§ Learn
A guided introduction to the stack: what agentic analytics is, how the five layers compose, which open standards matter at each layer, and where current systems break. Written for data practitioners.
§ Definition
Agentic analytics refers to a class of data analysis systems in which an autonomous agent receives a business question in natural language and executes a multi-step investigative workflow: planning queries, evaluating intermediate results, revising its approach based on what it finds, and synthesizing findings into a structured answer. The human provides the objective; the system determines the method.
The qualifier "agentic" is drawn from the broader AI agent paradigm, where a system exhibits goal-directed autonomy. In this context, that means the system decides what data to examine next based on prior observations. If an initial query surfaces an anomaly, the agent drills into it. If data quality issues are detected, it applies validation checks before proceeding. Answering the question requires joining across heterogeneous sources? The agent determines the join strategy independently.
This is a qualitative shift from prior paradigms. Traditional BI requires pre-built views. Self-service BI requires user-directed exploration. Even NL-to-SQL systems operate in a single-turn request-response loop. Agentic analytics introduces a feedback loop: the system observes, reasons, acts, and re-plans, mirroring the iterative workflow of a skilled analyst.
§ Taxonomy
The term has been co-opted across vendor marketing. A rigorous taxonomy distinguishes surface-level AI integration from architecturally agentic behavior.
Text-to-SQL that generates one query from one question. A chatbot wrapper on existing dashboards. Auto-generated chart titles. NL search over a BI catalog. Pattern-matched "insights." These systems add value. They are not agentic by the operational definition.
Multi-step investigation with branching logic. The agent decides which tables to query and why. Intermediate results change the investigation plan. Autonomous data quality checks mid-analysis. Root cause analysis that follows causal chains without additional user prompting.
The diagnostic test: does the system make autonomous decisions about what to investigate next, based on intermediate results, without additional human prompting? If every step requires a new user instruction, the system is assisted, not agentic. We maintain an open benchmark to make this distinction measurable rather than subjective.
§ The stack
An agentic analytics system is not a monolithic product. It is a composition of five layers with clearly delineated responsibilities. Each has its own open standards and failure modes. Each has its own subpage below.
The warehouse, lakehouse, and file formats the agent ultimately queries.
Standards: Iceberg · Parquet · Delta · Arrow · ANSI SQL Open →The contract between business meaning and database schema. The highest-leverage dependency in the stack.
Standards: MetricFlow · Cube · Malloy · SQL:2023 Open →Access controls, data quality validation, lineage, and audit. Prevents agents from amplifying errors at machine speed.
Standards: Polaris · Unity Catalog · OpenLineage · ODCS Open →Where the agent plans, calls tools, evaluates intermediate results, and decides what to do next.
Standards: MCP · OpenTelemetry GenAI · A2A Open →How agent results surface to humans and downstream systems: reports, alerts, dashboards, actions.
Standards: OpenAPI · AsyncAPI · CloudEvents · nbformat Open →§ State of the art
Based on publicly available information, vendor demonstrations, and early adopter reports, current capabilities fall into three tiers of readiness.
Production-viable today. Bounded investigative workflows over well-modeled data with a robust semantic layer. When the warehouse is clean, metrics are formally defined, and the question domain is constrained (revenue variance analysis, churn decomposition, campaign attribution), current systems deliver meaningful time savings compared to manual analyst workflows.
Inconsistent reliability. Multi-source investigations requiring cross-domain joins over data that is not co-modeled. Complex business logic that exists only in analyst intuition. Scenarios requiring institutional context that has not been formalized. Agents frequently produce confidently stated but materially incorrect conclusions. The failure mode is particularly dangerous because outputs appear well-reasoned.
Not production-ready. Fully autonomous closed-loop systems that detect anomalies, investigate root causes, and execute corrective actions without human review. The precision and reliability thresholds required for unsupervised operation are not met by current platforms. Vendor demonstrations here consistently rely on curated scenarios with pre-validated data, a pattern that does not generalize to production variability.
The central finding: the prerequisite for agentic analytics is not the agent itself but the data infrastructure beneath it. Organizations with mature semantic layers, governed warehouses, and formalized metric definitions are best positioned to extract value. Organizations without these foundations will find that agents accelerate error production, not insight generation.
§ Next
The maturity assessment positions you on the five-layer model and identifies the foundation work needed before agents deliver reliable value. Five minutes.