AI Agents for Analytics: What They Do and Which to Pick
September 17, 2026
AI agents for analytics don't just turn one question into one chart. They plan a multi-step analysis, write and run SQL (and sometimes Python), check what came back, and explain the result. There is no single most recommended one. Your safest pick is usually the agent native to where your data and metric definitions already live, because it can use your permissions and semantic model without copying either.
The short version:
- Databricks, Snowflake, Microsoft, and Google each ship an agent that works inside their platform and inherits its access controls.
- ThoughtSpot and Tableau sit on top of your data, with their own semantic layer doing the heavy lifting.
- All of them get better with curated context: metric definitions, synonyms, example queries, and test questions.
- The real risk isn't a failed answer. It's a confident, wrong number from a bad join or a drifted metric.
Quick comparison
| Service | Where your data must live | What the agent does | Governance hook |
|---|---|---|---|
| Databricks Genie Agents (with Genie One for business users) | Tables, views, or metric views registered in Unity Catalog | Chat mode writes a single SQL query; Agent mode plans, runs multiple queries, and returns a report | Unity Catalog permissions, evaluated as each end user |
| Snowflake Cortex Agents, which power Snowflake CoWork | Snowflake, with semantic views for structured data | Plans, calls tools (SQL via Cortex Analyst, search, Python sandbox, charts), reflects, responds | Snowflake roles and privileges, plus each tool's execution context |
| Fabric data agents and Copilot in Power BI | OneLake: lakehouses, warehouses, Power BI semantic models, KQL databases | Picks a source, generates read-only SQL, DAX, or KQL, validates and runs it | The requesting user's credentials, plus Microsoft Purview policies |
| Conversational Analytics in Looker and BigQuery data agents | A Looker model (LookML) over your database, or BigQuery | Answers from the LookML model; can write and run Python for advanced analysis (preview) | Looker model permissions; IAM in BigQuery |
| ThoughtSpot Spotter | Cloud data warehouses ThoughtSpot connects to live | Breaks broad questions into sub-questions, analyzes each, summarizes | ThoughtSpot Models, coaching, and row-level security |
| Tableau Next (Agentforce Tableau) and Tableau Agent | Data 360 objects for Tableau Next; Tableau Cloud data sources for Tableau Agent | Conversational Q&A, data prep, and visualization help | Tableau Semantics models; row- and column-level security policies |
What AI agents for analytics do that a BI copilot doesn't
Ask "why did gross margin drop in August?" A BI copilot writes one query and shows margin by month. An agent confirms the drop, lists likely causes (product mix, discounts, one big customer), queries each, drops the dead ends, and writes up what it found with the SQL attached.
Databricks describes its Genie Agent mode this way: confirm the spike, explore possible contributors, reflect on each query's results, and finish with a report that includes the underlying SQL. Snowflake's docs describe the same loop: plan, use tools, reflect and respond.
To do that reliably, an agent needs three things:
A semantic layer. A semantic layer maps raw tables to business terms like customer, product, and revenue. Without one, the agent guesses what "active customer" means from column names. Each vendor has its own: LookML in Looker, semantic views in Snowflake, the knowledge store and metric views in Databricks, Power BI semantic models, Tableau Semantics, and ThoughtSpot Models.
Governed access. Row filters and column masks must apply to the agent's queries, so it sees only what the person asking may see.
A way to check its work. Benchmarks, verified queries, and visible SQL. If you can't see how a number was produced, don't put it in a board deck.
If you're still working out which layer of the stack you're missing, our guide to AI analysis software for enterprise teams splits it into platform, transformation, interface, and generative layers.
Six AI agents for analytics, platform by platform
Databricks: Genie Agents and Genie One
Per the 2026 AI/BI release notes, Genie spaces are now Genie Agents, and the business-user experience is now Genie One. Agent mode in Genie Agents became generally available in July 2026, and its APIs followed in August.
Data must be registered in Unity Catalog. Authors teach Genie through a knowledge store (table descriptions, synonyms, join relationships), example SQL, and instructions. Unity Catalog evaluates data access as the end user, not the author. Benchmarks measure accuracy on a fixed set of test questions, including in Agent mode.
