AI Data Analyst: What It Is & How To Choose One In 2026
An AI data analyst is software that handles a data request from start to finish: it takes a business question in plain language, finds the data that answers it, writes and runs the queries, and explains the result with its work shown. What separates it from a chatbot with a file upload or a text-to-SQL box is context and control. It works on your live warehouse through your own metric definitions and permissions, asks when a question is ambiguous, and leaves a trail you can audit. You will also see it called an AI analyst, an AI business analyst or an AI agent for data analysis.
I'm a co-founder of GetDot, which builds one, so most of the evidence below is ours, and I label it that way. The checklist works for any vendor.
What an AI data analyst is, and what it is not
Four kinds of tools answer data questions asked in plain language. They differ mostly in where their context comes from, and that decides how often they are right.
Tool type | What it does | Where its context comes from | Where it falls short |
|---|---|---|---|
Text-to-SQL tool | Turns one question into one SQL query | The schema: table and column names | Chooses between look-alike tables by name, has no metric definitions, and stops after one query |
BI copilot | Answers questions inside one BI tool | That tool's semantic model and dashboards | Covers only what that tool models, and answers live in that tool's interface |
General AI assistant with a database MCP server | Runs SQL through a tool the assistant can call | The schema, plus whatever the person puts in the prompt | Permissions are whatever the connection's credentials allow, and definitions change with each person and prompt |
AI data analyst | Runs the whole request: clarifies, finds the data, analyzes, explains and delivers | Your warehouse, semantic layer, BI tools and docs, reviewed by your data team | Needs an owner who keeps that context current |
Conversational analytics is the interface: you ask in language and get an analysis back. Agentic analytics is the wider category of agents that plan, run several steps and act on data. An AI data analyst is the worker behind both. For named products, see our comparison of AI data analyst software, or a head-to-head such as Power BI Copilot vs Dot.
Many people first try an AI agent for data analysis through a database MCP server, which lets an assistant such as Claude or Cursor list tables and run SQL. That suits one technical person exploring a database they know, not many people who need answers that match the BI tool and rows limited by their permissions.
The 4-step workflow an AI data analyst takes over
Every data request goes through the same four steps, whether a person or an agent handles it. Here is what the agent takes over at each step, and what stays with people.
Step | Taken over by an AI data analyst | Stays with people |
|---|---|---|
1. Question | Asks one short clarifying question when two readings give different numbers, or states the assumption it made | Deciding which questions matter |
2. Data | Maps business terms to approved tables, metrics and joins, and applies the asker's permissions | Agreeing on definitions and owning them |
3. Analysis | Writes and runs SQL in the warehouse, iterates, breaks results down and tests many hypotheses in one pass | Judging whether the result makes business sense |
4. Explanation and action | Explains the result in the thread with SQL and lineage, builds the chart or deck, and schedules the recurring version | The decision |
The 12-point checklist for evaluating an AI data analyst
Run these checks on your data, with your questions and the people who will use the tool. Each point says what to test and what passes, then our evidence for Dot. Our playbook turns them into a 50-question eval set with thresholds, then shows how to run a two-to-four-week pilot and send an RFP checklist.
Accuracy and reasoning
Accuracy on your own questions. Ask about 50 questions whose answers you already trust, taken from dashboards nobody argues with, and score each one against a tolerance you set before the first run. Ask a few of them three different ways and require the same number every time.
Evidence: on Adyen's DABstep benchmark, a configured Dot answered 90% of a 30-task slice of easy and hard questions, scored with DABstep's own scorer. The same write-up puts Google's Data Science Agent, the strongest public agent at the time, at around 52% (weighted overall, November 2025, as our pricing page also cites), and trained Adyen analysts near 62% on the easier tasks after three or more hours of work. A 30-task slice is not your warehouse, which is why this check uses your own questions.
Reasoning past a single query. Include "why" and what-if questions that need several queries, and check that the numbers in the explanation reconcile with each other and with your dashboards.
Evidence: on DABstep, learning the warehouse's unwritten conventions took Dot to about 30%; the rest needed counterfactuals, multi-tier comparisons and joins that do not fit into one SQL statement (the write-up). At KRY, one open-ended question led to more than 20 hypotheses tested and a revenue opportunity of €800K a year identified and sized in 10 minutes (case study).
Improves when you configure it, and proves it. Take three wrong answers, fix the context behind them, and rerun the whole eval. A pass is a before-and-after score per question, with nothing that used to pass now failing.
Evidence: on DABstep the model stayed the same; the configuration around it took Dot from 10% to 90% on the slice. In the product, Train tests Dot against your trusted answers, fixes notes, relationships and table docs in an isolated environment, and reports accuracy before and after. Evaluations rerun on demand, including from CI.
Asks before it guesses. Put five or six ambiguous questions in the eval, such as "what's our churn rate?" when you track both logo and revenue churn. A pass is one short question back, or an answer that says which reading it used. A fail is a confident number for one reading with no mention of the other.
Evidence: Dot asks clarifying questions in the Duolingo and Choco rollouts. An answer that rests on a choice your workspace does not settle, such as an undefined metric, is marked "Assumption made" with what it assumed (how that works).
Trust and governance
Shows its work. Can the asker see the SQL, tables and definitions behind any answer, and can an analyst trace the number to its source?
