Phase 4 and successor MEC work, alongside 3GPP Release 19 AI/ML management activities, will address:

  • Digital twin telemetry data schema specifications for AI analytics compatibility

  • Real-time feature extraction specifications for digital twin sensor data streams

  • AI inference latency requirements for digital twin decision-support applications

  • Cross-domain digital twin data federation specifications

    (6) Agentic AI data pipeline governance: The emergence of agentic AI systems — autonomous agents executing multi-step data management tasks without per-step human supervision — creates new standards requirements for pipeline governance and safety:

  • Workflow authorisation protocol specifications for agentic data pipeline operations

  • Action audit trail specifications for autonomous data manipulation events

  • Rollback and compensating transaction specifications for agentic ML pipeline operations

  • Safety constraint specification frameworks for autonomous data management agents

    (7) Real-time AI data infrastructure standards: By 2030, real-time feature serving, continuous model retraining, and AI-driven data lifecycle management will be standard metaverse platform infrastructure components. Standards covering:

  • Sub-millisecond feature store serving specifications for real-time game AI workloads

  • Continuous learning pipeline monitoring specifications detecting performance degradation

  • Online model update coordination protocols for distributed metaverse AI services

  • Multi-vendor AI data pipeline interoperability specifications enabling ecosystem diversity

    (8) Explainable AI in data pipeline standards: Integration of Explainable AI requirements into data pipeline specifications — ensuring AI data quality decisions, feature engineering choices, and model selections are auditable and interpretable — will reflect pressure from the EU AI Act, financial sector AI governance (EBA AI governance guidelines), and healthcare AI deployment frameworks:

  • Explainability metadata specifications for automated data quality decisions

  • Feature importance documentation specifications for model technical files

  • Human-readable model decision trail specifications for high-stakes AI applications

  • Counterfactual explanation generation specifications for rejected or anomalous data records

    Formal Standards Landscape

    The ETSI Domain AI + Data Mgmt cross-domain marker sits at the intersection of several concurrent standards streams, each addressing different aspects of the AI-data management integration challenge. Navigating this landscape requires understanding both the normative standards that impose specific technical requirements and the informative technical reports that provide architecture guidance and terminology alignment.

    At the ETSI level, the ETSI ISG ENI (Experiential Networked Intelligence) series constitutes the primary ETSI technical reference for AI-driven data management in telecommunications and networked systems contexts. ETSI GR ENI 009 V1.1.1 defines the data processing mechanisms for ENI systems, establishing the baseline architecture (input processing, normalisation, AI module, data storage, output management) that subsequent work items extend and refine. ETSI GR ENI release 2 documentation (2022) introduced the Context-Aware Policy Management system and intent-aware network architecture, adding a policy representation layer atop the data management foundation. ETSI White Paper No. 64 “AI Technologies in Experiential Networked Intelligence to Increase Autonomous Operation” provides the conceptual framework situating ENI’s AI data management architecture within the broader autonomic operations literature.

    The ETSI ISG MEC Phase 4 work programme contributes to the domain through its focus on AI inference service exposure at the network edge — specifically the APIs and service specifications that enable AI workloads to consume and produce data within multi-access edge computing infrastructure. MEC Phase 4’s federation, multi-domain, and multi-tenancy work items address the data management challenges of shared edge infrastructure where multiple AI services compete for data access, processing capacity, and storage. The ISG MEC Terms of Reference v2.5 (October 2024) governs Phase 4 scope, with AI-native edge operations a priority area.

    The 3GPP AI/ML standards track, with which ETSI coordinates through joint activities, covers Management Data Analytics Services (MDAS) and AI/ML Life Cycle Management (LCM) frameworks directly relevant to the AI + Data Mgmt domain. 3GPP’s AI/ML roadmap spans 2017–2030 with progressive releases addressing data collection, model training coordination, model inference deployment, and feedback loop management for telecommunications network operations — constituting a parallel but aligned standards track to ETSI’s ENI work, with ETSI providing higher-level architectural context and 3GPP providing protocol-level implementation specifications.

