The AI landscape doesn't move in one direction — it lurches. Some techniques leap from experiment to table stakes in a single quarter; others stall against regulatory walls, technical ceilings, or organisational inertia that no amount of hype can dislodge. Knowing which is which is the hard part. The State of Play cuts through the noise with a rigorously maintained index of AI techniques across every major business domain — classified by maturity, evidenced by real-world adoption, and updated daily so you always know where you stand relative to the field. Stop guessing. Start knowing.
A daily newsletter distilling the past two weeks of movement in a domain or two — delivered to your inbox while the index updates in the background.
Each dot marks the weighted maturity of practices within a domain — hover for a brief summary, click for more detail
Comprehensive management of AI model inventories including documentation, model cards, versioning, and lifecycle tracking from development to retirement. Includes automated model card generation and deprecation workflows; distinct from model evaluation which assesses performance rather than managing metadata.
Model inventory, documentation, and lifecycle management is a leading-edge practice defined by an acute and widening gap between vendor tooling maturity and organizational adoption. The discipline covers maintaining authoritative registries of models in production, documenting capabilities and lineage via model cards, tracking versions, and governing the full lifecycle from development through retirement. By April 2026, every major cloud provider ships production-grade registries (AWS SageMaker, Databricks/MLflow, Microsoft Azure Databricks, Google Vertex AI), with regulatory compliance requirements now explicit: the April 2026 revised federal Model Risk Management guidance from the Federal Reserve, FDIC, and OCC establishes model inventory and governance as foundational requirements for all US banks >$30B assets. The tooling question is fully settled. The governance question has become critical. Research and audit findings in 2026 reveal a severe transparency and documentation crisis: Stanford's 2026 AI Index documents Foundation Model Transparency Index collapse (58→40/100 year-over-year), with 80 of 95 foundation models released in 2025 lacking training code or parameter disclosure. Meanwhile, regulatory examinations show most US banks have only partial inventory coverage (43% cannot update live models; 38% struggle sustaining governance across growing inventories), and shadow AI—unregistered business-unit tools, vendor-embedded models, proof-of-concept systems—has emerged as the dominant inventory gap. The binding constraint is now organizational implementation effectiveness and post-deployment lifecycle continuity, not technical capability. This is a practice where leading practitioners (Uber, Cisco, major financial institutions) operate sophisticated registries with automated governance, while the broader field faces acute governance-adoption velocity mismatch driving proliferation of shadow AI and organizational governance failures.
The vendor ecosystem continues maturing with multi-platform governance convergence. AWS SageMaker ML Governance suite (April 2026) provides automated model card population, DataZone integration, role-based access control, and Model Dashboard monitoring for customers including Cisco, Perplexity, and Salesforce. Databricks MLflow on Databricks (April 2026) integrates centralized model registry with Unity Catalog, deployment jobs with approval gates, and auditable activity logs—covering evaluation, approval, and deployment lifecycle stages. Microsoft Azure AI Foundry formalized lifecycle retirement governance with explicit phases and concrete deprecation timelines. Uber published production case study (April 2026) of centralized Model Catalog integrated into Michelangelo ML platform with automated metadata population from system signals and feature attribution (SHAP, TreeSHAP) directly linked to Model Cards for cross-functional governance. MLflow maintains ecosystem prominence with 30 million monthly downloads across 1000+ organizations. AWS demonstrated end-to-end lineage patterns (April 2026) combining DVC (data versioning) with SageMaker MLflow Apps to address audit governance for regulated industries.
Regulatory drivers have become binding. The April 17, 2026 revised Model Risk Management guidance from Federal Reserve, FDIC, and OCC replaces 2011-era SR 11-7 with principles-based framework requiring model inventory, validation, monitoring, and vendor management scaled to organizational risk profile—applying to all US banks with >$30B assets or significant model risk exposure. However, regulatory examinations reveal critical adoption failures: most US banks have only partial coverage across four governance components (inventory, approval, change management, retirement), with self-learning models and agent orchestration layers systematically missing. Shadow AI—unregistered business-unit tools, vendor-embedded models, proof-of-concept systems—has emerged as the dominant inventory gap in federal examinations. Consultancies now explicitly map model cards to EU AI Act compliance requirements (Articles 12-14), identifying governance elements like approval workflows, validity periods, and re-audit intervals as necessary for regulatory alignment.
