Use Claude or ChatGPT With HubSpot Data

Pasting an export into a chat works once. Making it work every week, with the same definitions, is a different problem. Three setups, what each costs, and where an MCP-connected data layer fits.

You can use Claude or ChatGPT with HubSpot data in three ways: paste or upload an export, use a connector that lets the assistant call HubSpot directly, or put a governed data layer between them. The first is fine for a one-off. The last is what keeps numbers consistent week after week. This post compares them and explains how Quiriz, which exposes an MCP server, plays the data-layer role.

Three setups, compared

1. Paste or upload an export. Export deals to CSV, drop the file into a chat, and ask questions. It takes five minutes and needs no setup. The costs are that the data is already stale when you upload it, large portals exceed what fits in a conversation, and the assistant picks its own formula for anything ambiguous.

2. Give the assistant a live connector. Some connectors let an assistant query HubSpot records through the API. Data is fresher, but the assistant is still computing metrics itself from whatever the API returns. Properties like stage and close date arrive raw, and the stage history that average-days-in-stage needs is not a simple field to fetch.

3. Put a governed data layer in between. HubSpot is connected to a system that already holds your datasets and your metric definitions, and the assistant asks that system for answers. The assistant does the conversation and the explanation. The data layer does the arithmetic from stored definitions.

Why the difference shows up as drifting numbers

Ask a general assistant for win rate by owner in Monday's chat and again in Thursday's. Unless the prompt says otherwise, the assistant decides whether open deals are in the denominator, whether to use close date or create date, and whether "won" means a stage name or the closed-won flag. All of those are reasonable. They produce different numbers, and nobody notices until a slide is quoted against last month's slide.

This is not a flaw unique to any one assistant. It is what happens when a definition lives in a prompt rather than in a place that persists. The fix is to move the definition out of the conversation.

What MCP adds

MCP, the Model Context Protocol, is a standard that lets an AI assistant call external tools. Quiriz exposes an MCP server. An assistant that supports MCP can connect to it and ask Quiriz a question about your connected data. Quiriz runs the query against the datasets, applies the metric definitions your admin stored in Company Context, and returns the answer with the work shown. The assistant then explains it, drafts the follow-up, or turns it into a summary.

In practice the split looks like this. You say, in Claude or ChatGPT where MCP is supported, "What is our win rate by deal owner this quarter, and which owner changed most from last quarter?" The assistant asks Quiriz. Quiriz returns win rate as closed-won over closed deals by close date. The assistant writes the narrative. Both assistants and the Quiriz web app return the same figure because it is computed in one place. The setup steps for your assistant are in the Quiriz help documentation; they vary by assistant and change over time, so we do not repeat them here.

If you want the comparison without MCP, there are write-ups of Quiriz versus ChatGPT and Quiriz versus Claude.

What HubSpot data is available

The HubSpot connector is read-only and refreshes daily. It supplies five datasets:

• [HubSpot] Deals with amount, stage, owner, closed flags, close and create dates, days to close, loss reasons and source fields.
• [HubSpot] Contacts with lifecycle stage, lead status and lifecycle dates.
• [HubSpot] Companies with industry, size and owner.
• [HubSpot] Deal stage history with entered and exited dates and days in each stage.
• [HubSpot] Pipeline snapshots, a dated copy of the open pipeline each day, which makes "what changed since last Monday" answerable.

Metrics ship as suggested definitions: Pipeline value, Win rate, Average deal size, Sales cycle (days), Closed-won revenue, Weighted pipeline, Lost deals, Average days in stage and Deals entering stage. An admin confirms or edits them once.

Questions that work well through an assistant

• "How has open pipeline changed since last Monday, and which deals moved?"
• "Why are we losing deals this quarter? Group by lost reason and summarize the pattern."
• "Which stage do deals sit in longest, and is it getting worse?"
• "Draft three bullets for the sales review from this quarter's win rate and cycle time."

The last one shows the division of labor. The numbers come from the data layer. The wording is the assistant's job.

Which setup to choose

A pasted export is the right choice for a one-time question you will never ask again, such as sizing up a spreadsheet inherited from a former colleague. Choose a data layer when the question recurs, when more than one person asks it, or when the answer goes in front of someone who will compare it with last month. A useful rule of thumb: if you would be embarrassed by two different answers to the same question, the definition should not live in a prompt.

It also helps to separate what you want from each tool. You want the assistant for explanation, drafting and follow-up. You want the data layer for counting. Mixing the two jobs in one chat is where most inconsistent CRM numbers come from.

Honest limits

• The assistant can still misread a question or summarize badly. Governed numbers make the figures consistent and the work visible, which helps you catch errors. It does not make the narrative infallible.
• MCP support differs by assistant, plan and client, and changes quickly. Confirm that yours supports remote MCP servers before planning around it.
• Data is as fresh as the last daily refresh.
• Hygiene still matters. Blank amounts or unassigned owners show up as gaps, whichever assistant asks.

For HubSpot-specific question lists see HubSpot AI: Ask Your CRM Questions in Plain English, and for source analysis see the HubSpot attribution report. The CRM workflows are at Quiriz for CRM teams.

Frequently asked questions