    At the regulatory level, the EU AI Act’s documentation requirements for high-risk AI systems (Article 11, Annex IV) mandate training data descriptions, data processing methodologies, and data governance measures — directly driving demand for standards in training data provenance (Data Lineage), dataset documentation (ML Bill of Materials), and data quality monitoring (continuous monitoring specifications). ETSI TS 104 008 (Continuous Auditing-Based Conformity Assessment) and ETSI EN 304 223 V2.1.1 (AI cybersecurity baseline) together provide the ETSI standards framework within which data management conformity assessment operates. ISO/IEC 42001:2023 (AI Management Systems) provides the international management system standard that data governance frameworks under ETSI Domain AI + Data Mgmt must align with, enabling organisations to satisfy both ISO 42001 certification requirements and EU AI Act technical documentation obligations through a coherent data management standards programme.

    Technical Variants and Pipeline Families

    The ML pipeline configurations that fall within ETSI Domain AI + Data Mgmt span several distinct architectural families, each requiring specific standards treatment:

    Batch ML pipelines: offline training and periodic batch inference pipelines that process data in large blocks on a scheduled basis. Characteristic of analytics use cases (weekly user behaviour summaries, monthly content performance reports) and model retraining workflows where data accumulates before each training cycle. Standards requirements include: batch data schema versioning specifications (ensuring schema compatibility across training cycles), batch processing latency requirements for scheduled jobs that must complete within defined maintenance windows, and data retention specifications governing how long batch training data must be preserved for model reproducibility and audit purposes.

    Stream processing ML pipelines: real-time processing pipelines that consume continuous data streams (event streams, sensor feeds, user interaction events) and produce inference outputs with low latency. Characteristic of anomaly detection, real-time personalisation, and adaptive quality-of-experience optimisation in metaverse contexts. Standards requirements include: stream processing latency thresholds for time-sensitive applications (sub-second for real-time anomaly detection, sub-100ms for interactive quality adaptation), event ordering and deduplication specifications for distributed stream sources, windowing function specifications (tumbling, sliding, session windows) for time-based feature aggregation, and exactly-once processing guarantee specifications for financial and compliance-critical stream applications.

    Lambda and Kappa architectures: hybrid data pipeline architectures that maintain parallel batch and stream processing layers (lambda) or unify processing through a single replayable stream (kappa), providing both real-time query capability and historical batch processing. Widely adopted in large-scale metaverse data platforms that must serve both real-time operational analytics and historical reporting workloads from a single data infrastructure. Standards requirements include: consistency specifications between batch and stream processing layers in lambda architectures, stream replay specifications for reprocessing historical data in kappa architectures, and state management specifications for long-running stateful stream processing operations.

    Federated Learning pipelines: distributed ML training pipelines where model updates are computed locally on distributed data sources and aggregated centrally without moving raw data. Particularly relevant to cross-platform metaverse analytics where user data sovereignty and GDPR compliance preclude centralising raw data. Standards requirements include: federated aggregation protocol specifications (FedAvg, FedProx, and variants), differential privacy calibration specifications for noise injection during local model update computation, communication efficiency specifications (gradient compression, update sparsification) for bandwidth-constrained edge deployments, and convergence quality specifications comparing federated model performance against centralised baselines.

    Online learning pipelines: ML pipelines where models are continuously updated from incoming data without periodic retraining cycles, adapting to distribution shifts in real time. Relevant to metaverse contexts where user behaviour patterns shift rapidly (new game releases, platform events, seasonal trends) and batch-retrained models quickly become stale. Standards requirements include: online learning stability specifications (preventing catastrophic forgetting of previously learned patterns), update rate specifications (balancing adaptation speed against stability), rollback specifications for reverting online updates that degrade model performance, and monitoring specifications for detecting when online learning has diverged from expected performance envelopes.

    Feature Engineering and Feature Store pipelines: specialised data management infrastructure for computing, storing, versioning, and serving feature vectors to ML models. Feature stores decouple the computationally intensive feature engineering step from the latency-sensitive model serving step, enabling feature reuse across multiple models and ensuring consistency between training-time and serving-time feature computation. Standards requirements include: feature store API specifications (point-in-time correct feature retrieval, batch feature materialisation, online feature serving), feature metadata schema specifications (feature name, data type, computation frequency, data lineage, access controls), feature versioning specifications for reproducibility, and feature freshness specifications defining acceptable staleness for different feature categories.

    Governance, Ethics, and Regulatory Context for AI Data Management

    The governance of AI data management systems in ETSI’s metaverse domain taxonomy is shaped by data protection law, AI regulatory requirements, and sector-specific governance frameworks that together define the compliance envelope within which standards operate.

