Snowflake Cortex Analyst vs Databricks Genie vs BigQuery Gemini: warehouse-native AI compared

Comparison · Warehouse-native AI · 12 min read

Last updated: November 2026. What changed: Snowflake Intelligence (now Snowflake CoWork), Genie One, and Conversational Analytics in BigQuery updated; MCP servers added for all three; Hex and ThoughtSpot added.

Snowflake, Databricks, and Google shipped the same answer to the same question: put AI inside the warehouse and let people ask in plain English. This year all three also shipped MCP servers, so Claude and ChatGPT can reach them. The products differ less than the marketing pages suggest, and they still share one blind spot. Each one governs only its own data.

TL;DR · warehouse-native AI compared, by the dimension that matters

• All three turn natural language into SQL against their own warehouse, and all three now publish an MCP server. Each MCP server reaches one vendor's data with one vendor's definitions.
• The biggest difference is where definitions live. Cortex Analyst reads Snowflake semantic views. Genie One reads Unity Catalog and Genie Ontology. Conversational Analytics in BigQuery reads Knowledge Catalog and verified queries. None of them moves with you if you change warehouses.
• Hex and ThoughtSpot sit one layer up. They query many warehouses, but their definitions govern only the queries that run through them.
• Run one warehouse? The native product is usually the better pick. Run two or more, or several AI assistants, and you need one semantic layer across all of them.

The category and the contestants

A "warehouse-native" AI product lives inside the warehouse vendor's stack. It runs natural-language questions against the warehouse's own tables, with the warehouse's own auth, governance, and pricing. All three contestants were renamed or reshaped in 2026, so here's the current map.

Snowflake Cortex Analyst is Snowflake's NL-to-SQL service inside Cortex AI. It generates SQL from a semantic view, a schema-level Snowflake object that holds metrics, dimensions, and relationships. Snowflake "recommends using semantic views for all new implementations." YAML files on a stage still work with Cortex Analyst through the legacy stage API. Billed in Snowflake credits.

Databricks Genie is Databricks's NL-to-SQL experience for business users. Databricks renamed it Genie One in June 2026, and Genie Spaces are now Genie Agents. Genie One reads Unity Catalog metadata and permissions, and it resolves business terms through Genie Ontology, which is in Public Preview. Genie One and Genie Agents usage is free through January 31, 2027.

BigQuery Gemini is Google Cloud's NL-to-SQL experience. Google now ships it as Conversational Analytics in BigQuery, generally available since June 30, 2026. Its data agents are grounded in Knowledge Catalog, verified queries, custom instructions, and BigQuery Graph. Teams publish agents to Gemini Enterprise, Data Studio, or their own apps through the Conversational Analytics API.

All three are generally available (Cortex Analyst, Genie One, Conversational Analytics in BigQuery). All three now reach AI assistants over MCP:

• Snowflake-managed MCP server, GA November 4, 2025. It exposes Cortex Analyst, Cortex Search, Cortex Agents, and SQL execution as tools. Snowflake documents setup for Claude, ChatGPT, and Cursor.
• Genie One MCP, GA September 22, 2026. The older Beta endpoint (/api/2.0/mcp/genie) reaches end of life on November 30, 2026. A workspace-wide Genie MCP for Microsoft Copilot Studio entered public preview in June 2026.
• BigQuery MCP server, a remote server at https://bigquery.googleapis.com/mcp. Its tools run SQL (execute_sql, execute_sql_readonly) and read metadata. Results cap at 3,000 rows. Google lists ChatGPT and Claude as clients.

Snowflake Intelligence (now Snowflake CoWork)

Snowflake Intelligence went GA on November 4, 2025. Snowflake renamed it Snowflake CoWork in June 2026, so you'll see both names. It's the business-user app on top of Cortex Agents.

CoWork agents connect to semantic views, Cortex Search services, and custom tools. Cortex Analyst writes SQL against the semantic views, and Cortex Search covers documents. Users ask in plain language and get charts or tables, with traceability back to the source queries.

For a CTO, the key point is the layering. Semantic views hold the definitions. Cortex Analyst writes the SQL. CoWork is the chat app. The Snowflake-managed MCP server exposes the same agents and tools to Claude, ChatGPT, and Cursor.

