Your brand managers already ask AI assistants for help with emails, plans and decks. Sooner or later they ask about research too, and the usual route is a copied chart or an uploaded export that has lost its weights and its context. MCP, the Model Context Protocol, offers a better route: the assistant asks the research platform directly and gets back an answer the platform calculated. This article explains MCP for insights leaders, not developers. It fits within our wider guide to AI market research tools, which covers the other ways teams bring AI into analysis.
Key takeaways
- MCP is an open standard that connects AI assistants to tools and data through a server, so users stop uploading files.
- The research platform does the calculation, so weights, significance tests and access control stay governed.
- An assistant can ask questions of survey data, fetch a table or chart, and draft a brief from cited results.
- Govern it like any data integration: authentication, access, logging, provider terms and human review.
- The main risks are uncited claims, broad permissions and hidden instructions in documents.
What is MCP in plain language?
MCP is a common plug between AI assistants and the systems that hold your data. The official documentation compares it to a USB-C port: one standard connector instead of a custom cable for every pairing.
For research teams, that matters because moving data by hand is where quality breaks. An exported crosstab rarely carries its bases, weighting or test results into the conversation. With MCP, the assistant asks the source system and the source system answers.
How does MCP work for research data?
A connection has four parts, and only one of them is the AI model.
- The assistant (host). The AI application your colleague is typing into.
- The MCP client. A component inside the assistant that holds the connection to one server.
- The MCP server. A program, usually run by the data owner, that offers a defined set of capabilities.
- The research platform. Where the studies, weights, tests and permissions live. The server calls it on the user’s behalf.
The server exposes two main kinds of capability. Tools are actions the assistant can call, such as “run this crosstab.” Resources are context it can read, such as a study’s question list. The assistant decides when to call a tool; the platform decides what the answer is.
What can you do with MCP, and what stays in the platform?
Through a well-built research server, a user can:
- Ask questions of governed survey data, such as “How did consideration move among 18 to 34s last quarter?”
- Fetch a chart or table the platform has already calculated, with its base and test results.
- Draft a brief from cited results, where every number points back to a study, wave and base.
What stays in the platform is everything that makes the number correct: weights applied before results leave, significance tests on the platform’s rules, and access control over which studies a user can see. Treat the assistant as a new front door to the same governed house. The same principle sits behind natural-language survey analysis.
How does MCP compare with file uploads, APIs and built-in platform AI?
| Route | Where people ask | Who calculates the numbers | Weights, tests and access | Best for |
|---|---|---|---|---|
| Uploading files to an assistant | The assistant | The assistant, from the file | Lost unless rebuilt by hand | One-off exploration of a small, clean export |
| Platform API | A custom app your developers build | The platform | Kept, if the app enforces them | Fixed integrations and data pipelines |
| Built-in platform AI | Inside the research platform | The platform | Kept | Analysts and researchers working in the platform |
| MCP connector | The assistant people already use | The platform, through the server | Kept in the source system | Business users who want answers where they work |
The upload route is where most problems start, as our guide to using ChatGPT for survey analysis explains. Connectors and built-in AI are complements: researchers work in the platform, stakeholders ask from their assistant.
Can ChatGPT and Claude connect to survey data with MCP?
Yes, both assistants can connect to MCP servers, so they can reach survey data when your research platform offers one. As of November 2026, Anthropic’s help center says custom connectors using remote MCP servers are available in Claude on its Free, Pro, Max, Ultra, Team and Enterprise plans, with admins controlling who can add connectors in Team and Enterprise workspaces. OpenAI’s developer documentation, as of November 2026, describes adding a custom server to ChatGPT, with account and workspace policies governing who can add and use one.
Both providers warn that they do not verify custom servers and advise connecting only to ones you trust. The connection gives the assistant access; it does not make it a statistician. Answer quality depends on what the server’s tools calculate and return.
What does an MCP research query look like in practice?
This workflow is illustrative; the study and numbers are invented.
A brand manager asks their assistant: “Did consideration for our brand change last quarter, and for whom?” The assistant calls the research connector’s tracker tool for consideration, Q2 versus Q3, total and by age. The platform weights the data, runs the test and returns this table.
| Group | Q2 wave (base) | Q3 wave (base) | Change | z | Significant at 95%? |
|---|---|---|---|---|---|
| Total | 34% (n=1,012) | 38% (n=998) | +4 pts | 1.87 | No |
| Age 18–34 | 29% (n=310) | 37% (n=305) | +8 pts | 2.12 | Yes |
| Age 35+ | 36% (n=702) | 38% (n=693) | +2 pts | 0.77 | No |
For the 18–34 row, a two-proportion test gives a standard error of √(0.29 × 0.71 ÷ 310 + 0.37 × 0.63 ÷ 305) = √(0.000664 + 0.000764) = 0.0378. So z = 0.08 ÷ 0.0378 = 2.12, above the 1.96 needed at 95% confidence. The total’s z of 1.87 falls short.
The assistant replies: “Consideration rose from 34% to 38% overall, but that change is not significant at 95%. The real movement is among 18 to 34s, up 8 points. Source: Brand Tracker, Q2 and Q3 waves, weighted, bases shown.” A pasted export would likely have produced “consideration up 4 points.”
What should IT and research check before connecting a server?
- Authentication. Each user signs in as themselves. Avoid shared credentials, which give everyone the same access.
- Access by study and team. Users see only the studies, markets and clients the platform already allows them.
- Logging. Record which tools were called, by whom and with what results.
- Data provider terms. Check that panel and syndicated data contracts allow results to pass to an AI assistant, and at what level of detail.
- Human review. Decide which outputs need a researcher’s approval before they reach stakeholders.
What are the limits and risks of MCP for research?
- Claims outside the tool results. The assistant can add its own explanation that no tool returned. Ask for a source on every number, and treat uncited claims as opinion.
- Over-broad permissions. A server with wide access makes every connected assistant a wide door. Expose only the tools users need, read-only for analysis.
- Prompt injection from documents. A file, web page or open-ended response can hide instructions aimed at the assistant. Keep write actions off research connectors and connect only trusted servers.
What are the common mistakes?
- Exposing respondent-level data when aggregates would do. Most business questions need a weighted table, not raw records.
- Letting the assistant run its own statistics. If the server returns counts and the assistant computes significance, you have rebuilt the upload problem.
- Dropping the base from the answer. Without it, people over-read small groups.
- Treating a fluent answer as an approved one. Anything that reaches leadership still needs a researcher’s review.
Where should you start with MCP?
Pick one tracker and one group of business users, and pilot a read-only connector with the checklist above agreed in advance. In mTab, Halo MCP on the mTab Halo platform gives access to governed Halo answers and outputs from ChatGPT, Claude and other AI assistants your teams already work in. Inside the platform, AI survey analysis answers plain-language questions with crosstabs and charts that are weighted, significance tested and traced to study, wave and base. Once the questions repeat, our guide to an AI report generator shows how to turn them into a regular readout.