    Data protection and privacy: The EU General Data Protection Regulation (GDPR) and the UK Data Protection Act 2018 establish the primary legal framework for personal data processing in AI data pipelines. AI data pipelines in metaverse contexts process substantial volumes of personal data — user location history, interaction events, behavioural patterns, biometric identifiers (voice, facial appearance for avatar generation) — requiring technical implementation of GDPR principles (data minimisation, purpose limitation, storage limitation, accuracy, integrity) within data pipeline architectures. Standards requirements include: purpose-binding specifications for data collected in metaverse contexts (ensuring data used for analytics cannot be repurposed for training without appropriate consent), retention period enforcement specifications for automated data deletion after defined retention periods, and pseudonymisation and anonymisation specifications for privacy-preserving analytics. The ETSI ISG ENI architecture’s context-aware policy management system provides a relevant technical model for implementing purpose-binding policies within AI data pipelines, with policies specifying permitted data uses, retention limits, and access controls.

    AI system documentation and auditability: The EU AI Act’s Article 11 and Annex IV requirements for high-risk AI system technical documentation impose specific data management obligations: records of training data (origin, collection methods, selection criteria, data cleaning methodologies), data preprocessing documentation, and training data governance measures. These obligations directly drive adoption of Data Lineage tooling, ML Bill of Materials frameworks, and dataset documentation schemas — all within the scope of ETSI Domain AI + Data Mgmt standardisation. The ETSI TS 104 008 Continuous Auditing-Based Conformity Assessment framework provides the audit architecture within which these documentation obligations are verified, specifying the audit trail and evidence collection requirements for conformity assessment of AI data management systems.

    Data quality and bias governance: AI systems in high-stakes metaverse applications (identity verification, content moderation, personalised recommendation with commercial implications) must maintain data quality standards that prevent the introduction of systematic bias through data pipeline failures. Standards requirements include: data quality metric specifications (completeness, accuracy, consistency, timeliness) for AI training datasets, bias assessment methodology specifications for evaluating distributional representativeness of training data, and data quality monitoring specifications for detecting data drift (shifts in data distribution between training time and serving time that can degrade model performance and introduce unintended bias). ISO/IEC 23053 (Framework for Artificial Intelligence Systems Using Machine Learning) and ISO/IEC TR 24027 (Bias in AI Systems) provide the international standards vocabulary for bias and data quality governance that ETSI AI + Data Mgmt standards must align with.

    Model governance and lifecycle accountability: Beyond data governance, AI data management standards must address the governance of models themselves: who approved a model for production deployment, what data was used to train it, how its performance is monitored, when it should be retrained or retired, and what the evidence trail is for each lifecycle event. The ML Bill of Materials (ML-BOM) concept — analogous to the Software Bill of Materials (SBOM) in software supply chain security — catalogues models, datasets, parameters, training configurations, and third-party component dependencies, enabling downstream users to assess supply chain risk and regulatory compliance. ETSI standardisation in this domain includes ML-BOM schema specifications, model card documentation standards (following the Mitchell et al. model card framework), and model deployment authorisation workflow specifications that ensure appropriate governance gates before AI models enter production metaverse platform operations.

    Key Terminology

• ML Pipeline Classification: systematic taxonomic framework for categorising machine learning pipeline configurations by topology (batch, stream, hybrid), learning paradigm (offline, online, federated), data modality, hardware deployment target, and lifecycle stage; enables interoperability assessment and conformity evaluation of ML pipeline components across different vendor implementations.

• Intelligent Analytics Categorization: structured taxonomy for AI-powered analytics systems distinguishing descriptive (reporting), diagnostic (root cause), predictive (forecasting), and prescriptive (recommendation/automation) analytics; provides the classification vocabulary for specifying analytics capabilities in metaverse platform SLAs and conformity assessment documentation.

• Machine Learning Operations (MLOps): the engineering discipline encompassing the practices, tools, and cultural norms for operationalising machine learning models at scale — including data pipeline management, model versioning, automated testing, deployment orchestration, performance monitoring, and model lifecycle governance; the primary operational framework within which AI + Data Mgmt standards are applied.

• Feature Store: a specialised data management component that maintains pre-computed, versioned, and governed feature vectors for efficient serving to AI inference workloads, decoupling the computationally intensive feature engineering step from the latency-sensitive model serving step; critical for achieving consistent model performance between training and serving environments.

• Data Lineage: automated recording of data provenance chains across pipeline stages, tracking the origin, transformations, and movements of data from source systems through feature engineering, model training, and inference; required for EU AI Act technical documentation, debugging, and compliance audit trails.

