Salesforce Put Its Permission Model Behind Claude
One admin click, 37 skills, and a governance layer that decides what an AI agent is allowed to do. If your product was a wrapper around CRM data, the wrapper just shipped.
One admin click, 37 skills, and a governance layer that decides what an AI agent is allowed to do. If your product was a wrapper around CRM data, the wrapper just shipped. What happened Salesforce and Anthropic announced Claudeforce , an expanded strategic partnership that exposes Salesforce data, workflows, business logic, actions and governance to Claude (Salesforce newsroom). The first product is Salesforce in Claude , a Plugin shipping with 37 prebuilt sales skills — meeting prep, deal health review and pipeline review are named; the other 34 are not. The integration runs over MCP, Anthropic's Model Context Protocol, with Salesforce's AIforce harness enforcing business rules and permissions on every request (Salesforce newsroom). The authentication model is the part worth reading twice. Salesforce states: "An admin connects Salesforce in Claude a single time, with authentication and permissions managed centrally, and every seller on the team gets access from day one — no per-user setup, no new permissions model to build, no re-auditing account by account" (Salesforce newsroom). And on visibility: "Claude sees what the user is authorized to see, and can do what that user is authorized to do, no more." It is live with select pilot customers now; Salesforce expects to launch open beta in September 2026, with additional skills from Q3 2026 (investor release says "late 2026"). Salesforce's product page also claims zero data retention on Sonnet, Opus and Haiku — that is a vendor claim on a marketing page, not in the press release, and not independently verified. Pricing is not disclosed. Why it matters The direction of the integration is the news. For two years the pattern was: SaaS product hosts a chat widget, widget calls a model. Claudeforce inverts it. The assistant is outside your product, and your product supplies governed capabilities to it. The user never opens the CRM. If you are a CPO , the roadmap question this quarter is no longer "what does our AI chat interface look like." It is: what governed actions do we expose, and what must the user see before one fires? If your differentiation was "we wrap an LLM around CRM data," the two vendors who own both halves just shipped that wrapper themselves — with the permission model included. If you are a CEO , this is a governance story before it is a productivity story. The board-legible version: AI actions inherit the same role-based controls as the CRM, and every action is auditable. That moves the conversation from can we let an agent touch the CRM to how fast do we roll it out . If you are a startuper selling into Salesforce shops , the per-seat onboarding tax you fought is gone — and so is any differentiation you built on top of it. Central auth means your users are provisioned the moment an admin connects the Plugin. Assume Claudeforce is present in your customers' orgs from early 2027 and plan around it. We wrote about this on 7 August , when we argued that the model itself was becoming a