Analytics Agent

The analytics agent that doesn't write SQL

Ask about revenue, a funnel, retention or an A/B test in plain English. Mitzu's agent composes each analysis against the semantic layer your data team governs — and a deterministic query engine generates the SQL on your warehouse, so the same question gets the same answer.

What is an analytics agent?

An analytics agent is an AI system that answers business questions by planning and running analyses on your data, then returning the answer with the evidence behind it. A dashboard answers the question it was built for; a chatbot that turns one sentence into one SQL query is a translation layer. An agent takes the follow-up question and investigates from several angles before it answers.

Mitzu's analytics agent answers questions, explains why metrics moved and runs deep dives on your governed metrics. It builds analysis specifications that a deterministic engine turns into SQL, so the same question gets the same answer. It runs on the data already in Snowflake, BigQuery, Databricks, Redshift, ClickHouse and other warehouses, and needs one there: events, revenue and CRM tables in one place are what let it answer across them.

Read the full definition, the main architectures and what makes an agent trustworthy.

How it works

The agent composes the analysis. The engine writes the SQL.

Mitzu's Analytics Agent does not write SQL. It composes analysis specifications — funnels, retention, segmentations — against a semantic layer specialised for product analytics, and a deterministic query engine turns each specification into SQL that runs in your data warehouse. That is why funnel, retention, and cohort methodology stays correct instead of depending on a model getting the SQL right.

1

The agent interprets the question

It maps your words to real events, properties, and cohorts in the semantic layer that the Configuration Agent built from your warehouse — filters are suggested from sampled values that actually exist, not invented.

2

It composes an analysis specification

A metric and its breakdown, a funnel's steps and conversion window, a retention cohort and its return event, an experiment and its metrics. The agent fills in these parameters. It never authors query text.

3

A deterministic engine generates the SQL

The same specification always produces the same SQL, executed in your warehouse. Funnel, retention, and cohort methodology lives in the engine, so the model cannot get it wrong — and the SQL is shown for your analysts to verify.

The same architecture powers the agent in the app, in Slack, and over MCP, on top of Mitzu's warehouse-native architecture — read the full mechanism reference at docs.mitzu.io: how the agent works.

Analytics agent vs dashboards, text-to-SQL and BI chatbots

Most tools now answer a question in plain English. They differ in who writes the query, what data it can reach and whether the answer holds up when you ask again.
How an analytics agent compares with static dashboards, text-to-SQL tools and agents inside product analytics tools
DimensionStatic dashboardsText-to-SQL and BI chatbotsAgents inside a product analytics toolMitzu's analytics agent
Who writes the SQLAn analyst, once per chartThe language model, on every questionThe vendor's query engine, over its own event storeA deterministic engine, from the agent's analysis specification
Same question, same answerYes, for the questions the chart was built forNot guaranteed: the query can change between runsYes, inside the toolYes: same specification, same SQL
Follow-up and “why” questionsA new request to the data teamOne query per question; the investigation is up to youYes, on the data the vendor ingestedRoot cause, impact analysis and deep dives
Funnel and retention methodologyHand-built for each chartWritten freehand; conversion windows and cohort bucketing are easy to get wrongBuilt inBuilt into the engine
Data it can reachThe modelled tables behind each dashboardThe warehouse it is attached toEvents sent to the vendor; billing, CRM and support data sit elsewhereEverything in your warehouse: events, revenue, CRM and experiment exposures, joined
Metric definitionsIn the BI modelA semantic layer you write and maintain, or noneThe vendor's ownGoverned by your data team and shared with every dashboard
Where people use itThe BI toolA warehouse console or chat windowInside the vendor's appThe Mitzu app, Slack, and MCP clients such as Claude and ChatGPT

The longer version, with a checklist for evaluating tools in a pilot: analytics agents vs text-to-SQL and BI chatbots.

Where teams use an analytics agent

The same agent and the same definitions, from a product diagnosis to a revenue question.

Product teams

“Why did week-2 retention drop after the v2.3 release?”

Diagnose funnel drop-offs, cohort behaviour and activation friction without waiting for an ad-hoc query.

Marketing and growth

“Did the email campaign move conversion?”

Connect campaigns and channels to product outcomes and revenue, on the same definitions the rest of the company reports.

Leadership and operations

“Revenue dropped last week. Which funnel step, and which users?”

Go from a KPI to the funnel step and the users behind it in one question, because the KPI and the funnel share one set of definitions.

Data teams

Define net revenue once.

Keep governance while everyone self-serves: the agent, dashboards and Slack answers all use the metrics and permissions you set.

Get started in minutes, not weeks.

From connection to insight in under 10 minutes. No complex setup, no data engineering required.
  1. 01/2 min

    Connect your data warehouse

    Agent config scans your warehouse and creates semantic layer.
  2. 02/5 min

    Create a semantic layer

    No need to write YAML at all. The produced data catalog is designed specifically to work with product analytics.
  3. 03/Instant

    Ask anything

    Our deterministic SQL engine makes sure there are no hallucinations.

Built for questions that demand deep diagnosis.

