Picture the dashboard your team built last quarter. The one with the careful layout, the brand colours, the eleven tiles of revenue, spend, CAC, and conversion rate. Now answer honestly: when did anyone last open it, notice something odd, and act on it before the monthly review forced them to?
That gap is the whole problem. A dashboard is a beautifully organised answer to a question nobody is currently asking. It reports what happened and then waits. The work that actually matters, noticing the anomaly, finding the cause, and writing the story, still falls on a person who has fourteen other tabs open. Agentic analytics is the attempt to hand that work to software. Not the charting. The looking.
Origin: the dashboard era solved the wrong half of the job
Business intelligence spent two decades getting very good at display. Self-service tools let anyone build a report without filing a ticket. The result was more dashboards than anyone could read. Even now, BI adoption sits around 30 to 35 percent of employees at a typical organisation, a ceiling the industry has bumped against for years. The other two thirds were handed self-service analytics and quietly declined to use it.
The reasons are not mysterious. Reading a dashboard well is a skill, and it is also a chore. You have to know which numbers are normal, spot the one that is not, and then resist stopping there. A dashboard shows you that conversion fell. It does not tell you why. To get the why you open five more queries, segment by device, by channel, by region, and by the date the tracking change shipped, and somewhere in that drill-down you find the cause or you give up. Analysts know this tax. Survey work has long found data professionals spend less than half their time actually analysing data, with much of the rest lost to finding and preparing it.
So the dashboard era automated the easy half of analytics, the reporting, and left the hard half, the noticing and the explaining, exactly where it was. The chart got faster to build. The insight did not get any easier to reach. Agentic analytics starts from that observation and inverts the model.
Present: an agent that does the noticing
The cleaner way to describe agentic analytics is by contrast. Traditional BI is reactive and query driven. You go to the data with a question. Agentic analytics is proactive: the system goes through the data continuously and brings the question to you. Databricks frames it as autonomous agents that explore data, generate insights, and take context-aware actions with little human prompting, a step beyond augmented analytics, which only assists an analyst who is already working.
Strip the marketing language and a read-only analytics agent does four concrete things. It monitors metrics continuously instead of waiting to be opened. It detects an anomaly by learning what normal looks like for that metric, including its seasonality and its day-of-week rhythm, rather than firing on a fixed threshold. It runs the diagnostic drill-down automatically, the segment-by-segment dig an analyst would do by hand. And it delivers the result as a written narrative: what changed, by how much, and the most likely driver. A chart shows a dip. An agent writes the sentence a human would have written after an hour of work.
This is already shipping inside mainstream tools. Tableau Pulse, generally available since the February 2024 release, monitors metrics you follow and pushes plain-language insights to Slack and email when something moves. Its design has a detail worth noticing: the system uses deterministic statistical models to establish the facts first, then uses language generation only to phrase them, so the numbers are computed, not guessed. ThoughtSpot launched Spotter in November 2024 and calls it an autonomous agent for analytics, one that reasons through a question and holds context across follow-ups. Microsoft put Copilot into Power BI for natural-language questions and report-wide summaries, with Fabric data agents acting as domain-specific virtual analysts underneath. Google did the equivalent with Gemini in Looker. The pattern is now standard: every serious BI platform has a chat box, and the chat box is becoming an agent.
Why read-only is the right place to start
The smartest thing about the current wave is what these agents deliberately do not do. They observe and explain. They do not spend money or change a campaign. That restraint is a feature, and it changes the risk maths entirely.
Think about the failure mode. When an agent only reads, the worst thing it can produce is a wrong explanation. Someone reads it, checks it against what they know, and discards it. The cost is a few minutes and a little credibility. When an agent can act, the worst thing it can produce is a wrong action: a budget moved, a bid raised, a campaign paused, all at machine speed and possibly overnight. A wrong explanation is an annoyance. A wrong action is a loss. Starting with read-only agents means you get the upside of continuous monitoring while the downside stays survivable. Improvado draws the same line, separating observing agents that surface insights from acting agents that execute changes only with permission. Earn trust on the cheap failure mode before you allow the expensive one.
Three use cases shipping now
Three read-only patterns are already concrete and useful.
The first is the weekly performance narrative. Instead of an analyst spending Monday morning assembling last week's numbers into a deck, an agent pulls the data, computes the week-on-week moves, and writes the summary: revenue, the channels behind it, what beat plan, what missed, and the one chart that explains the week. The human edits and adds judgement rather than starting from a blank page.
The second is anomaly alerting with context. A bare alert (conversion rate down 9 percent) creates work; someone still has to investigate. A good agent does the first pass of that investigation and attaches it. A useful alert reads more like a finished thought: monthly sales fell because one product category underperformed in one region. Tableau Pulse phrases its insights in roughly that shape, naming the driver, not just the dip.