Pick it when all your governed data lives in Snowflake and you'll invest in semantic views.

Watch for lock-in at the definition layer. Semantic views are Snowflake schema objects. On the MCP server, access to the server doesn't grant access to its tools; each tool needs its own grant.

Where each wins

Honest strengths of each, drawn from their own docs and from customer work.

Cortex Analyst wins on:

• In-warehouse latency. Cortex inference runs next to warehouse compute, so the round-trip is short.
• Definitions as schema objects. Semantic views are created with CREATE SEMANTIC VIEW, so they deploy and grant like any other Snowflake object. Snowflake can also draft one from a YAML spec, a Tableau file, or a Power BI file.
• Credit-based billing. Predictable cost; the same Snowflake credits as the rest of the AI stack, so finance can model spend without a separate SKU.

Databricks Genie wins on:

• Unity Catalog inheritance. Existing Unity Catalog policies (row-level security, column masking, attribute-based access) flow into Genie. Genie One only uses context from assets the user has permission to see. The warehouse's governance is the AI's governance.
• Reach into everyday tools. Genie One works in Teams and Slack. The workspace-wide Genie MCP brings it into Copilot Studio, in public preview since June 2026.
• Context it builds for you. Genie Ontology extracts context from metric views, dashboards, and SQL queries, including queries from dbt. Teams define and certify the rest through Unity Catalog metric views.

BigQuery Gemini wins on:

• Looker pull-through. If you already run Looker, Conversational Analytics in Looker reuses your LookML rather than asking you to author a new semantic layer. An existing semantic-layer investment carries forward.
• Reach beyond the warehouse. Conversational Analytics works with lakehouse-managed Iceberg tables. Google also names Databricks Unity, AWS Glue, SAP, and Salesforce as sources one conversation can analyze.
• Grounding options. Knowledge Catalog adds glossaries and context bundles, verified queries pin known-good SQL, and BigQuery Graph handles multi-hop questions.

Where each loses

Honest shortcomings, drawn from vendor docs and from customer escalations we've seen.

Cortex Analyst loses on:

• Cloud lock-in. Works only against Snowflake. If your data is in BigQuery, Redshift, or Postgres, Cortex Analyst isn't on the table.
• Definitions that stay in Snowflake. Semantic views are Snowflake objects. You author them, you maintain them, and you rebuild them if you ever leave Snowflake.
• Per-tool grants on MCP. Each tool on the Snowflake-managed MCP server needs its own grant. That's sound security, and it's work your platform team repeats for every tool it exposes.

Databricks Genie loses on:

• Cloud lock-in (same). Works against Databricks data. The "Lakehouse" framing implies cross-cloud, but Genie is bound to the Databricks platform.
• Moving names and endpoints. In 2026, Genie became Genie One, Genie Spaces became Genie Agents, and the Beta MCP endpoint reaches end of life on November 30, 2026. Integrations built early need rework.
• Cost after the free window. Usage is free through January 31, 2027, then pay-as-you-go. Budget controls don't apply during the free period, so track usage yourself now.
• A context layer still in preview. Genie Ontology is in Public Preview, and Databricks may change its pricing with notice.

BigQuery Gemini loses on:

• Cloud lock-in (mostly). Agents can reach other sources, but the definitions live in Google Cloud. Knowledge Catalog entries and verified queries don't move to another warehouse.
• Two surfaces, two answers. Conversational Analytics in BigQuery and in Looker can give different answers to the same question when they read different definitions. Resolving that takes governance work.
• SQL-level MCP. The BigQuery MCP server runs SQL and reads metadata, with a 3,000-row cap. An assistant using it writes its own SQL against your tables.

Two clouds, two of everything. The bigger the cloud-strategy commitment, the higher the cost of being wrong about which cloud to bet on.

Hex and ThoughtSpot

CTOs often shortlist two BI-side products next to the warehouse vendors. Neither is a warehouse. Both query many warehouses, and both now publish MCP servers.

Hex

Hex is a notebook and analytics workspace. Its AI runs in Threads, which answer data questions in plain language.

Hex semantic models define metrics, dimensions, and relationships once. Teams build them in Hex's Modeling Workbench or sync them from dbt MetricFlow, Cube, or Snowflake semantic views. Admins can turn on Governed mode to limit Hex agents to semantic models only.

