MCP for data analysis: How to connect AI assistants to governed business data

Business teams already take analytical questions to Claude and ChatGPT. With MCP for data analysis, a Model Context Protocol (MCP) server gives those assistants a standard way to query live business data. Nobody has to paste exports into a chat window and hope the numbers are right.
Whether those answers can be trusted depends on the server. The assistant should see only what the signed-in user is allowed to see, and it should answer from governed data models and metric definitions. That way, the number in the chat can match the number on the dashboard.
Key takeaways
- MCP is an open standard that lets assistants like Claude and ChatGPT reach business data through one server per system, with no custom connector for each assistant.
- The protocol doesn't guarantee safety. Authorization is optional in the specification, so the server you choose decides whether each user sees only their own data.
- A chat answer that contradicts the dashboard often traces back to a pasted export, a shared service account, or a raw table with no metric definitions.
- Start with a few curated, well-described sources and test access with real roles before anyone connects an assistant.
What is MCP for data analysis?
MCP for data analysis is the use of the Model Context Protocol to give AI assistants such as Claude and ChatGPT a structured way to find, describe, and query business data from inside a conversation.
MCP is an open standard for connecting AI applications to external systems. Anthropic introduced it and later donated it to the Linux Foundation's Agentic AI Foundation. ChatGPT, Gemini, Microsoft Copilot, Cursor, and Visual Studio Code have all adopted it, which lets one integration reach many of them.
MCP clients, servers, and tools
The assistant application is the host. It runs one MCP client per server. Under the current specification, each request is self-contained, and no persistent session stays open. The server is the program on the other side, and it exposes capabilities through three primitives:
- Tools are functions the model can call, such as a database query.
- Resources are data that provides context, such as file contents or database records.
- Prompts are reusable templates that structure an interaction with the model.
For data analysis, tools carry most of the load, because searching for a source and running a query are both tool calls.
MCP vs custom API integrations
Without a shared protocol, each pairing of assistant and system needs its own integration. For example, five assistants and 10 data systems could need up to 50 separate connectors to build and maintain. A shared protocol reduces that to one server per system and one client per assistant. A single remote MCP server can then serve Claude, ChatGPT, and code editor agents at once.
That shared server also carries the security work, because the specification makes authorization optional and leaves enforcement to each implementation.
How MCP connects AI assistants to the data warehouse
In MCP for data analysis, the assistant answers a data question by calling tools on an MCP server. The server authenticates you, locates a relevant source, runs a query against the warehouse, and returns the result to the conversation. Follow-up questions build on the same thread.
Authentication as the signed-in user
With per-user authorization, you sign in once, and the assistant reaches the server as you. Under the MCP authorization specification, the server must reject tokens issued for other services. It also can't pass your token on to downstream systems. Whether the warehouse then enforces your permissions depends on how the server is built.
Search and description of data sources
Discovery starts with a request for the server's tool list. Each tool comes back with a name, a description, and an input schema. Data tools then return dataset and table names along with column-level metadata. BigQuery's managed MCP server is one example that works this way.
Queries on live warehouse data
The server sends the query to the warehouse when the question is asked. The warehouse does the computation and sends back a result set, and the underlying tables stay where they are. Tool errors come back inside the result, and the model can read what went wrong and correct its next query.
Answers saved as reusable analysis
A chat answer scrolls away. Some servers let the assistant save the analysis as an artifact a colleague can open, inspect, and extend, such as a workbook built on the governed source it queried. A saved artifact lets the next person check the query behind a number instead of trusting a screenshot of a chat.
Where ungoverned AI access to business data hits limits
Routes through pasted exports or shared service accounts often skip the permissions and metric definitions the rest of the analytics program depends on. The resulting answers can expose data a user shouldn't see or contradict the dashboard.
Exports pasted into chat
A CSV dropped into ChatGPT works once. Then the export goes stale like a spreadsheet emailed around on Friday, and by Monday nobody knows which version is current. The file keeps none of the source's row-level or column-level controls, and pasted exports often include sensitive data.
Shared service accounts on raw tables
Pointing a generic database MCP server at raw tables through one service account removes the per-user identity your security model relies on. When every request shares one service account and session role, a Snowflake policy keyed to CURRENT_ROLE can't tell the people asking apart. It filters every user's rows the same way.
A read-only flag is also a thin defense. A SQL injection flaw in Anthropic's reference Postgres MCP server, since archived, let callers bypass its read-only restriction. Back any read-only setting with database-level privileges.
Metric definitions left to the model
A raw schema doesn't say that your fiscal quarter starts in February or which revenue column finance certifies. Without that context, the model guesses at business meaning, and the same question can return different numbers across runs. In dbt Labs’ 11-question benchmark, accuracy reached 98% to 100% on questions a well-modeled semantic layer covered. Once the model picked the right metric, the layer wrote the query.
What a governed MCP server enforces
A governed MCP server preserves your analytics stack's existing access controls. The assistant inherits the requesting user's access, draws on governed models and definitions, and leaves a query trail the data team can monitor.
