How to Connect Customer Feedback to Claude or ChatGPT (MCP, Files, and RAG)

Write down five to ten questions your team actually asks, and sort them into three kinds:

• Lookup: "Find the call where a customer described the SSO bug." Every option handles these.
• Summary: "What do customers dislike about onboarding?" Needs the whole period, not a sample.
• Count or trend: "How many enterprise accounts raised permissions issues in 90 days, and is that rising?" Needs option 5, or option 1 with code over tagged rows.

Pick the lightest option that answers them as often as you ask. A question you repeat monthly will outlive any export. If your feedback still sits in five tools, start with how to centralize customer feedback.

The AI can only cite what it can identify. For exports and projects, use one row per feedback item: ID, date, source, account, plan or ARR band, the customer's words, and a link to the original.

Strip signatures and quoted replies, and replace emails and phone numbers with tokens such as [EMAIL]. Put the date range in file names, such as zendesk-2026-08-to-10.csv, so stale data is obvious. For export steps by tool, redaction tips, and each provider's training rules, see how to analyze support tickets and sales calls with AI.

Paste or Upload an Export

Claude takes up to 20 files per chat, including CSV; XLSX needs code execution turned on (Claude Help). ChatGPT also reads CSVs and can analyze them with code.

The prompt matters more than the upload. Ask the model to tag each row with a theme, count rows and unique accounts per theme with code, show you the code, and quote rows word for word with their IDs.

Set Up a Claude Project or ChatGPT Project

Projects keep files and instructions across chats. Claude Projects are on every plan, five on Free, with files up to 30 MB each (Claude Help). On paid plans, Claude switches to RAG mode as project knowledge nears the context limit; past that point, treat any total as an estimate.

Put your recurring questions and theme definitions in the instructions. Replace last month's export instead of adding beside it, or rows get counted twice.

Add an MCP Server in Claude

Custom connectors work on every Claude plan, but Free allows only one. On Team and Enterprise, Owners and Primary Owners add them, plus any Enterprise custom role with Manage access to Libraries (Claude Help).

1. Open Customize → Connectors, or Organization settings → Connectors on Team and Enterprise.
2. Choose Add custom connector and paste the server URL, such as https://mcp.getdeck.io/mcp.
3. Sign in to the vendor in the browser window that opens.

Claude Code adds servers from the command line; see how to give coding agents your customer evidence.

Add an MCP Server in ChatGPT

OpenAI's docs describe connecting a server "to ChatGPT as a plugin," with read and write tools and OAuth sign-in. Writes need your confirmation by default, and any tool not marked read-only counts as a write (OpenAI). OpenAI's help center lists full MCP support, writes included, for Business, Enterprise, and Edu on the web; Pro can connect a server with read and fetch access only (OpenAI Help).

1. Open ChatGPT on the web. Workspace permissions apply, so on Business and Enterprise check that your admin allows custom servers.
2. On the Plugins page, select the plus button, then Add custom MCP server; enter a name and URL, and pick OAuth.
3. Review the risk warning, create the plugin, sign in, and check the tools ChatGPT found.

OpenAI has renamed this feature more than once, so older guides call it connectors or apps.

Build Your Own RAG Pipeline

• Store rows, not documents. Keep each feedback item as a database row with metadata and an embedding column. Postgres with pgvector is a common start.
• Split the jobs. Answer counts with SQL and quotes with vector search. Never ask a vector index for a total.
• Theme before you search. Have an LLM extract and label feedback first, so you are not searching raw tickets.
• Wrap it in MCP. One server built with the official SDKs works in Claude, ChatGPT, and coding agents.

Rule of thumb: one engineer can build a prototype in a few days. The lasting cost is upkeep: syncs, theme drift, and access control.

Ten minutes of checks prevents the worst outcome: a confident, wrong number in a roadmap review.

1. Count check. Ask for a number you counted by hand, such as accounts that reported one bug last month. If it is off by more than about 10%, do not use the setup for counts.
2. Quote check. Ask for three quotes with links on a known topic. Open each one and confirm the words match.
3. Freshness check. Ask about something a customer said yesterday. The answer shows your lag.
4. Absence check. Ask about a feature no customer has requested. A good setup says it found nothing; a weak one invents demand.

Rerun the checks whenever you change the source, the prompt, or the model.

• Start read-only. Turn on write tools after a week of reading, and approve each write.
• Decide who can connect. Claude's owner roles, Productboard's MCP roles, Canny's Owner-first setup, and Deck's Settings → MCP all control this.
• Use business accounts. Personal plans can have different training rules. Ask whether the server masks names, emails, and phone numbers in its responses.
• Remember the handoff. Once an answer reaches Claude or ChatGPT, that client's data policy applies, not the vendor's. For the rest of a vendor review, use the customer feedback platform security checklist.