• ML Bill of Materials (ML-BOM): a specialised inventory framework for AI systems cataloguing models, datasets, training configurations, parameters, and third-party dependencies in a standardised machine-readable format; analogous to Software Bill of Materials (SBOM) in software supply chain security, enabling supply chain risk assessment and regulatory compliance verification for AI systems.

• Feature Engineering: the process of constructing informative, relevant, and discriminative feature representations from raw data for use as inputs to ML models; in metaverse contexts, encompasses extracting behavioural signals from user interaction event streams, computing spatial features from virtual world geometry data, and deriving economic indicators from virtual economy transaction data.

• Data Governance: the collection of policies, processes, technical controls, and organisational roles that ensure data quality, security, privacy, and regulatory compliance throughout the data lifecycle; in AI data management contexts, specifically includes training data governance (provenance, quality, bias assessment) and serving data governance (access control, purpose limitation, retention enforcement).

• Data Mesh: a socio-technical architectural pattern that decentralises data ownership to domain teams, provides a self-serve data platform for data product creation, and applies federated governance to ensure cross-domain data interoperability and quality; relevant to large-scale metaverse platforms with multiple product teams managing domain-specific data assets.

• Model Registry: a versioned repository for storing, tracking, and governing trained ML models, recording training metadata (dataset versions, hyperparameters, evaluation metrics) and deployment history (which model version served which workload at which time); a core component of MLOps infrastructure and a prerequisite for EU AI Act-compliant AI system technical documentation.

• Data Lake: a centralised repository that stores raw data at any scale in its native format (structured, semi-structured, unstructured) enabling future processing by diverse analytics and ML workloads; common in large-scale metaverse platforms as the primary landing zone for high-volume telemetry, user-generated content, and environment state data.

• Anomaly Detection: AI techniques for identifying unusual patterns, outliers, or deviations from expected behaviour in data streams or datasets; deployed in metaverse contexts for infrastructure failure detection, fraud and cheating identification in virtual economies, policy violation detection in user-generated content streams, and data quality monitoring in AI training pipelines.

• ETSI ISG ENI: the ETSI Industry Specification Group on Experiential Networked Intelligence, which defined the ENI closed-loop AI architecture for intelligent network service management and produced the primary ETSI technical references for AI-data management integration in telecommunications systems contexts; the foundational ETSI technical group for the AI + Data Mgmt domain.

  Standards References and Key Documents

  The following are the primary ETSI and cross-body standards documents within scope of ETSI Domain AI + Data Mgmt:

  ETSI primary documents:

• ETSI GR ENI 009 V1.1.1: Experiential Networked Intelligence (ENI) — Definition of data processing mechanisms

• ETSI GR ENI Series: ENI system architecture, context-aware policy management, intent-aware network architecture

• ETSI EN 304 223 V2.1.1 (2025-12): Securing Artificial Intelligence (SAI) — Baseline cybersecurity requirements (includes data pipeline security)

• ETSI TS 104 008: Continuous Auditing-Based Conformity Assessment Framework (aligned with EU AI Act)

• ETSI TS 104 223 V1.1.1 (2025-04): Securing Artificial Intelligence (SAI)

• ETSI TR 104 128 V1.1.1 (2025-05): Securing Artificial Intelligence (SAI) — AI Threat Ontology

• ISG MEC Terms of Reference v2.5 (October 2024): Phase 4 scope definition including AI-native operations

• ETSI Work Programme 2024–2025: AI/ML data management standardisation work items

  3GPP aligned standards:

• 3GPP TR 28.809: Management Data Analytics Services for AI/ML in network management

• 3GPP TR 28.812: Study on enhancement of AI/ML LCM for network management (Life Cycle Management)

• 3GPP TR 28.908: Management and Orchestration aspects of AI/ML network operations

  Cross-body standards in scope:

• ISO/IEC 42001:2023: Artificial intelligence management system — Data governance requirements

• ISO/IEC 22989:2022: Artificial intelligence — Concepts and terminology (MLOps vocabulary)

• ISO/IEC TR 24027:2021: Bias in AI systems — Data representativeness assessment

• ISO/IEC TR 23053:2022: Framework for AI systems using machine learning — Data requirements

• NIST AI RMF 1.0 (2023): AI Risk Management Framework — Data governance alignment