Account and connection permissions
Access control starts above the table. The server should check whether the user has an account, which warehouse connections they can use, and which tools they can call. In the Snowflake-managed MCP server, access to the server doesn’t grant access to what its tools use. Each Cortex Search service, semantic view, or agent behind a tool needs its own grant.
Row-level and column-level security
Row-level and column-level security rules need to evaluate the person asking, at query time. On Databricks, the caller's identity decides what the Databricks SQL MCP server can read. A user who signs in with their own account queries as themselves, while machine-to-machine authentication runs every request as one service principal.
Governed data models and definitions
The assistant should query curated models with reusable metric definitions, not raw tables. Define each metric once in the model, with its filters and joins. The assistant then starts from the logic your reports use and doesn't need per-prompt instructions to reach the same number.
Query monitoring
The data team should log assistant queries as agent activity where it already reviews query history, with each query tied to a person. Databricks, for example, logs each managed MCP request as an audit event with the tool name and, for SQL calls, the statement ID. A separate tag on assistant queries also shows what the assistant costs in warehouse compute.
How to roll out MCP for data analysis
A rollout of MCP for data analysis starts with a few governed sources, described in plain language. Confirm access and connect each user through OAuth before you widen coverage based on the questions people ask.
Start with a small set of governed sources
Pick a handful of curated tables or models that answer one team's recurring questions. Every extra source is another option the model can pick wrongly, so keep the list short and describe each source clearly. Hydrolix started its rollout with four curated tables of Salesforce data in a Sigma workbook, keeping only the columns its team needed.
Add plain-language context for tables, columns, and metrics
Write descriptions for someone who has never seen your schema. "Sales table" tells the model nothing. A better description says the table holds daily sales by product line, aligned with forecasted revenue, and when to use it. Add synonyms for the business jargon people actually type.
Hydrolix wrote a plain-English description for every column it kept and built a Claude skill with its standard query patterns. Its go-to-market queries grew 250% after launch.
Confirm permissions before connecting the assistant
Test as several real users with different roles before anyone connects Claude or ChatGPT. Check that each person sees the right rows, that columns hidden by column-level security stay hidden, and that no default role grants broader access than intended. Include a user with no access to the source, who should get an error rather than a partial answer.
Connect through OAuth
Use per-user OAuth instead of a shared key. Claude and ChatGPT both support custom remote MCP connectors. On Claude Team and Enterprise plans, an Owner adds the connector for the organization, and each member then connects with their own account.
Monitor usage and expand coverage
Review the query log for the questions people ask, the sources they hit, and the requests that fail. Cap spend where your warehouse allows it. BigQuery, for example, can fail a query that exceeds a byte limit you set before it incurs a charge. Add the next source based on the questions the assistant couldn't answer.
How Sigma supports MCP for data analysis
Permissions, governed metrics, and query logs are easiest to keep when the MCP server runs on the platform that already governs your analytics. Sigma is the AI apps and analytics platform that lets users build agentic analyses and applications directly on live cloud data warehouse data. It runs live warehouse queries on Amazon Redshift, BigQuery, Databricks, Snowflake, and others. Snowflake has named Sigma its top business intelligence partner four years running, and Databricks has named it Business Intelligence Partner of the Year twice.
The Sigma MCP Server brings that governed layer to the assistants your teams use. A question asked in Claude runs under the rules that govern a question asked in a workbook.
Sigma permissions inherited by the assistant
The Sigma MCP Server connects Claude, ChatGPT, and other MCP clients through each user's OAuth sign-in and applies that user's existing Sigma permissions. When the warehouse connection itself uses OAuth, queries run under each user's warehouse identity. On service-account connections, Sigma's own row-level and column-level security narrows what each user sees. Sigma validates queries before they run, and malformed SQL doesn't reach the warehouse.
IT keeps the guardrails it built, and business teams get answers in the assistant they already have open.
Answers from your data models and certified metrics
The assistant can search your Sigma content before it runs a query, and it uses your data models, certified metrics, and endorsed workbooks as context for the answer. When a question maps to a certified metric, the revenue number a finance lead gets in Claude comes from the metric the dashboard uses. Sigma also tags MCP queries, which lets the data team track assistant usage when it reviews warehouse cost.
Search, analysis, and workbook builds from Claude or ChatGPT
From Claude or ChatGPT, users can search and query governed sources on live warehouse data, then build workbooks that stay under Sigma's governance. Workbooks created through the server land in the user's personal document folder, leaving visualizations and underlying tables fully adjustable on the canvas. Generating new workbooks requires a Build license along with the "Create, edit, and publish workbooks with Sigma MCP" permission, though search and query tools remain available to all MCP-enabled users. Those agents can’t use a tool or action the user isn’t allowed to use, and their queries run as the user. The one exception is a workbook set to run with service account credentials, where agent queries run as that service account.
Try MCP for data analysis with Sigma
Pick one recurring question your team already takes to Claude or ChatGPT, such as weekly pipeline by region. Connect the assistant to a governed source, then have three people with different access levels ask it. Each answer should reflect that person's access and match the dashboard. When both hold, your MCP setup is ready to scale.
Sigma runs those queries in your warehouse under the permissions you already set, so the assistant's analysis can carry straight into a workbook or app. Get a demo, or watch the setup webinar to see it end to end.


