How to Connect Claude to Salesforce and Gong Deal Data with MCP
You connected Claude to Salesforce, asked which commit deals had no economic buyer, and got a different list when you asked again the next morning.
Two answers to the same question
Say you're the RevOps engineer who wired it up. Claude can see opportunities, contacts and activity through an MCP server, and you gave it Gong transcripts as well. The first answer listed six deals. The second listed eight, including two the first run had called fully qualified.
Nothing in the CRM changed overnight. Claude read the same notes and transcripts, reasoned through them from scratch, and landed somewhere slightly different. For a lookup that wouldn't matter much. For a commit call it matters a lot.
What MCP actually connects
The Model Context Protocol is an open standard Anthropic released in November 2024 for connecting AI assistants to outside systems. An MCP server exposes a set of tools, such as search opportunities or get transcript, and Claude decides when to call them while it works on your request.
Connecting Claude to deal data therefore means running or installing MCP servers that can reach your CRM and your call recordings, then adding them as connectors in Claude Desktop, Claude Code or your organization's Claude settings.
Connecting Claude directly to Salesforce and Gong
Salesforce
Salesforce has announced MCP support for its platform, and open-source MCP servers for Salesforce also exist. Either way, the server authenticates as a Salesforce user, so Claude sees exactly what that user can see. Start with a dedicated integration user, keep its permissions narrow, and respect field-level security from day one.
Decide early how much Claude may write through a raw connection. Many teams hold back write permissions at first, because an agent that updates fields from its own reading of a transcript can put opinion into the CRM very quickly.
Gong
For call data, check whether your Gong plan includes an MCP connection. Where it doesn't, teams wrap the Gong API in a small MCP server of their own that returns calls by account, date or participant, with transcripts on request.
Test against a sandbox
Point both servers at a sandbox or a narrow slice of real data first. Log every tool call Claude makes during the first week, so you can see which questions trigger which lookups and how much transcript text each answer pulls in.
📊 Through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. — Gartner |
What a direct connection does well
Quite a lot. With live CRM and call access, Claude can pull up a deal before a meeting, summarize the last three calls, draft a follow-up that quotes what the buyer said, and answer factual questions like when your team last spoke to the CFO. These are lookups and drafting, and Claude handles them well.
It saves real prep time too. A rep who used to open five tabs before a call can ask one question.
RevOps gets something out of it too: ad hoc questions about pipeline counts, stage history or meeting activity no longer wait for a report request.
📊 60% of revenue and ops leaders worry about AI agents acting on the wrong records. — LeanData, 2026 State of AI GTM Readiness Report (157 revenue and ops leaders) |
Where judgment questions go sideways
Trouble starts when the question needs a judgment. Is this deal qualified? Who is the real champion? Which commit deals are at risk? Each one asks Claude to read raw fields and transcripts, decide what counts as evidence, and reach a conclusion, fresh on every request.
Answers can vary between runs, and rep notes sit beside buyer statements with nothing to tell them apart, so an assertion can weigh as much as a confirmation. Cost is the quieter issue. Every judgment re-reads transcripts, so the token bill grows with your call volume.
Picture a deal where the rep logged a strong champion after a dinner, while that contact has spoken for four minutes across three recorded calls. If Claude leans on the notes, the champion box gets ticked. If it leans on the transcripts, it stays open. Which source wins can change from one run to the next.
What you ask Claude | Raw CRM and call data over MCP | Through Spotlight's MCP |
|---|---|---|
When did we last meet the CFO? | Answers well from activity and calendar data | Answers well, with the meeting already matched to the right deal |
Is this deal qualified? | Claude judges from notes and transcripts on each run | Returns MEDDIC status already scored on evidence |
What is missing on this deal? | Depends on what Claude notices in the text | Returns the deal's scored qualification gaps |
Who matters on the buying side? | Inferred from who spoke most on recent calls | The Deal Inspection Skill shows the contact power map |
Update the deal after the call | Possible with write access, using Claude's reading as the input | Agents can write through Spotlight's MCP on evidence CDAM has already checked |
What changes when Claude reads a qualified deal through Spotlight
Spotlight.ai is the Enterprise Revenue World-Model, and its MCP server gives Claude a different starting point. The documented tools include getDeal, which returns a deal's scores, MEDDIC qualification status and pipeline stage, and searchMeetings, which finds meetings by date range, deal, title or participants, with optional summaries and transcripts. The server also writes, so agents can act on deals through it.
Behind those tools sits Spotlight's Calibrated-Decision AI-Model (CDAM). Each qualification question is answered inside a closed answer space, stored as a typed value with its source, and weighted by your win/loss history and industry benchmark. The deal score is computed in code from those values, so when Claude asks about the same deal twice it gets the same answer twice, and the decision layer produces zero output tokens.
The Spotlight Deal Inspection Skill renders all of it inside Claude as an interactive widget: deal details, qualification scores, gaps, recent meetings and the contact power map in one view. It fires on requests like inspect this deal or what's the MEDDIC status.
One connection serves the whole revenue team. Sales reps prep calls from a deal already qualified on evidence. RevOps engineers build agents on complete, evidence-backed deal data. Value teams pull the buyer's own metrics into business cases, and Spotlight has already generated 5K sales assets autonomously. Customer success inspects a renewal the same way the original deal was inspected.
The Spotlight MCP plug-in is a paid package that sits alongside the platform.
Setup checklist for Claude and deal data
Use a dedicated integration user. Scope it to the objects and fields Claude needs.
Sort your questions into lookups and judgments. Raw connections handle prep and drafting well; route qualification and risk questions to scored data.
Ask each judgment question twice. Different answers with no new evidence mean the variance is coming from the setup.
Watch transcript volume. Log how much text each answer reads in the first week, since that is where the token cost sits.
Agree which fields agents may update. Require evidence behind every change an agent writes.
Keep a readable trail for reps. A rep should see why a field changed and what the buyer said that supports it.
List the ten questions your team asks Claude most
Write down the ten deal questions your team puts to Claude most often. Mark each one as a lookup, meaning a fact that lives in a field or a transcript, or a judgment, meaning a conclusion about the deal. The lookups can stay on your direct connection.
The judgments are the list to bring to a Spotlight.ai demo. We can run them through Spotlight's MCP and the Deal Inspection Skill on your own deals, so you see what Claude returns when the deal arrives already qualified.
FAQs About Connecting Claude to Salesforce and Gong
Can Claude connect to Salesforce?
Yes. Through an MCP server that authenticates as a Salesforce user, Claude can query opportunities, contacts and activity, limited to what that user is permitted to see.
How do I connect Claude to Gong call data?
Check whether your Gong plan includes an MCP connection. If not, a small custom MCP server that wraps the Gong API can return calls and transcripts to Claude.
What is an MCP server?
An MCP server implements the Model Context Protocol, the open standard Anthropic released in November 2024. It exposes tools that an AI assistant like Claude can call to read from or act on an outside system.
Why does Claude give different answers about the same deal?
When Claude judges qualification from raw notes and transcripts, it reasons from scratch on each request and can weigh the text differently. Scored deal data computed from stored values returns the same answer for the same evidence.
What does Spotlight's MCP server do?
It gives Claude and other agents access to deals Spotlight has already qualified and scored. Documented tools include getDeal and searchMeetings, and agents can also write to deals through it.
What is the Spotlight Deal Inspection Skill?
It is a Claude agent skill that renders a deal as an interactive widget inside Claude, showing deal details, qualification scores, gaps, recent meetings and the contact power map.
Comments