Most agentic analytics tools wait for simple questions. Mitzu tackles the hard ones, working from multiple directions and returning findings built on product analytics methodology: funnels, retention windows, and cohorts, applied consistently every time.

Did that release actually move conversion? Did the campaign change behaviour? Mitzu compares cohorts, finds correlations, and tells you whether that feature, campaign, or product change made a real difference.

#analytics
Natural Language

Ask questions like you're talking to a colleague

Type the question the way you would ask an analyst. The agent picks the analysis, runs it on your warehouse and answers with the chart and the SQL behind it.

  • Revenue, funnel, retention and A/B test questions in one conversation
  • Filters come from values that exist in your data, not guesses
  • Follow-up questions keep the context of the thread
  • Every chart can go straight onto a shared dashboard
Try AI Queries
Scheduled Agents

A dashboard you have to remember to check is worse than one that watches itself

Scheduled agents run your questions on a cadence and email you only when the results meet a condition you set — a metric drop, a threshold crossed, a cohort behaving differently.

  • Runs on the schedule you choose
  • Alerts only when your condition is met
  • Same semantic layer and deterministic engine as every other surface
  • Findings arrive with the supporting charts
See Automated Insights

Works with your modern data stack.

Native integrations with the data platforms you already use. Connect in minutes, not weeks.
Snowflake
Databricks
BigQuery
Redshift
AWS Athena
Postgres
Trino
Firebolt
ClickHouse
Fabric

Trusted by leading companies

BrokerChooser
Prezi
Khatabook
Fluenta
Guided eLearning
Raptor
Munch
Suunto
52 Entertainment
Colossyan
Nansen
Shapr3D
Transfr

0

Lines of SQL written by the agent

1

Deterministic engine behind every answer

11

Warehouses queried in place

3

Surfaces: app, Slack, and MCP

Analytics agent FAQ

What teams ask before they put an analytics agent on their warehouse.

How is an analytics agent different from a BI chatbot or text-to-SQL?

A BI chatbot or text-to-SQL tool has a language model write one query per question. An analytics agent plans an investigation, runs several analyses and synthesises the result. In Mitzu the agent does not write SQL at all: it composes funnel, retention, metric and experiment specifications, and a deterministic engine generates the SQL, which is shown for review.

What can an analytics agent do?

Mitzu's agent answers quick questions such as net revenue last week, explains why a metric moved, measures the impact of a release or a campaign, and runs deep dives that fan out into many analyses and return a report. Charts it builds can go straight onto shared dashboards, and scheduled monitoring agents review KPIs and email the team when a condition you describe is met.

Can the agent answer funnel, retention, and cohort questions reliably?

Yes — that is what it is built for. The agent composes the analysis specification (funnel steps and conversion window, retention cohort and return event, segmentation filters), and a deterministic query engine turns it into SQL using product analytics methodology. The same specification always produces the same SQL and the same answer.

How does Mitzu prevent hallucinated SQL?

Mitzu's agent does not write SQL, so there is no model-authored query to hallucinate. The deterministic query engine generates the SQL from the agent's analysis specification, grounded in a semantic layer with sampled filter values from your warehouse. The generated SQL is shown for your analysts to verify.

Should we build our own analytics agent or buy one?

Building one on an LLM with MCP or a framework such as LangChain is a reasonable project for a narrow use case. The slow part is not the chat loop but the methodology and governance: conversion windows, cohort bucketing, experiment statistics, metric definitions and permissions. Mitzu ships those in its engine and semantic layer and exposes them over MCP, so a team that wants its own assistant can use Mitzu as the governed backend.

How do you measure whether an analytics agent is working?

In a pilot, track three things: whether answers match analyst-reviewed SQL on a fixed set of questions, whether asking the same question again returns the same answer, and how many questions get answered without a ticket to the data team. Mitzu shows the SQL behind every answer, so the first check is a review rather than a rebuild.

Does an analytics agent replace data analysts?

No. It takes recurring and exploratory questions off the analysts' queue, so the data team spends its time on models, definitions and the investigations that need judgement. The aim is to multiply the team's analytics capacity, and the definitions the agent uses still belong to the data team.

Which warehouses does the analytics agent run against?

Mitzu connects to Snowflake, BigQuery, Databricks, Redshift, Athena, PostgreSQL, ClickHouse, Trino, Starburst, Firebolt, and Microsoft Fabric — querying data where it already lives, without copying events out of your warehouse.

Can teams use the agent in Slack or their own AI assistant?

Yes. Teams can ask @mitzu in Slack, or use Mitzu from Claude, ChatGPT, Cursor or GitHub Copilot over MCP. Every surface uses the same governed definitions and Mitzu roles, so each person sees only the data their role allows.

How do we try it, and how is it priced for multiple teams?

Start a 14-day free trial, connect your warehouse and check the answers and the SQL against your own questions. Pricing is seat-based with no per-event metering; talk to sales for a rollout that covers analysts, PMs and executives.

Try the analytics agent on your own warehouse

Connect Snowflake, BigQuery, Databricks or another warehouse and ask your first question. The SQL behind every answer is there to check.