Adoption barriers remain acute and structural. Post-deployment governance represents a critical lifecycle gap: documentation decays after launch, classification drift is missed when use cases expand, ownership dissolves when project teams disband, and vendor model updates often bypass reassessment workflows. Hawk/Chartis survey of 125 financial institution leaders (April 2026) found 43% cannot update live models, 38% struggle to sustain governance across growing inventories, and 70% report model performance degradation left unaddressed. Stanford 2026 AI Index documents critical transparency crisis: Foundation Model Transparency Index collapsed from 58 to 40/100 year-over-year, with 80 of 95 foundation models released in 2025 lacking training code or parameter disclosure—indicating that documentation comprehensiveness is deteriorating despite governance framework adoption at vendor level. Multi-sourced 2026 surveys reveal governance-adoption velocity mismatch: 23% of companies moderately using agents projected to reach 74% within two years, but only 21% have mature agent governance models; 96% using agents report sprawl, yet only 12% implement centralized platforms. Independent reviews document steep learning curves and opaque pricing across major platforms. Cross-platform integration reveals friction: MLflow Model Registry in Microsoft Fabric exposes API limitations around aliases and metrics. The constraint has shifted from technical capability to post-deployment lifecycle continuity, documentation comprehensiveness at scale, and ability to manage rapidly proliferating unregistered AI systems.
— Uber's production-scale deployment of centralized Model Catalog with standardized Model Cards, auto-populated metadata from Michelangelo ML platform, integrated feature attribution—demonstrates enterprise implementation of model inventory and documentation at scale.
— Databricks analysis of April 2026 revised MRM guidance shows shift to risk-based, lifecycle-oriented governance with continuous monitoring; architecture implications and platform integration strategies for regulatory compliance.
— Databricks MLflow 3 with Unity Catalog integration providing centralized model registry, lifecycle deployment jobs with approval gates, and auditable activity logs—production-ready infrastructure for model inventory and lifecycle governance.
— Multi-sourced 2026 surveys show 23% companies moderately using agents growing to 74% in 2 years but only 21% have mature governance; 96% using agents report sprawl; only 12% implement centralized management—demonstrates urgent adoption pressure for inventory infrastructure.
— AWS technical pattern combining DVC (data versioning) with SageMaker MLflow Apps to track complete model lineage from data→training→deployment, addressing audit and governance requirements for regulated industries (healthcare, finance).
— Stanford CHAI 2026 AI Index documents Foundation Model Transparency Index collapse (58→40/100), 80 of 95 models lack training code disclosure—signals critical transparency/documentation gap despite governance framework adoption (ISO/IEC 42001 36%, NIST RMF 33%).
— Joint April 2026 revised Model Risk Management guidance from FDIC, OCC, Federal Reserve replaces 2011 guidance with principles-based framework requiring model inventory, validation, monitoring, and vendor governance—major regulatory driver for adoption.
— Critical analysis of model card maturity gap: published cards (designed 2019 for fairness) fail to document operational specs (latency, context degradation, failure rates); Stanford CRFM 2026 Index confirms declining transparency (24/100–32/100)—negative signal on documentation practices.
2020: Major vendors (AWS, Databricks, SAS) released model registry and lifecycle management products, signaling early ecosystem maturity; open-source implementations showed technical maturity gaps with documented integration failures and UI limitations.
2021: Academic standardization efforts emerged (HuggingFace, GEM, Model Card Toolkit) indicating consensus on documentation templates; vendor tooling continued to mature with documentation updates. However, industry survey data revealed critical adoption barriers: financial services organizations with 270+ models in production rated inventory processes as only 25% effective. MLflow and open-source tools struggled with platform compatibility (Windows performance issues) and integration reliability, limiting adoption beyond cloud-native environments.
2022-H1: Vendor ecosystem consolidated with AWS, Microsoft, and Databricks all shipping production-grade model registry and lifecycle management features; Vanguard deployed SageMaker Model Registry at Fortune 500 scale with 100% automated deployment. However, adoption barriers intensified on the organizational side: HuggingFace documentation study showed only 40% of models have any documentation despite years of tooling availability; industry surveys found 85-90% of ML models never reach production, with lifecycle management delays and organizational bottlenecks cited as primary causes. Tooling had matured; adoption had not.
2022-H2: Vendor expansion accelerated with Azure ML Registries entering public preview and Google Vertex AI Model Registry growing to named production deployments (ZOZO). AWS reported tens of thousands of SageMaker customers managing millions of models and generating hundreds of billions of predictions. Documentation standards continued to advance (NVIDIA Model Card++). However, organizational adoption barriers persisted: talent shortages (29% of decision-makers cited lack of talent as key challenge) and organizational complexity remained the binding constraints, not tooling maturity.