• EU Regulation 2024/1689 (EU AI Act): Article 11 (technical documentation), Annex IV (training data)

• EU GDPR Articles 5, 17, 25: Data quality, erasure, and privacy-by-design for AI data pipelines

• UK Data (Use and Access) Act 2025: AI training data transparency review provisions

  Key technical references:

• ETSI White Paper No. 64: AI Technologies in ENI to Increase Autonomous Operation

• ETSI White Paper No. 59: Enabling MEC in IoT — AI data service deployment context

• AI Standards Hub ETSI/ISG ENI Archive: https://aistandardshub.org/search_committee_reference/etsi-isg-eni/

• ML-BOM Community Specification (SPDX ML Profile): Machine Learning Bill of Materials schema

  Research and Literature

  1. Sculley, D., Holt, G., Golovin, D., Davydov, E., Phillips, T., Ebner, D., Chaudhary, V., Young, M., Crespo, J.F., & Dennison, D. (2015). Hidden technical debt in machine learning systems. NeurIPS, 28.
  2. McMahan, B., Moore, E., Ramage, D., Hampson, S., & Agüera y Arcas, B. (2017). Communication-efficient learning of deep networks from decentralized data. AISTATS 2017.
  3. Zaharia, M., Chen, A., Davidson, A., Ghodsi, A., Hong, S.A., Konwinski, A., … & Xin, R. (2018). Accelerating the machine learning lifecycle with MLflow. IEEE Data Engineering Bulletin.
  4. Kephart, J.O., & Chess, D.M. (2003). The vision of autonomic computing. IEEE Computer, 36(1), 41–50.
  5. Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., & Bouchachia, A. (2014). A survey on concept drift adaptation. ACM Computing Surveys, 46(4).
  6. Bayram, F., Ahmed, B.S., & Kassler, A. (2022). From concept drift to model degradation: An overview on performance-aware drift detectors. Knowledge-Based Systems, 245, 108632.
  7. ETSI. (2022). ETSI GR ENI 009 V1.1.1: Experiential Networked Intelligence (ENI) — Definition of data processing mechanisms. European Telecommunications Standards Institute.
  8. ETSI. (2025). ETSI EN 304 223 V2.1.1: Securing Artificial Intelligence (SAI) — Baseline cybersecurity requirements. European Telecommunications Standards Institute.
  9. ETSI. (2025). ETSI TS 104 008: Continuous auditing-based conformity assessment framework. European Telecommunications Standards Institute.
  10. ETSI. (2025). ETSI Work Programme 2024–2025: Advancing Standards. https://www.etsi.org/images/files/WorkProgramme/ETSI-Work-Programme-2024-2025.pdf
  11. ETSI. (2025). ETSI White Paper No. 64: AI Technologies in Experiential Networked Intelligence to Increase Autonomous Operation.
  12. IEEE ComSoc Technology Blog. (2022). ETSI Release 2 of Experiential Networked Intelligence (ENI) Explained. https://techblog.comsoc.org/2022/01/15/etsi-release-2-of-experiential-networked-intelligence-eni/
  13. Montojo, J. (2025). Overview of AI/ML related work in 3GPP. ETSI AI Conference, 11 February 2025. Qualcomm / 3GPP RAN presentation.
  14. OECD AI Observatory. (2022). ETSI GR ENI 009 V1.1.1. OECD AI Tools and Resources Catalogue. https://oecd.ai/en/catalogue/tools/etsi-gr-eni-009-v-111
  15. ResearchGate. (2019). Network management and orchestration using artificial intelligence: Overview of ETSI ENI. IEEE Communications Standards Magazine.
  16. Zhou, Z., Chen, X., Li, E., Zeng, L., Luo, K., & Zhang, J. (2019). Edge intelligence: Paving the last mile of artificial intelligence with edge computing. Proceedings of the IEEE, 107(8), 1738–1762.
  17. Li, L., Xu, G., Du, L., et al. (2021). The next decade of telecommunications artificial intelligence. arXiv:2101.09163.
  18. Governing Cloud Data Pipelines with Agentic AI. (2024). arXiv:2512.23737.
  19. ISO/IEC. (2023). ISO/IEC 42001:2023 — Information technology: Artificial intelligence management system. International Organisation for Standardisation.
  20. NIST. (2023). AI Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology.
  21. Gartner. (2025). Data lineage market forecast: Enterprise adoption projections 2023–2026. Gartner Research.
  22. EWS. (2025). AI and Data Governance: The Essential 4-Pillar Framework for 2025. https://www.ewsolutions.com/ai-and-data-governance/
  23. Quinnox. (2025). Data governance for AI in 2025: Challenges, best practices and solutions. https://www.quinnox.com/blogs/data-governance-for-ai/
  24. AI Standards Hub. (2025). ETSI/ISG ENI archives. https://aistandardshub.org/search_committee_reference/etsi-isg-eni/
  25. Kellton. (2026). AI Tech Stack 2026: Frameworks, MLOps & IDEs Guide. https://www.kellton.com/kellton-tech-blog/ai-tech-stack-2026
  26. European Parliament. (2025). Generative AI and copyright: Training, creation, regulation. PE 774095.
  27. Azumo. (2026). Top 10 MLOps platforms for scalable AI in summer 2026. https://azumo.com/artificial-intelligence/ai-insights/mlops-platforms
  28. UCL Centre for Artificial Intelligence. (2025). AI and Machine Learning research themes. https://www.ucl.ac.uk/engineering/ai-and-machine-learning