The Hex MCP server (https://app.hex.tech/mcp) lets Claude, ChatGPT, Cursor, and other MCP clients search projects, run Threads, and query data. It's available on the Team and Enterprise plans.

Pick it when your analysts already work in Hex and you want AI answers next to the notebooks that produced them.

Watch for where definitions live. If they sync from dbt or Snowflake, the source of truth is upstream, and Hex is one more surface reading it.

ThoughtSpot

ThoughtSpot is a search and AI BI platform. Spotter is its AI analyst. Spotter 3 plans, validates assumptions, and self-corrects; picks the most relevant data model for a question; and adds a Research mode for multilayered investigations.

ThoughtSpot's MCP server exposes its analytics to Claude, ChatGPT, Gemini, and custom agents. It authenticates with OAuth or a bearer token and enforces ThoughtSpot's row-level and column-level security. It's an add-on for Enterprise Edition and ThoughtSpot Embedded.

ThoughtSpot has handled join traps for years. Its docs say it protects you from chasm-trap over-counting when fact tables join through a shared dimension.

Pick it when you want a full BI platform with governed search, and you're ready to model your data in ThoughtSpot.

Watch for the model living inside ThoughtSpot. Answers are governed when they go through ThoughtSpot, not when an assistant queries the warehouse another way.

The feature comparison

The same eight dimensions every Track Competitor post uses, now with all five products.

In May, the bottom row decided this table: none of the three connected natively to Claude or ChatGPT. Now all five do. The row that decides it today is the semantic layer. Each product governs answers only inside its own boundary.

What none of them do well

Every product here now reaches the assistant. Each one still governs only its own data.

That changes the question a CTO should ask. It's no longer "can my AI assistant reach the data?" It's "how many definitions of revenue will my assistants see?"

If you run one warehouse and one BI tool, that's fine. Most companies don't. A team with Snowflake for finance, Postgres for the product, and Databricks for ML connects three MCP servers to Claude. The assistant now holds three semantic formats with three ideas of "active customer."

The "warehouse-native" framing is a feature when you're a one-warehouse shop. It's a tax when you're not: two or three semantic models, governance frameworks, and pricing meters, and answer conflicts on the same question.

Where a cross-warehouse semantic layer fits

A cross-warehouse semantic layer sits between the assistants and all of your sources. It does three jobs:

1. One definition per metric. Revenue is signed off once, by a named owner, and every assistant gets the same definition.
2. Join safety across sources. Grain and join cardinality are declared, so fan and chasm traps get flagged before anyone acts on the number.
3. One audit trail. Every answer carries a receipt with the SQL, the joins, and the model version, whichever assistant asked.

Agami AI is a governed MCP server and semantic layer that gives Claude, ChatGPT, and Copilot trusted answers from your databases. It's AI-native: AI agents build and maintain the semantic layer and the MCP server, and your data team approves every change. They start from what you already have, reading metadata from your BI apps, documents, and spreadsheets to draft the semantic model.

Agami works with any SQL database, including Snowflake, Databricks, BigQuery, Redshift, and Postgres, and queries each one in place. It runs in your cloud. Row-level and column-level security stay in your database, where you already set them.

The wedge isn't that Agami writes better SQL than Cortex. Cortex is excellent on its in-warehouse path. The wedge is one governed model across every warehouse and every assistant, owned by your team rather than by a warehouse vendor.

When you don't need it: one warehouse, users who are happy in that vendor's app, and no plan to change. The native product is the better fit there.

When you'd pick each

One decision matrix. Honest about where Agami isn't the answer.

Frequently asked questions

Q: How should a CTO compare Snowflake Intelligence, Databricks Genie, Hex, and ThoughtSpot?

A: Start with where your definitions live. Snowflake and Databricks govern their own warehouses. Hex and ThoughtSpot govern queries that run through them. If you have more than one warehouse or AI assistant, add a cross-warehouse semantic layer so every tool gets the same metric.

Q: Is Snowflake Intelligence the same as Snowflake CoWork?

A: Yes. Snowflake renamed Snowflake Intelligence to Snowflake CoWork in June 2026. It's the business-user app built on Cortex Agents, which use Cortex Analyst for structured data.