2023-H1: AWS launched SageMaker Collections for hierarchical model organization; Microsoft and Google expanded model registry capabilities with tutorials and production case studies. However, documentation quality deteriorated: HuggingFace analysis found 80% of models lack sufficient docs (vs. 40% in 2022), 88% of model cards inflated performance claims, 96% omitted bias/limitations. Only 1 in 10 ML models operationalized; 64% of organizations require 1+ months for deployment. Research pivoted toward automated model card generation to address documentation labor bottleneck. Vendor feature expansion continued while organizational adoption stagnated.
2023-H2: Vendor ecosystem continued expanding with AWS SageMaker deployment approval workflows (November) and Databricks MLflow lifecycle examples. Research on automated model card generation (arXiv 2309.12616) advanced documentation automation with 500-example QA datasets, though findings revealed LMs struggling to understand documentation requirements. RMA survey showed two-thirds of 53 financial institutions using lifecycle management IT applications, signaling sustained adoption in regulated sectors despite tool complexity. However, Kubeflow user survey (July) found model registry remained a top gap (44%) across open-source platforms, and Azure ML integration issues surfaced real-world deployment complexities. Vendor tooling matured incrementally while organizational barriers—documentation quality, tool interoperability, and platform compatibility—persisted as binding constraints on broader adoption.
2024-Q1: Vendors continued operationalizing governance features: AWS SageMaker model registry automated approval and promotion workflows (with Merck pharma as production case study), and Azure Databricks launched wind farm forecasting examples for lifecycle management. Research community advanced automated model card generation with LLM-based approaches and large datasets (NAACL-HLT 2024, CardBench with 4.8k model cards). Open-source ecosystem expanded with Kubeflow Model Registry entering alpha. However, platform reliability remained a blocker: MLflow integration failures with Azure ML and model registration bugs surfaced in production deployments, indicating that despite vendor maturity, organizations still face real-world obstacles to seamless inventory management. Documentation incompleteness persisted as a structural challenge requiring automation.
2024-Q2: New vendor expansion accelerated: Snowflake announced GA of its Model Registry (May), and Valohai released centralized registry features for its MLOps platform. Research on automated model card generation published in NAACL 2024 (CardGen paper) demonstrated LLM-based generation of model and data cards from cardBench dataset of 4.8k examples, addressing documentation labor bottleneck. Regulatory perspectives strengthened: OSFI (Canadian financial regulator) research paper advocated adoption of model ownership, documentation, and challenge principles from financial model risk management to AI systems. However, integration maturity gaps persisted: MLflow registry integration failures in Ultralytics YOLO and Vertex AI SDK deployment issues for BigQuery ML models revealed ongoing real-world adoption barriers despite vendor tooling expansion.
2024-Q3: Vendor feature consolidation accelerated: AWS advanced SageMaker Model Registry with automated approval workflows incorporating governance checks (quality, bias, feature importance) for multi-account organizations; Azure ML confirmed GA MLflow integration for workspace-level lifecycle management; Valohai released Model Hub with versioning, lineage, and automated approval. Critical ecosystem shift emerged: Databricks deprecated its workspace model registry in favor of Unity Catalog, signaling major platform reorganization toward centralized cross-workspace governance. However, real-world deployment barriers persisted: practitioner analysis documented fundamental misalignment between idealized lifecycle models and organizational chaos in implementation, with hidden dependencies and provisioning failures in Azure ML ecosystem undermining platform reliability. Documentation automation research advanced but organizational adoption gaps remained structural.
2024-Q4: Platform consolidation continued: AWS released GA of cross-account model sharing via SageMaker Model Registry with AWS Resource Access Manager (November), enabling enterprise governance at scale. SAS and open-source projects (model-card-generator) advanced automated model card generation to reduce documentation labor. Databricks workspace model registry formally moved to legacy status with migration to Unity Catalog, completing the platform reorganization toward centralized governance. However, ecosystem voices remained critical: vendors acknowledged documentation complexity (SAS plea for simpler cards) and real-world deployment barriers persisted despite expanded feature sets. Tooling reached clear maturity—cross-account governance, automated documentation, deep audit trails—but organizational adoption and documentation quality remained below leading-edge expectations, with talent, process complexity, and interoperability gaps persisting as binding constraints.
