Agentic analytics is the use of AI agents that carry out analytical work in multiple steps: given a question, the agent plans an approach, runs queries, looks at the results, notices problems, and revises — the way an analyst works, rather than the way an autocomplete works.
The term is young and marketing departments have been generous with it, so it’s worth pinning down what’s actually different. Plain text-to-SQL is one shot: question in, query out, hope for the best. An agent runs a loop. Asked “why did signups drop last week?”, it might check the trend, break it down by channel, find the anomaly concentrated in one source, query that source’s data for the same window, and come back with a narrowed hypothesis. Each step’s output shapes the next step — that feedback loop is the substance behind the word “agentic.” A dashboard filter is not agentic. A chatbot that runs one query is not agentic. Vendors will disagree.
What agents need to work
Under the hood, an agent is a language model with access to tools — typically the ability to inspect schemas and execute SQL, increasingly connected through standard interfaces like the Model Context Protocol. Three things determine whether the result is useful:
- Grounding. The agent only knows what your schema and metadata tell it. A semantic layer with pinned-down metrics matters even more here than for single-shot generation, because an agent makes many small judgment calls per task, and each ambiguity is a chance to drift.
- Feedback it can read. Clear errors, sensible result shapes, and documented tables let the agent notice when it’s wrong. Iteration without a signal is thrashing, not investigation.
- Evaluation. Multi-step work fails in multi-step ways, so serious deployments test agents against questions with known answers and track pass rates, cost per task, and latency over time.
The sober assessment
The compounding-error problem is real: a small misreading at step two quietly shapes every step after it, and a fluent five-step narrative is more persuasive than a single wrong number — which makes it more dangerous, not less. Agents are genuinely useful for exploration, triage, and the first pass at “why did this move?”; they are not a substitute for review on decisions that matter. The practical posture is the same as for a capable junior analyst: give them good tools and clean definitions, check the work, and expand trust as the track record earns it.
Metabase’s position here is deliberately incremental: Metabot lets people ask questions in natural language and generates the queries, grounded in the models and metrics your team has already curated. That groundwork is the unglamorous prerequisite for the whole category — an agent pointed at an undocumented warehouse mostly automates the production of plausible wrong answers.
Related terms
Put it to work
- LLM analytics — Overview
- Eval pass rate — Metric
- LLM cost per user — Metric
- LLM spend overview — Dashboard