Evidence: since August 2026 every Dot answer can show its lineage: a map from the warehouse tables through the exact SQL to the answer, dbt models included, plus any team note that steered the query.
Leaves an audit trail in your own logs. Take one query from your warehouse's query history and ask the vendor to trace it to the person and conversation behind it.
Evidence: every SQL query Dot runs carries a comment naming the user, the workspace, the trigger (a question, a schedule or a dashboard refresh) and a link to the chat; on BigQuery these also land as job labels in INFORMATION_SCHEMA and the billing export. Enterprise plans add audit logs of user actions and data access (security).
Uses your semantic layer instead of re-deriving it. Ask for three metrics defined in your semantic layer and compare the answers with your BI tool's numbers for the same filters and dates.
Evidence: Dot answers through your LookML, the dbt Semantic Layer, Cube, Malloy and Power BI semantic models rather than writing its own versions of your measures, and since August 2026 it applies each Looker explore's default filters instead of overriding them.
Governs changes to context. Who can change what "active customer" means? Check that every change is reviewed, versioned, testable before it goes live, and reversible.
Evidence: Dot's Context Agent pulls definitions from dbt, BI tools and docs, but nothing goes live on its own: each change is a proposal with a readable diff that you can test in its own environment, merge in part or reject, and every file keeps its history. Environments can mirror to Git for teams that want pull requests.
Security, delivery and cost
Enforces permissions and row-level security. Sign in as a regional manager and ask for revenue by country, then for a table that person cannot see. A pass is their rows only and a refusal on the restricted table, in every channel.
Evidence: Dot applies the warehouse's own permissions and row-level security, with role-based access inside Dot, and in MCP clients answers with the user's existing Dot permissions. The Team plan and above add SSO (Okta, Entra ID, Google) and Dot's own row-level rules, which can key on user attributes synced from Looker. A user with no value is denied rather than shown everything, and "Test as user" previews the exact filter a person's queries would carry (September 2026 release).
Handles data and prompts safely. Ask where the data goes, which LLM providers see prompts, what they retain, and for the audit report.
Evidence: Dot queries your warehouse in place and surfaces only the results you request; 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 and GDPR compliant, with EU data processing and a DPA (security and trust).
Works where questions are already asked. Check that the agent is where your pilot users ask data questions today.
Evidence: Dot answers in Slack, Microsoft Teams, email, its web app and MCP clients such as Claude, ChatGPT and Cursor, and turns a recurring question into a schedule. Placement drives use: Ramp's analyst agent, which lives in Slack, answered 1,476 questions in four weeks while their old #help-data channel answered 66 (Ramp engineering). We compare the chat tools in AI data analysts for Slack and Teams.
Costs scale with use, and can be capped. Estimate the cost per answered question at pilot volume and at ten times that, and check whether you can cap spend per person or team. Per-seat pricing makes you decide who may ask. Usage pricing makes you decide how much asking is worth.
Evidence: as of October 2026, Dot has unlimited users and bills on credits (usage), not seats. Free is $0 with 300 one-time credits and full Pro features. Pro is $180/month billed annually ($200 month-to-month) with 150 credits/month. Team is $720/month billed annually ($800 month-to-month) with 800 credits/month. Enterprise is custom. Admins can set weekly credit limits per user or as shared group pools, and decide who may use each model tier (pricing).
Will an AI data analyst replace your analysts?
It replaces the queue, not the people who know what the numbers mean. I wrote the long version in Will AI replace data analysts? A data founder's perspective. The evidence since then adds three things.
- Demand grows when asking gets cheap (Ramp, check 11), and every answer still rests on a definition someone has to own.
- Accuracy decays unless someone tends it. Anthropic's own analytics agent lost accuracy within a month as the data model kept changing (the playbook has the numbers and the weekly routine that catches it). Keeping context true is analyst work.
- Analysts move up the stack. In our case studies, Duolingo's analysts "reclaim time for deeper modeling and experimentation", and Choco's analytics team moved its time to "architectural and structural data problems."
The analyst who answered "what was revenue in EMEA last month?" ten times a week now owns the definition of revenue, reviews the agent's context changes, keeps the eval set current, and takes the questions that need judgment. A role that only pulled the same numbers from the same tables does shrink, and it is better to say so plainly.
FAQ
It takes a data request from question to answer: it clarifies the question, finds the governed tables and metrics, runs SQL in your warehouse, and explains the result with its SQL where the question was asked.
In analytics software, "AI data analyst", "AI analyst" and "AI business analyst" describe the same kind of product. "AI business analyst" is also used for business analysts who use AI for requirements, process maps and documentation; for that meaning, see our list of AI tools for business analysts.
As accurate as its context, so measure it on your own questions (check 1). Track trust separately, because users rate a confident wrong answer far higher than expert review does (the playbook has LinkedIn's numbers).
No. A semantic layer makes computing an agreed metric deterministic, but it does not decide which metric a question meant. Connect yours if you have one; if not, you do not need one to start.
Yes, if the vendor queries in place, applies your permissions and row-level security, runs its LLM providers with zero data retention, and shows a SOC 2 Type II report. Ask for all four in writing (checks 9 and 10).
Vendors charge per seat, per unit of usage, or by custom contract. Dot's usage-based plans are in check 12 and on the pricing page, and our AI data analyst software comparison covers other vendors' prices.
The fastest way to run this checklist is on your own data. Try Dot free with 300 one-time credits and full Pro features.
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.