2025-Q1: Vendor tooling maturation continued: AWS unified SageMaker Model Cards directly with Model Registry to streamline governance workflows; empirical research showed MLflow adoption driving significant improvements in development cycle times, reproducibility, and deployment efficiency across organizations. Sectoral adoption accelerated in healthcare with Coalition for Health AI (CHAI) launching a model card registry with Providence, Cleveland Clinic, and Kaiser Permanente participation, standardizing documentation for healthcare AI procurement. However, real-world deployment barriers persisted acutely: Kubeflow Model Registry UI defects and Azure Databricks/MLflow authorization failures in production deployments revealed ongoing integration maturity gaps; critical assessments documented persistent governance and lifecycle management deficiencies in custom AI solutions despite vendor tooling maturity. Open-source and commercial ecosystems continued divergence: while Tier 1 vendors achieved technical maturity and sectoral adoption signals, organizational barriers—integration failures, documentation quality, and production reliability gaps—remained binding constraints on broader industry adoption.
2025-Q2: Documentation crisis intensified alongside continued vendor tooling maturity. IEEE Requirements Engineering Conference research (June 2025) analyzed 26 ethics guidelines and 10 model cards, finding developers overwhelmingly emphasize capabilities and reliability while systematically overlooking fairness, explainability, and user autonomy—negative signal on model card comprehensiveness despite vendor automation efforts. Major vendor transparency failure emerged: Google released Gemini 2.5 Pro (March 2025) without safety report or model card, violating public commitments to US government and international AI safety summits; similar gaps reported at OpenAI and Meta. Vendor tooling continued maturity trajectory but deployment-first practices demonstrated that regulatory and transparency commitments remained subordinate to rapid release cycles. The practice remained technically advanced but organizationally fractured: Tier 1 vendors provided mature, feature-rich registries and governance automation; yet deployment practices and documentation completeness continued deteriorating as model volume and urgency accelerated.
2025-Q3: Vendor tooling expanded unified lifecycle capabilities while organizational adoption barriers persisted. AWS SageMaker HyperPod launched model deployment (July 2025) enabling unified training-to-inference on same infrastructure with named customers (Perplexity, Hippocratic, Salesforce, Articul8) demonstrating real-world adoption across foundation model development. Microsoft formalized lifecycle retirement governance in Azure AI Foundry (September 2025) with explicit phases and concrete timelines for model deprecation and replacement. Regulatory compliance requirements emerged: EU AI Act mapping accelerated with consultancies (2B Advice September 2025) explicitly linking model cards to compliance, identifying governance elements (approvals, validity periods, re-audit intervals) needed for regulatory alignment. However, open-source tooling quality regressed: MLflow 3.0 introduced UI regressions (Source run link disappearance in Model Registry, July 2025) signaling quality control gaps despite continued development. Documentation incompleteness and organizational implementation barriers remained binding constraints despite leading-edge vendor feature maturity. The practice embodied persistent technical maturity alongside organizational stagnation: sophisticated registries coexisting with systematic documentation gaps, transparency failures, and deployment velocity outpacing governance infrastructure.
2025-Q4: Vendor ecosystem consolidated continued investment in model lifecycle tooling while documentation quality research revealed persistent comprehensiveness gaps. Microsoft Azure AI Foundry and Azure ML continued MLflow integration (November 2025), confirming platform commitment to lifecycle management though with explicit limitations documented (no model renaming, no organizational registries, no cross-workspace operations). Academic research emerged with mixed signals: Patra Model Card framework (November 2025) advanced documentation beyond static reports with dynamic, runtime-aware systems for edge AI environments, yet peer-reviewed analysis of 90 model cards (WEBIST 2025) found pervasive structural variance, missing ethical reporting, and inconsistent transparency—documenting that documentation practice quality remained far below vendor tooling capability. Industry analysis (December 2025) reported 87% of data science projects never reach production with poor data lifecycle management as primary culprit, underscoring that organizational adoption barriers and lifecycle practices themselves—not vendor tooling—remained the binding constraint. By year-end 2025, the practice had reached a plateau: registries achieved leading-edge technical maturity with formalized retirement governance and expanded cloud platform integration, yet documentation completeness, organizational implementation effectiveness, and deployment velocity management remained unresolved structural challenges limiting broader adoption despite years of vendor investment.