  Interoperability Considerations

  AI data pipeline interoperability is critical for multi-vendor metaverse AI ecosystems where different platform components source, process, and consume data through pipelines built from heterogeneous technology stacks. Key interoperability dimensions in the AI + Data Mgmt domain include:

• Data schema interoperability: AI data pipelines consuming data from multiple metaverse platform components must handle heterogeneous data schemas. Standards must specify schema registration, versioning, and compatibility protocols enabling AI pipeline components to consume data from sources with different schema versions without manual reconfiguration.

• Feature store interoperability: Multiple AI services may share feature computations from a common feature store. OpenAPI-based feature store serving API specifications, and open feature metadata schemas (building on Feast’s feature metadata model), must enable cross-vendor feature store interoperability — allowing an AI model from vendor A to consume features computed and stored by a feature store from vendor B.

• Model registry interoperability: AI models trained in one environment (cloud training cluster) must be deployable in another (MEC edge node) using interoperable model serialisation formats and registry API schemas. MLflow’s model registry API and ONNX model serialisation format provide the primary open interoperability reference points; ETSI standards must specify which format profiles and API versions are required for compliant AI data management systems.

• Data lineage interoperability: OpenLineage provides an open standard for lineage event emission, enabling lineage data from pipeline components from different vendors to be collected in a unified lineage store. ETSI AI + Data Mgmt standards must reference OpenLineage or equivalent open lineage specifications, ensuring that lineage evidence collected for EU AI Act conformity assessment is platform-independent.

• Pipeline monitoring interoperability: AI pipeline observability data — metrics, logs, traces — must be emittable in interoperable formats (OpenTelemetry, Prometheus, structured JSON logging) to enable platform-independent monitoring and alerting in multi-vendor AI data pipeline deployments. OpenTelemetry’s semantic conventions for ML pipeline telemetry provide the primary reference for ETSI monitoring interoperability specifications.

• Federated learning protocol interoperability: Cross-platform federated learning — enabling multiple metaverse platform operators to jointly train models on distributed data — requires interoperable federated learning coordination protocols. The Flower (flwr) federated learning framework and the proposed IETF federated learning protocol drafts provide the primary reference implementations for cross-vendor federated learning protocol interoperability.

• ML-BOM interoperability: Machine Learning Bills of Materials produced by different AI development tools must be consumable by downstream compliance, procurement, and vulnerability management systems. The SPDX ML Profile extension to the industry-standard SPDX SBOM format, and the CycloneDX ML BOM extension, provide the primary interoperability reference specifications; ETSI AI + Data Mgmt standards must designate which profile formats are required for conformant AI data management systems.

• Continuous monitoring interoperability: Model performance monitoring data — inference latency histograms, prediction accuracy samples, drift detection statistics, data quality metrics — must be emittable in interoperable observability formats (OpenTelemetry ML semantic conventions, Prometheus ML metrics format) to enable unified observability platforms to monitor heterogeneous AI data pipeline deployments across multi-vendor metaverse infrastructure.

• AI audit trail interoperability: Audit records produced by AI data management systems for compliance purposes — data access logs, model update authorisation records, training data usage logs — must conform to interoperable audit log schemas enabling third-party auditors and regulatory authorities to consume audit evidence from systems from different vendors and platforms within a common evidence framework.

Provenance