Agentic Analytics: What It Is & How To Do It In 2026
Agentic analytics is data analysis done by an AI agent that works in a loop: it plans the steps, picks its tools (SQL, Python, search), checks what comes back and revises before it answers. A chat box that turns one question into one query is conversational analytics. An analytics agent can take "why did conversion drop?", split the change into its drivers, test the segments, mark what the evidence cannot confirm, and run the same check again next Monday without being asked, inside your permissions.
What is agentic analytics?
Gartner describes the category as the evolution of augmented analytics through applying AI agents to data analysis. In augmented analytics, AI helps an analyst who stays in charge of every step. In agentic analytics, the agent carries the analysis from question to finding, and a person reviews what it produced.
Each step depends on the last result: a dip concentrated in one region sends the agent into that region, a failed query gets fixed rather than reported, and a finding the data cannot support gets labeled as a hypothesis.
Agentic is not a synonym for better. A loop costs time: the d0 answer Vercel showed took 45 steps and 4 minutes 20 seconds, and an OpenAI screenshot of its agent reads "Worked for 6m 7s". Last quarter's revenue should be a lookup in seconds, so a good system decides how much loop a question deserves.
Seven criteria that separate an agent from a chat box
Ask the vendor to show each one live on your data rather than a sample dataset.
Criterion | What it looks like | How to test it in a demo |
1. Plans multi-step work | Breaks a broad question into hypotheses, cuts and checks, and shows the plan while it runs | Ask "why did [metric] drop last month?" Do you get one chart, or an investigation with steps? |
2. Chooses its tools | Decides when to write SQL, when to run Python for statistics or a forecast, when to read docs or search the web, and when to hand a branch to a sub-agent | Ask a question that needs a significance test or a forecast on top of a query |
3. Checks its own results | Fixes failed queries, re-checks a surprising number with another cut, and refuses to call a correlation a cause | Ask for a number you already know, then ask how it verified it |
4. Asks clarifying questions | When a question could go materially different ways, asks one short question before starting instead of guessing | Ask about "revenue" when your company has a gross and a net definition |
5. Shows its work | SQL, sources, lineage and the strength of evidence behind each claim, so a reviewer can check it | Click any number. Can you see the query and the tables it came from? |
6. Acts on schedules and triggers | Runs without a person starting it: scheduled analyses, monitors with conditions, alerts | Ask it to "tell me on Mondays, but only if signups fell more than 10%" |
7. Operates under governance | Applies your warehouse permissions and row-level security, logs who asked, and changes shared definitions only through review | Log in as a restricted user and ask for data they should not see. Then tell it to "remember" a new metric definition and see who has to approve it |
The combination matters more than any single row. A tool that passes 1 to 5 but not 6 and 7 is a good analyst you have to remember to ask. A tool that runs on triggers (6) without checking itself (3) or showing its work (5) sends confident numbers to people who did not ask for them and cannot check them.
Is agentic analytics a new category or just marketing?
Reddit asks it bluntly: is this BI plus a semantic layer plus an LLM with a new name? My answer: the word is mostly marketing, and the thing underneath is real.
The marketing part. "Agentic" now gets attached to chat boxes that run one query per question. Gartner's must-have capabilities for the market (data source connectivity, data preparation, agent workflow orchestration, automated insights and natural language query) are mostly what good BI tools already offered; agent workflow orchestration is the new item. Treat the word as a claim to test.
The real part. A one-shot system lives or dies on its first query. A loop can recover. LinkedIn added a stage that checks generated SQL and sends errors to a researcher agent, and invalid tables and columns fell from 23% to 1%. Vercel rewrote d0 as an agent that explores its semantic layer with file reads and bash, and its success rate went from 80% to 100%.
The part nobody sells. The loop does not fix missing context. LinkedIn's agent scored 9% correct or near-correct with table schemas alone and 53% with its knowledge graph. Pinterest's table search found the right table 40% of the time without documentation and 90% with it. Agentic analytics on an undocumented warehouse is a faster way to be wrong.