The third is the most underrated: attribution-shift analysis. Channels quietly steal credit from each other. When branded paid search captures conversions that organic would have won anyway, paid looks like a hero and organic looks weak, and the budget follows the wrong signal. The tell is a pattern humans rarely catch in time: pause one channel and watch a different channel's conversions rise by almost the same amount while the total barely moves, which means the first channel was cannibalising the second, not creating demand. An agent watching the full channel mix can flag that drift as it starts, instead of a quarter later when someone finally runs the holdout test. For the deeper logic of why a channel mix needs causal proof, not just reported numbers, the four types of analysis sets out what each layer of analytics can and cannot tell you.
The honest limits
An agentic analytics agent is genuinely useful and genuinely limited, and pretending otherwise helps nobody.
It can hallucinate an explanation. Language models are pattern matchers, and a confident, fluent, wrong paragraph reads exactly like a correct one. The risk is sharpest in the query step. The Spider 2.0 benchmark tested language models on realistic enterprise databases, the kind with over 1,000 columns and multiple SQL dialects, and a strong code agent built on o1-preview solved only 21.3 percent of the tasks, against 91.2 percent on the older, simpler Spider 1.0. The gap is the point. An agent that nails clean demo data can fall apart on a real warehouse. And the failure is quiet: a query that runs and returns a tidy number that happens to be wrong is far more dangerous than one that errors out, because a wrong answer that looks right gets believed and acted on. Designs that compute facts deterministically and let the model only narrate them, as Tableau Pulse does, are one sensible guard against this.
It inherits every problem in the data underneath it. An agent built on messy, inconsistent inputs explains the mess with total confidence. Vendors who track this find that grounding agents in real business definitions and quality metadata is what separates accurate output from invented output, and that without that context an agent will simply make something up that sounds authoritative.
And it cannot tell correlation from causation, because it was never built to. A language model is an associational engine; it finds what moves together, not what causes what. So an agent will reliably report that two metrics fell in the same week and may imply one caused the other when a third factor moved both, or when it is pure coincidence. The drill-down it produces is a strong hypothesis, not a proven cause. That distinction is the most expensive one in marketing analytics, and an agent does not remove the need for a human to apply it.
The honest summary: an agentic analytics agent is a fast, tireless first-pass analyst. It is not a final authority. It does the looking and the first draft of the explaining. A person still judges.
Future and impact: from explaining to recommending to acting
The trajectory is not hard to read. Today's agents explain. The next step is recommending: not just "conversion fell because paid social weakened", but "and here is the budget shift that would recover it". The step after that is acting: the agent makes the change itself. Gartner expects 40 percent of enterprise applications to embed task-specific AI agents by the end of 2026, up from under 5 percent in 2025, so analytics agents will soon be ambient rather than novel.
Each step up that ladder raises the stakes. An explaining agent that is wrong wastes a few minutes. An acting agent that is wrong moves real money before anyone reviews it. This is exactly why agentic media buying, covered in agentic MMM, is being built with the statistical model kept under human control while the agent orchestrates around it. The same caution belongs in analytics. The discipline is to advance one rung at a time: let an agent explain until its explanations are reliably right, let it recommend until its recommendations are reliably good, and only then let it act, with limits and a human on the consequential calls.
The market is already pricing the danger of skipping steps. Gartner predicts over 40 percent of agentic AI projects will be cancelled by the end of 2027, blaming escalating costs, unclear value, and weak risk controls. Governance is not paperwork here. It is the thing that decides whether an acting agent is an asset or a liability.
This is the work Perform Digital does with clients: starting where the failure mode is cheap, with read-only analytics agents that monitor, explain, and write the narrative, then adding recommendation and action only behind clear limits and human review. An agent that reads your dashboards so a person does not have to is a real productivity gain. An agent that acts on what it read before anyone checked is a different kind of system, and it needs to earn that trust first.
The dashboard era gave every team more data than it could read. Agentic analytics is the correction: software that does the reading, the digging, and the first draft of the story, and leaves the human to do the judging. Used with that division of labour clear, it is one of the most genuinely useful things AI has brought to analytics. Used as an oracle, it is just a faster way to be confidently wrong.
Council summary
The post argues that the dashboard era automated the easy half of analytics, the reporting, and left the hard half, the noticing and the explaining, untouched, and that agentic analytics is the correction: software that monitors metrics, runs the diagnostic drill-down, and writes the narrative a human would have written by hand. Its sharpest point is that read-only is the right place to start, because a wrong explanation costs a few minutes and a little credibility while a wrong action moves real money at machine speed. It is honest about the limits, that an agent can hallucinate a fluent wrong answer, inherits every flaw in the data beneath it, and cannot tell correlation from causation, so it is a fast first-pass analyst and never a final authority. The reader's takeaway is a discipline: adopt read-only analytics agents now for the genuine productivity gain, advance from explaining to recommending to acting one rung at a time, and keep a human on every consequential call.
Comments