2026-Jan: Vendor tooling advancement continued with AWS S3-based SageMaker AI Project templates (January 2026) enabling version-controlled, decentralized project management, and research reframed model cards using system safety methodologies at ICSE 2026. Empirical MLOps tool evaluation (arxiv January 2026) independently assessed Metaflow, Airflow, and Kubeflow alongside MLflow, measuring installation complexity and ML scenario implementation barriers. Critical research emerged: ADAS framework (January 2026) explicitly critiqued model cards as providing only descriptive information without binding deployment decisions, calling for machine-readable authorization standards. Real-world Cisco deployment showcased SageMaker Model Registry efficiency gains with programmatic lifecycle management. Market adoption signals remained positive: Grand View Research projected MLOps market at $16.6B by 2030 (+40.5% CAGR), with MLflow 3.x governance evolution and enterprise adoption of comprehensive lifecycle practices. However, the structural tensions identified in 2025 intensified: tooling maturity continued advancing at vendor level, yet the documentation crisis and organizational adoption barriers remained unresolved despite three years of research, automation attempts, and regulatory pressure. Model inventory registration tooling had become a commodity feature across Tier 1 vendors—the binding constraint shifted toward integrated governance, documentation completeness, and organizational implementation effectiveness.
2026-Feb: AWS released continued enhancements to SageMaker in 2025 review post, emphasizing improved observability with granular metrics and serverless MLflow integration (February 2026). MLflow maintained ecosystem prominence with 30 million monthly downloads across 1000+ organizations. However, critical adoption barriers persisted: independent platform review (TrueFoundry February 2026) documented opaque pricing, steep learning curves, and vendor lock-in penalties for multi-cloud strategies in SageMaker ecosystem. Platform integration challenges emerged: Microsoft Fabric integration of MLflow Model Registry revealed API limitations (alias support, metrics accessibility), indicating maturity gaps in cross-platform lifecycle management. By February 2026, the practice maintained leading-edge technical capability in vendor tooling but faced unresolved organizational adoption barriers, platform integration friction, and pricing/complexity burdens limiting deployment velocity among practitioners.
2026-Mar: Regulatory drivers intensified with OCC examinations explicitly requiring model inventory compliance under SR 11-7, extending financial model risk management to AI systems. Examinations revealed most US banks have only partial governance coverage: 43% cannot update live models, 38% struggle sustaining governance across growing inventories, and self-learning models/agent orchestration layers often missing entirely. Post-deployment governance emerged as a critical lifecycle gap: documentation decay, classification drift misses, ownership dissolution when teams disband, and vendor model updates bypassing reassessment workflows. ServerWorks deployed production model lifecycle management on SageMaker MLflow, demonstrating end-to-end operationalization. Model lineage formalized as a foundational EU AI Act compliance requirement, with datasets, code, hyperparameters, training conditions, and deployment targets forming an auditable provenance chain.
2026-Apr: Regulatory drivers reached critical mass with revised federal Model Risk Management guidance (April 17, Federal Reserve/FDIC/OCC) replacing 2011 guidance, establishing principles-based governance requiring model inventory, validation, monitoring, and vendor management across all US banks. AWS unified ML Governance suite matured with Model Cards autopopulation, DataZone integration, and integrated Model Dashboard monitoring. Databricks MLflow on Databricks formalized lifecycle management with Unity Catalog integration and approval-gated deployment jobs. Uber published production case study of Model Catalog (centralized inventory with auto-populated Model Cards and feature attribution integrated into Michelangelo ML platform), demonstrating enterprise-scale operationalization. Stanford 2026 AI Index documented critical transparency gap: Foundation Model Transparency Index dropped 58→40/100 year-over-year, with 80 of 95 2025 model releases lacking training code disclosure. Multi-sourced 2026 surveys revealed governance-adoption paradox: 23% companies moderately using agents projected to reach 74% in two years, but only 21% have mature governance models; 96% of organizations using agents report sprawl, yet only 12% implement centralized control platforms. Hawk/Chartis survey of 125 financial leaders found 70% report model performance degradation unaddressed, with shadow AI (business-unit tools, vendor-embedded models, PoCs) as primary inventory gap in regulatory examinations. By April 2026, vendor tooling maturity is unambiguous and multi-sourced (AWS, Databricks, Microsoft, Uber), but organizational adoption barriers remain acute: documentation comprehensiveness declining despite governance framework adoption, post-deployment lifecycle gaps (ownership dissolution, classification drift), and governance-velocity mismatch driving shadow AI proliferation.