So if a product does none of the seven things above, it is conversational analytics, which is fine for plenty of questions. If it does most of them, the new name is earned.
How Gartner frames the market in 2026
- Agentic analytics on Peer Insights. Gartner Peer Insights has an agentic analytics market page where users review products. Most vendors there have only a handful of reviews, so read the scores with that in mind. Dot is listed in that market, with four reviews as of October 2026.
- Decision intelligence platforms. Gartner published its first Magic Quadrant for Decision Intelligence Platforms on 26 January 2026; Aera Technology, FICO, IBM, Quantexa and SAS are among the vendors that announced they were named Leaders (see Quantexa's release and Aera's). Gartner frames that category as decision modeling, analytics and AI combined to augment and automate decisions: agentic analytics does the analysis, decision intelligence turns it into repeatable decisions. Our decision intelligence software guide covers that side.
What nine in-house data agents teach
The best evidence comes from companies that built their own data agent and published what happened. One lesson per company, with the criterion it teaches and our teardown of each (all fourteen published builds are in How 14 Companies Built AI Data Analysts):
Company and agent | The lesson | Criterion it teaches |
Shrink the agent's decision surface: fewer overlapping tools, more context prepared in advance | Chooses its tools | |
Decompose: domain, then tables, then columns, then SQL | Plans multi-step work | |
Golden queries before every deployment; human validation for executive numbers | Checks its own results | |
A fix-the-query stage beats a smarter first draft | Checks its own results | |
Two general tools over a complete semantic layer beat eleven specialized ones | Chooses its tools | |
Context as code: reviewed, tested and versioned in dbt | Operates under governance | |
Find the right tables first, and let the user confirm them | Asks clarifying questions | |
Autonomy is a portfolio: fixed workflows for audited reporting, agents for exploration | Acts on schedules and triggers | |
Discovery, not SQL, was the failure mode; fix it upstream | Operates under governance |
OpenAI. Its data platform serves 3,500 internal users across 600 PB and 70,000 datasets. Under the write-up's first lesson, "Less is More", the team consolidated overlapping tools, and six layers of context mean the agent makes fewer choices at runtime.
Uber QueryGPT. One large prompt lost accuracy as tables were added; four narrow agents (intent, tables users can review, column pruning, SQL generation) fixed it. In a limited release it averaged 300 daily active users, 78% said it cut their query-writing time, and Uber estimates about 140,000 hours saved a month.
Uber Finch. A supervisor agent routes each finance question in Slack to a SQL writer or a document reader, and an OpenSearch index maps aliases such as "US&C" to real column values. Executive-level answers get a "Request validation" button that sends them to a subject-matter expert.
LinkedIn SQL Bot. Only 53% of answers scored correct or near-correct on 133 internal questions, yet 95% of users rated them passing or better, because the tool helps them find tables and iterate. Its most-used feature, at 80% of sessions, is a "Fix with AI" button on failed queries.
Vercel d0. The rewrite reads a YAML semantic layer with ls, grep and cat, and runs 3.5 times faster with 37% fewer tokens. It works because every entity in that layer is documented, with descriptions, sample values, joins and example questions.
Ramp Research. The agent in Slack drew 1,476 questions in four weeks, against 66 in the old help channel, and answered more than 1,800 questions for 300 users. Its tools are plain; the context lives in dbt, with glossaries as seeds, domain documents as doc blocks and tests for the agent's behavior.
Pinterest. With hundreds of thousands of tables, the second version retrieves candidate tables, has an LLM re-rank them and asks the user to confirm before writing SQL. First-shot acceptance rose from about 20% to above 40%.
DoorDash. It maps four stages (deterministic workflows, single agents, deep agents, agent swarms), keeps deterministic workflows for reporting where auditability matters, and gives agents work such as "Investigate the drop in conversions in the Midwest last week". Generated SQL goes through EXPLAIN-based validation.
Stripe Hubert. About 900 Stripe employees use it each week. Stripe's verdict: "The biggest failure mode has been data discovery." The fix was upstream: deprecate low-quality tables, write human-owned documentation for canonical datasets, and give Hubert the asker's org as context.
