What Is Agentic Analytics?
Agentic analytics is an approach where an AI agent plans and runs multi-step analysis end to end, replacing waiting-on-analysts with asking-and-acting. You ask the question. The agent chooses the method, runs the queries, checks the results, and writes the narrative. The unit of work stops being the ticket and becomes the answer.
The term matters because "AI analytics" has come to mean almost anything with a chat box. Agentic analytics names something specific.
From dashboards to chat boxes to agents
Analytics tooling has moved through four stages, and each one broke in a predictable place.
Static dashboards answered predicted questions fast and unpredicted questions never (the full argument is in why dashboards fail brand teams). Self-service BI handed the query builder to the business user, which mostly proved that drag-and-drop is still a query language. Chat boxes that turn a typed question into a single database query removed the query language and kept the deeper problem: one prompt becomes one query, unverified, with every judgment call (measure, period, product level) made silently.
Each stage moved the bottleneck without removing it. Dashboards moved it to the build queue. Self-service moved it to the business user's tool skills. Chat boxes moved it to trust: answers arrived instantly and nobody could tell which ones were right, so the analyst re-checked everything and the "time saved" evaporated. A tool that produces answers you must verify by hand has only automated the first draft of analysis, which was never the expensive part.
Agents are the fourth stage, and the break they fix is structural. In Anthropic's terms, earlier tools were workflows, systems that move along steps scripted in advance, while agents direct their own process and tool use to reach a goal [1]. OpenAI draws the same boundary: if the model does not control the workflow, it is not an agent [2]. The agent is the first stage where the system, not the user, is responsible for getting the analysis right end to end.
The ask-and-act loop
Stripped of jargon, agentic analytics is a loop with five verbs:
- Plan. Decompose "why is my share down at Kroger?" into the analyses that answer it.
- Query. Run each step against the data with the correct provider math.
- Check. Compare each result against what the data actually returned before building on it. Anthropic calls this the crucial discipline: gaining real feedback from the environment at each step rather than assuming the last action worked [1].
- Assemble. Turn verified results into charts, tables, and written narrative.
- Iterate. Notice the anomaly (share fell but velocity rose) and pull the thread.
In Sous the loop produces a workbook: a living document that holds the queries, the charts, and the narrative together. Because the workbook remembers how it was built, it can be saved as a template and re-run when the next period's data lands, so the answer does not rot.
Watch what the loop does to a real workday. Under the old model, the buyer question that arrives Monday morning gets scoped Monday afternoon, queued behind the trade deck, pulled Wednesday, sanity-checked Thursday, and answered Friday, by someone who spent half the intervening time on data preparation rather than thinking. Under agentic analytics the same question is answered in minutes, verified, and the human time moves to deciding whether the answer changes the Thursday meeting strategy, and what to ask next. Agentic analytics does not remove the analyst from the process. It removes the queue.
Why the timing is now
CPG is adopting AI faster than it is scaling it. In McKinsey's 2024 survey of 63 CPG leaders, 71% said they had adopted AI in at least one business function, up from 42% a year earlier, and 56% said they regularly use generative AI. The same research notes that no CPG player has truly scaled it yet [3]. McKinsey's estimate of the prize is not small: generative AI could add $400 billion to $660 billion a year to retail and CPG, equal to roughly 1.2 to 2.0% of annual revenues [4]. Both figures are consultant estimates, worth reading as direction rather than destiny, but the direction is unambiguous.
The gap between 71% adoption and zero at scale is exactly where agentic analytics lives. A chat box is easy to adopt and impossible to trust with real decisions, which is why so many CPG AI deployments stall at the pilot: the demo dazzles, the first wrong number in a buyer meeting ends the experiment. An agent that plans, verifies against provider rulebooks, and shows its work is the version of AI a brand can actually run its reporting on. Scaling AI in CPG analytics is a trust threshold to clear rather than a change-management problem to push through, and verification is what clears it.
Why "agentic" is more than a chat box
Three properties separate agentic analytics from an AI model with a text field, and they are the same three to test in any demo.
Autonomy: the system carries the analysis end to end, including the judgment calls, instead of translating one sentence into one query. Verification: every result is checked against the data source and the provider's rules before it ships (the argument for why this is non-negotiable on syndicated data is in "how accurate are AI agents on syndicated data?"). Memory: the system knows your brand, your categories, your retailers, and your language, and it gets better every period it runs.
A chat box has none of the three. An agent without verification has one of them, which on syndicated data is worse than none: it produces wrong numbers faster and with more confidence.
The distinction shows up in unglamorous places. An agentic system notices that the new period's file has a renamed column and stops to ask before loading it. It remembers that your team excludes the food-service line from velocity reads. It re-runs 14 workbooks at 6 a.m. when the drop lands and sends one notification. That work demos poorly next to a clever chat reply, and it is the actual product.
The definition, restated plainly: agentic analytics is analysis run by a system that owns the whole loop, plan through verified answer, and gets better at your business every period it runs. Start with what an AI agent is for the base definition, then judge every tool you evaluate against the loop.