What this module covers: The complete pattern for any write-capable Salesforce action you expose to Claude — query, confirm, act, verify — demonstrated on the connection you built in Module 4.
This walkthrough assumes the Module 4 setup is complete and working — verified as of July 2026.
The Four-Step Pattern
Every write-capable MCP workflow should follow the same shape: ask Claude to query the relevant data, let it confirm the intended action with you, watch it act, then verify the result somewhere Claude didn't render it. This is the exact pattern to reuse for any Salesforce workflow you build going forward.
1
Query First
Find a real record's Salesforce ID (format a0X...) — from Setup or the record's own URL. Ask Claude to act on it directly:
Prompt
"Approve quote demo a00fj00001kR7mTAAS"
2
Claude Confirms Before Acting
Because "Require user confirmation" is enabled on the Agent Action (Module 4, Step 3), Claude pauses and checks the intended action with you before touching the record — it does not execute the write on the first message.
Claude's response
Claude confirms the intended action and asks whether to proceed, before making any change.
3
Confirm, and Claude Executes
Reply to confirm. Claude then calls the tool and reports the result:
Claude's response
"Approved — quote a00fj00001kR7mTAAS is now set to Approved."
Claude calling the tool and confirming the result
4
Verify Outside the Chat
Open the same record in Salesforce Lightning. This is the proof step — checking the result somewhere Claude didn't render it confirms the change genuinely hit the database, not just a chat response.
Record QD-0001: Status now Approved, Amount $750,000.00, Territory West
Why the Confirmation Step Isn't Optional
Step 2 is the single most important design decision in this whole pattern. An agent that can read data and take action with no checkpoint is a liability the moment it misreads intent — approving the wrong quote, rejecting the wrong record. Every write-capable MCP tool you build should default to requiring confirmation unless you have a specific, reviewed reason not to.
Key Points
- Query, confirm, act, verify — reuse this exact pattern for any write-capable Salesforce workflow
- "Require user confirmation" on an Agent Action genuinely pauses execution for a human check, including when called from an external MCP client like Claude Desktop
- Always verify a write action somewhere the agent didn't render the result — a Lightning record, a report, a second independent query
Next Up
Module 6 covers governance and API limits in depth — how permissions travel with the agent, and the daily API math you need to watch before scaling this pattern beyond a demo.