Across all nine, the loop helps, but context, discovery and verification decide the result. These companies staffed platform teams for all three; a product has to deliver them without one.
How accurate is an analytics agent?
Our best public evidence is Adyen's DABstep benchmark: multi-step questions on real payment data that usually need three or four tables and business rules the warehouse never wrote down. From our full write-up:
System | Accuracy | Scored on | Time per question |
Dot, configured (our run, Feb 2026) | 90% | A 30-task slice, easy and hard, scored with DABstep's own scorer and reproduced on the public Hugging Face leaderboard | About 60 seconds |
Google's Data Science Agent (DS-Star), top public agent at the time | About 52% | Full benchmark, weighted (Nov 2025) | Not reported |
Trained human analysts at Adyen | About 62% | Easy tasks only (DABstep paper) | 3+ hours |
Original state of the art (DABstep paper) | About 15% | DABstep paper | Not reported |
Our 90% is on a 30-task slice, not the full leaderboard, and the comparison figures come from different task sets. What the run shows is how much the system around the model matters. With the same underlying model throughout, the first configuration scored 10%. Writing down the warehouse's unwritten conventions (fees stored in basis points, NULL meaning "matches everything", lists stored as text) took it to about 30%. Reusable skills and an AI agent that read the failures, changed the configuration and re-measured did the rest, over 14 versions, peaking at 90% on version 13.
LinkedIn and Pinterest show the same: accuracy moves with context more than with agent cleverness. LinkedIn's users add the other half: people accept an imperfect first answer they can check, not a confident wrong one they cannot.
How GetDot does agentic analytics
Dot introduced Deep Analysis in March 2025 to break a broad question into a plan, pull the data for each part and write up recommendations. In May 2026 it became much more thorough on "why" questions, backing each key driver with statistical confidence. As of October 2026 it is no longer a separate mode: the product page calls it Root Cause Analysis, and it answers "why" and "where" questions in the app, Slack or Teams. Quick lookups come back in seconds, investigations usually take a few minutes, and for the hardest questions you type "go deep" or pick the Frontier energy mode.
A published run: one question at KRY
The clearest public run is our case study with KRY, a European digital healthcare provider. In about a month of setup, Dot connected to KRY's Looker semantic layer (LookML), the team chose the explores, synced its metric glossary, and gave Dot its revenue engine framework for reasoning about growth.
Then Claire Bertrand, KRY's Data and Analytics Lead, asked: "Where are the growth opportunities for France?" According to the published case, Dot analyzed a full year of data, tested about 20 hypotheses across age, specialty, funnel step, market and booking type, and surfaced three findings, each sized in euros:
- Elevated cancellation rates in one specialty, which Dot classed as structural, not a leak, after checking the booking patterns behind them.
- Unexpectedly strong growth in one specialty: not directly actionable, but worth following up.
- A patient segment converting much worse at one booking step: many who picked the catch-all option needed a service the flow did not list. The estimated opportunity: about €800K a year, from a small change to the booking flow.
The case reports about 10 minutes to identify, analyze and size the opportunity, where testing one hypothesis by hand took KRY's team 2 to 4 hours. Claire: "We could have cut the funnel by age, by gender, by cohort, by dozens of dimensions. We just never had the time to do it systematically. Now we can." Weigh these as a vendor case study, not an independent test.
The steps GetDot follows on a "why" question
The order is deliberate: arithmetic before statistics, and statistics before stories, so a plausible correlation is not reported as a cause.
- Measure the change. Pin down the metric, its definition and the two periods being compared. If any of these is ambiguous, ask one clarifying question first.
- Split it. Break the change down by funnel step and by segment, test every segment rather than only the ones someone thought of, and rank each by how much of the change it explains.
- Find when it changed. Locate the day the metric stepped, and check seasonality and release dates around it.
- Hypothesize last. Only then write possible explanations, kept separate from findings the data supports.