Fits you if your lakehouse and metric definitions already sit in Databricks.
Snowflake: Cortex Agents and Snowflake CoWork
Snowflake CoWork is the new name for Snowflake Intelligence. It's the business-user app, and Cortex Agents are the engine underneath. An agent can call Cortex Analyst to generate SQL over semantic views, Cortex Search for unstructured documents, a Python sandbox for calculations, and a charting tool.
Access runs through Snowflake privileges and each tool's execution context, and CoWork inherits row access policies and column-level security. Billing covers orchestration tokens plus the tools an agent uses, and you can set an orchestration budget per agent.
Fits you if Snowflake is your warehouse and you want one agent across tables and documents.
Microsoft: Fabric data agents and Copilot in Power BI
A Fabric data agent is generally available and connects to up to five sources: lakehouses, warehouses, Power BI semantic models, KQL databases, ontologies, or Microsoft Graph. It writes SQL, DAX, or KQL, validates the query, and runs it using the requesting user's credentials. Microsoft Purview policies apply on top.
Three limits matter: queries are read-only, chat responses are capped at 25 rows and 25 columns, and only English is supported for now. Configurations can go into Git and move through deployment pipelines.
For report users, the standalone Copilot in Power BI (in preview) finds the right report or semantic model and answers from it. Microsoft warns answers may be limited or inaccurate unless the author has used Prep data for AI (AI data schemas, verified answers, AI instructions) and marked the model Approved for Copilot.
Fits you if your organization already runs on Fabric capacity and Power BI semantic models.
Google Cloud: Looker Conversational Analytics and BigQuery data agents
Conversational Analytics in Looker uses the LookML model as its source of truth. Explore conversations and Explore data agents are generally available, and a data agent can connect to up to five Explores. Advanced analytics, where the agent writes and runs Python, is in preview. Sharing an agent doesn't grant access to the Explores behind it.
Conversational analytics in BigQuery is still in preview. Its data agents take instructions, verified queries, and glossary terms, and Google's own docs tell you to validate all output.
Fits you if you already maintain LookML, or your warehouse is BigQuery.
ThoughtSpot: Spotter
Spotter connects live to your cloud data warehouse. It breaks a broad question into specific questions, analyzes each, and summarizes the findings. SpotterModel helps data engineers build governed semantic models, with human review before anything ships. Data teams coach Spotter with reference questions, business terms, and instructions, and ThoughtSpot supports row-level security.
Fits you if you want one analytics layer on top of your warehouse that business teams across the company share.
Tableau: Tableau Next and Tableau Agent
Tableau Next is built on the Salesforce Platform. Its data is always a Data 360 object, its semantic layer is Tableau Semantics, and Agentforce provides the Data Pro, Concierge, and Inspector skills.
Tableau Agent is the assistant in Tableau Cloud. It needs a Tableau+ site and a Creator or Explorer role, and it respects row- and column-level security policies. It helps build views and calculated fields, but per Tableau's help page it can't create dashboards or pick a data source for you.
Fits you if your CRM data lives in Salesforce, or your analysts already work in Tableau.
How to evaluate AI agents for analytics
Demos use clean data and friendly questions. Judge each agent on these five things instead.
| What to check | How to test it | Red flag |
|---|---|---|
| Accuracy on your own questions | Run 30 to 50 real questions with known, finance-approved answers | The number is close but not exact, with no explanation |
| Semantic model quality | Ask about your three most disputed metrics | Two phrasings of the same question return different totals |
| Audit and lineage | Open the SQL or query plan behind each answer | You can't see which tables and filters were used |
| Permissions | Ask the same question as two users with different row-level access | Both users get the same total |
| Cost model | Map how usage is billed (tokens, compute, capacity, seats) and whether you can cap it | No way to set a budget or see usage per agent |
Failure modes to test for
Confident wrong joins. An agent joins orders to order lines to shipments and double-counts revenue. The answer looks tidy and is wrong. Define join relationships in the semantic layer; Databricks' knowledge store and Snowflake's semantic views both capture them.
