How-To

Pipeline Analysis in ChatGPT or Claude With Your Own Data

Three ways to get your data in front of an assistant, how to check its numbers, and what to settle before you trust an answer about pipeline.

5 min read

Asking an assistant "which deals are at risk?" only works if it can see your deals. There are three ways to make that happen, and they suit different situations.

Three Ways to Connect Your Data

RouteHow it worksGood forWatch out for
Upload a fileExport a CSV and attach itA one-off questionThe data is stale the moment you export it
Live connectorThe assistant queries the tool directly through MCPQuestions you ask every weekAccess and permissions need thought
Warehouse layerThe assistant queries a governed copy of your dataLarger teams with a data functionSetup time and cost

For most small teams the live connector is the right middle. See what MCP is for how it works.

Agree on Definitions First

Before you ask anything, write down what your stages mean, what counts as qualified, and which date field is the real close date. Give that to the assistant at the start. Most wrong answers about pipeline are right answers to a different definition.

Questions That Work Well

  • Which open deals have had no activity in two weeks?
  • Which channel produced the most paying customers last quarter?
  • Which accounts became more active on our site this month, and are any of them open deals?
  • What did this account do before it upgraded?

The third and fourth need data your CRM does not hold: what accounts did on your site and in your product. That is why connecting analytics alongside the CRM changes what you can ask.

Check the Numbers

Do this every time until you trust the setup.

  1. Ask one question you already know the answer to. Last month's new customers, for instance.
  2. Compare it to the report in the source tool.
  3. Ask how it got there. Which tool it called, and with what filters.
  4. Fix the mismatch before asking anything you do not know the answer to. It is usually a date range, a definition, or a filter.

Access and Privacy

A connector can do whatever its tools allow. Anthropic's documentation notes that connectors can read, create, modify, and delete depending on the tools offered, and that you can turn individual tools off. OpenAI's documentation says ChatGPT asks for confirmation before write actions by default, and that OpenAI does not build or verify custom servers itself.

So:

  • Use read-only tools for analysis. You do not need write access to answer a question.
  • Prefer official servers from the vendor whose data it is.
  • Check your assistant's data settings. Whether conversations are used for training differs by plan and provider. Look at the current policy for the plan you are on.
  • Remember that CRM data includes people. Names and emails pulled into a chat are still personal data.

How Leapio Fits

Leapio's MCP server gives an assistant the account side of the picture: events, funnels, companies, and revenue, at the company level. Tools are read-only unless an owner or admin turns writes on, access is scoped to your organization, and analytics that a visitor declined are never counted. It works in Claude, ChatGPT, Cursor, and other assistants. Reading your data this way is included on every plan. Start with the setup guide.

Get Started

See Your First Events in Minutes

Free to start, no credit card. Add signals and actions when you are ready.

AI-Native Go-to-Market Analytics That Drives Revenue and Action