- Label the evidence. Every finding carries how strong its evidence is, from an exact decomposition down to a hypothesis.
It can read free text such as support tickets alongside the numbers. It starts from a known change, so it is not an anomaly detector (scheduled monitors do that), and it is not a causal-inference engine: observational data shows where a change happened far more reliably than why.
Where GetDot meets the seven criteria
As of October 2026, with sources. It is our own product, so check it in a trial.
Criterion | What Dot does | Source |
Plans multi-step work | Plans an investigation for "why" and "where" questions, shows the plan, and takes corrections while it runs | |
Chooses its tools | SQL against your warehouse or semantic layer, Python for forecasts, regressions and significance tests, and outside sources such as web search, each one cited | |
Checks its own results | Labels each "why" finding with the strength of its evidence (steps above); Train and Evaluations score Dot against trusted answers and propose fixes that change nothing until merged; dashboard migrations are compared tile by tile with the source | |
Asks clarifying questions | Asks a short clarifying question when a question is ambiguous, instead of guessing; an answer that rests on an unsettled choice is marked "Assumption made" | |
Shows its work | The SQL behind every answer; a lineage map from warehouse tables through the queries and dbt models to the answer; an evidence tag on each root-cause finding | |
Acts on schedules and triggers | Schedules with a work gate ("only run if new orders arrived today") and a result gate ("only deliver if revenue dropped more than 5%"); plain-language alerts; each scheduled run can send fresh numbers, a brief, or a new analysis to Slack, Teams or email | |
Operates under governance | Applies the warehouse's own permissions and row-level security; stamps every query with who asked and what triggered it (on BigQuery also as job labels); "remember this" for a shared definition becomes a proposal an admin reviews and merges | August 2026 and May 2026 changelogs; Security |
The context comes from Dot's Context Agent, which pulls definitions from dbt, BI tools and docs and sends every change for review. Dot queries your warehouse in place; no data is copied or moved. All LLM providers run with zero data retention, and prompts are never stored or used for training. Dot is SOC 2 Type II, GDPR compliant with EU data processing and a DPA, and supports SSO (Okta, Entra ID, Google) with role-based access.
Where Dot is the wrong choice: if you have a platform team and want an agent shaped exactly to your stack, build one; the teardowns above are the manual. If every question lives in one BI tool whose assistant already answers it well, start there; our comparisons with Power BI Copilot, Looker and ThoughtSpot lay out the trade-offs.
FAQ
Conversational analytics is the interface: you ask in plain language and get an answer, often from one query. Agentic analytics is the behavior behind it: planning, tool choice, self-checking and acting without being asked. Most agentic products are also conversational, but plenty of conversational products are not agentic.
Mostly, yes. Companies that build their own tend to say "data agent" (OpenAI calls its internal tool one); vendors tend to say "analytics agent" or "AI data analyst". Judge any of them by the seven criteria above, not by the name.
They replace the repetitive first pass: testing every segment, writing the summary, re-running last month's analysis. Analysts still decide which questions matter, judge what a finding means, and own the definitions the agent relies on. More in Will AI Replace Data Analysts?
Build if you have engineers for the agent framework, analytics engineers for the context, domain owners who keep documentation current, and a reason your agent must be different, as Ramp, Uber and OpenAI did. Otherwise buy the same pattern (Slack and Teams delivery, versioned context, evaluations, governance). Either way, the context work is yours.
Run 30 to 50 questions your team already knows the answers to, a few of them ambiguous, on your own data: the golden-query method OpenAI and Uber's Finch use, and what Dot's Train feature automates. Score the numbers, whether it asked when it should have, and whether you could verify each answer.
Want to run the seven-criteria test on your own warehouse? Try Dot free (300 one-time credits with full Pro features), connect a read-only warehouse user, and ask it why your last dip happened.
Rick Radewagen
Rick is a co-founder of Dot, on a mission to make data accessible to everyone. When he's not building AI-powered analytics, you'll find him obsessing over well-arranged pixels and surprising himself by learning new languages.