Metric drift. "Revenue" in the agent's instructions means gross; on the finance dashboard it means net of refunds. Keep one definition in the semantic layer and point both at it.
Silent truncation. A capped result set gets summarized as if it were complete. Ask for totals, not lists.
Ambiguous time. "Last quarter" can mean calendar or fiscal. Put your fiscal calendar in the instructions and add a test question for it.
Nondeterminism. Microsoft and Databricks both describe their agents as nondeterministic, so the same question can get different answers. Rerun your benchmark more than once.
For keeping AI inside policy more broadly, see our roundup of tools for managing AI governance in workflows.
A 30-day pilot plan for an analytics agent
A template; adjust the timing to your team.
| Week | What you do | What you have at the end |
|---|---|---|
| 1 | Pick one domain (say, sales pipeline) and the agent native to that data. Collect 30 to 50 real questions with approved answers. | A benchmark set and a named owner |
| 2 | Add context: metric definitions, synonyms, joins, fiscal calendar, a few example queries. Limit the agent to the tables it needs. | A scoped agent with documented context |
| 3 | Run the benchmark, read the SQL for every miss, and fix the context, not the questions. Test permissions with two users. | An accuracy baseline and a permissions check |
| 4 | Open it to a small group. Log questions, flag wrong answers, review usage and cost. | A go or no-go decision backed by evidence |
Two rules keep a pilot honest: don't expand scope until the benchmark stabilizes, and make "show me the query" a habit for anyone sharing an agent's number outside the team.
If the result still becomes a chart someone rebuilds by hand, see our guide to AI data visualization. And if your bottleneck is the workflow around the analysis rather than the query, Taku is an AI-native desktop workspace where you can mirror an AI workflow someone already got working, run it, and save it so next month's analysis doesn't start from a blank prompt. Taku is in Beta, and the Mac app is available now.
Key points
- AI agents for analytics plan, query, check, and explain. A BI copilot answers one question with one query.
- There's no single most recommended service. Start with the agent native to where your data and semantic model already live.
- Databricks Genie Agents, Snowflake Cortex Agents, Fabric data agents, and Google's Conversational Analytics inherit their platform's access controls. ThoughtSpot Spotter and Tableau lean on their own semantic layers.
- A semantic layer and governed access do most of the work. Missing definitions and joins are a common source of wrong answers.
- Test on your own questions, check permissions with two users, read the SQL, and watch for wrong joins, metric drift, and truncation.
- Pilot one domain for about a month before you roll anything out.
FAQ
What is the most recommended agentic analytics service in data analytics?
There isn't one that fits everyone. The most sensible default is the agent native to your data platform: Genie Agents on Databricks, Cortex Agents on Snowflake, Fabric data agents on Microsoft Fabric, or Conversational Analytics on Looker and BigQuery. Consider ThoughtSpot Spotter or Tableau when your analytics layer already lives there.
What are AI agents for analytics?
They're AI systems that take a business question, plan the analysis, generate and run queries or code against your data, evaluate the results, and explain what they found. Unlike a single text-to-SQL call, they can run several queries in a row and change course based on what each one returns.
Do AI agents for analytics need a semantic layer?
They run without one, but accuracy suffers. Every major vendor asks you to supply business context: LookML, Snowflake semantic views, Databricks' knowledge store, Power BI's Prep data for AI, Tableau Semantics, or ThoughtSpot Models. That context tells the agent what your metrics mean and how your tables join.
Can an analytics agent change or delete my data?
It depends on the product and on the permissions you grant. Microsoft states that Fabric data agents generate only read queries. Other platforms let agents call custom tools or take actions in connected apps, so review each agent's tools and give its identity the least access it needs. Our guide to AI agents for enterprise covers access design in more depth.
How accurate are AI agents for analytics?
A general figure can't tell you how an agent will do on your data. Accuracy depends heavily on your semantic model, your instructions, and how messy your schema is. Build a benchmark of real questions with approved answers, and measure it yourself before and after tuning.