Q: Is Databricks AI/BI Genie the same as Genie One?

A: Yes. Databricks renamed Genie to Genie One in June 2026, and Genie Spaces became Genie Agents. The Genie One MCP went GA on September 22, 2026.

Q: Can Claude and ChatGPT use these tools now?

A: Yes. Snowflake, Databricks, Google, Hex, and ThoughtSpot all publish MCP servers. Each one reaches only its own platform's data and definitions.

Q: Do I still need a semantic layer if my warehouse has an AI assistant?

A: Not if you run one warehouse and stay in that vendor's app. Yes if several warehouses or assistants need the same numbers.

Q: Are Cortex Analyst, Genie, and BigQuery Gemini direct competitors to each other?

A: Not really. Each is bound to its own cloud, so a single customer rarely shops between two of them. The real competition is on cloud strategy: which warehouse the customer commits to. Once that's decided, the warehouse-native AI is a follow-on choice within that ecosystem.

Q: Why are Hex and ThoughtSpot in a warehouse-native comparison?

A: Because CTOs shortlist them together. Hex and ThoughtSpot are BI-side products that query many warehouses, while the warehouse products govern one. Comparing all five shows where each one keeps its definitions.

Q: What if I want all three (Cortex, Genie, and BigQuery)?

A: Multi-cloud teams sometimes do exactly this. Think a semantic view in Snowflake for finance, Genie One for product analytics, and a BigQuery agent for marketing. The cost is reconciling three semantic models, three governance frameworks, and three pricing meters. That's the gap one semantic layer across warehouses closes.

Q: What's the latency story across the three?

A: All three slow down on complex questions or large schemas. The variance comes from the query plan more than the AI; the AI is the small share of the budget.

Q: Are any of these portable across clouds?

A: No. Each binds its definitions to its platform. The closest thing to portability is LookML, which travels to other Looker deployments, and Hex, which syncs definitions from dbt MetricFlow, Cube, or Snowflake. In both cases the source of truth still lives in one vendor's format.

Q: How does Agami handle the single-warehouse case?

A: Agami works fine against a single Snowflake, BigQuery, Databricks, or Postgres warehouse. It doesn't require multiple warehouses; it just doesn't require one. If you stay on one warehouse and one vendor app, the native product is often enough. Agami earns its place when you want a model your team owns and the same answers in Claude, ChatGPT, and Copilot. It also puts an audit trail on every answer.

References

1. Snowflake, "Cortex Analyst documentation."
2. Snowflake, "Semantic views vs. the legacy stage API."
3. Snowflake, "Overview of semantic views."
4. Snowflake, "Overview of Snowflake CoWork."
5. Snowflake release note, "Snowflake CoWork (formerly Snowflake Intelligence) general availability," November 4, 2025.
6. Snowflake release note, "Snowflake-managed MCP server general availability," November 4, 2025.
7. Snowflake, "Snowflake-managed MCP server."
8. IT Brief, "Snowflake renames AI tools amid broader platform push," June 3, 2026.
9. Databricks, "The Genie One MCP is now generally available," September 22, 2026.
10. Databricks, "Genie One MCP documentation."
11. Databricks, "Genie Ontology."
12. Microsoft Learn, "AI/BI and Genie One release notes 2026."
13. Microsoft message center, "AZ564868: Azure Databricks workspace-wide Genie MCP for Microsoft Copilot Studio," June 2, 2026.
14. Google Cloud, "Conversational Analytics in BigQuery now GA," June 30, 2026.
15. Google Cloud, "Use the BigQuery MCP server."
16. Hex, "MCP server."
17. Hex, "Intro to semantic models."
18. ThoughtSpot, "MCP integration."
19. ThoughtSpot, "Spotter versions."
20. ThoughtSpot, "Chasm traps."
21. Agami blog, "The Semantic Model Is Having Its Moment."
22. Open Semantic Interchange, "v1 specification."

Have data in more than one warehouse? Run the same question against each.

Bring a question that crosses two of your warehouses. We'll trace it through the native products and Agami side by side, so you see where the definitions differ. The harder question for the buying committee isn't "which warehouse-native AI wins." It's "how many definitions of the same metric will our assistants see?"

Book a demo or try agami-core (source-available) →