Anthropic Brings Claude Inside Salesforce With 37 Built-in Sales Skills
Anthropic's new plugin drops live Salesforce pipelines, accounts, and 37 prebuilt sales skills into Claude, letting sellers act on CRM data without leaving the chat.
- Anthropic launched Salesforce in Claude in open beta with 37 prebuilt sales skills.
- Plugin covers call prep, deal scoring, pipeline dashboards, and CRM writebacks with seller approval.
- Part of the broader Claudeforce partnership announced alongside Salesforce Q2 FY27 earnings.
- GitLab, Siemens, and Legora already deployed it to roughly 7,000 sellers.
- Runs on Salesforce's AIforce harness and MCP; permissions inherit from existing Salesforce roles.
- Available on all paid Claude plans; more skills for other teams shipping in coming months.
Anthropic has opened beta access to Salesforce in Claude, a plugin that brings Salesforce records and actions into Claude. Its 37 prebuilt sales skills can assemble account context, analyze opportunities, update pipeline data, and draft forecasts. Claude requests the seller’s approval before writing changes to Salesforce.
A seller can ask Claude to prepare for a call, review a deal, build a pipeline dashboard, or summarize the week. The plugin combines permitted data from Salesforce, Slack, and email, reducing the manual work of finding records and reconciling conversations across applications.
Claudeforce links three surfaces
Salesforce and Anthropic announced the beta as the first released product from Claudeforce, an expanded partnership unveiled alongside Salesforce’s Q2 FY27 earnings. The agreement covers three integration paths:
| Integration | Role |
|---|---|
| Salesforce in Claude | Exposes permitted Salesforce data and governed CRM actions inside Claude. |
| Claude in Agentforce | Adds Claude as a reasoning model for Salesforce’s platform for building and deploying AI agents. |
| Claude in Slack | Extends Claude’s access to conversations and workflows inside Slack. |
The partnership also includes reciprocal internal deployments. Salesforce is adopting Claude Code and Claude Enterprise for engineering work, while Anthropic uses Salesforce as its primary CRM.
Thirty-seven workflows, packaged
The initial skill library follows the recurring work of an account executive:
- Morning brief: Produces a daily digest of meetings, opportunities nearing their close dates, at-risk deals, and unread threads. On Fridays, it summarizes the week and drafts a manager update.
- Call preparation: Collects open opportunities, recent Slack discussions, unanswered messages, and unresolved questions from earlier calls. When it identifies stakeholders missing from Salesforce, it proposes contact records for the seller to approve.
- Deal review: Scores an opportunity against the team’s sales methodology, identifies qualification gaps and missing stakeholders, and drafts a business case and dated mutual close plan.
- Post-call follow-up: Converts notes or a transcript into an email, a Slack deal-channel summary, and proposed changes to next steps, stage, and close date.
- Pipeline analysis: Builds an interactive view of coverage by stage and opportunities likely to slip, then drafts a forecast narrative in the format used by sales leadership.
Salesforce permissions stay in charge
The access model reuses Salesforce identity and authorization rather than creating a separate permissions layer inside Claude. Its main controls are:
- Sellers authenticate with their Salesforce credentials.
- Claude can read only the objects, fields, and records available to that user.
- Salesforce remains the system of record for CRM data.
- Claude requests approval before committing a write by default.
- Administrators connect Salesforce at the organization level and assign access to selected groups.
MCP supplies the plumbing
Salesforce in Claude runs on AIforce, Salesforce’s enterprise integration layer for connecting business data and workflows to AI agents. It uses APIs, command-line tools, and Model Context Protocol servers. MCP is a standard that allows an AI client to discover external tools, retrieve authorized data, and invoke supported actions.
AIforce builds on Salesforce’s Headless 360 architecture, which exposes data, workflows, agents, and governance without requiring users to navigate the standard Salesforce interface. The Salesforce MCP connector is available through Claude’s marketplace. Administrators seeking the complete plugin and its skill library must request beta access through AgentExchange.
Access, data terms, and cost
| Availability | Beta access requires administrator approval through AgentExchange. |
|---|---|
| Claude plans | Anthropic says the plugin supports all paid Claude plans. |
| Model training | On Team and Enterprise plans, Anthropic does not train models on customer data by default. |
| Plugin pricing | No separate price has been announced. |
| Inference contract | Customers contract separately with Anthropic for Claude usage. |
Anthropic says GitLab, Siemens, and Legora have deployed the plugin, with about 7,000 sellers using it in production. Additional prebuilt skills are scheduled to begin arriving in late 2026. The first release concentrates on sales, with skills for other teams expected later.
Agent quality follows CRM quality
The release gives enterprise developers a concrete pattern for placing an AI agent over operational systems: inherit the source application’s permissions, gather context across connected services, package repeatable workflows as skills, and require human approval at write boundaries.
The architecture also exposes predictable failure modes. Stale records, inconsistent stage definitions, missing contacts, and overly broad permissions all affect the context Claude receives and the actions it proposes. A beta evaluation should therefore test record-level access, field-level restrictions, write approvals, stale data, cross-application identity matching, and recovery when one connected service is unavailable.
As CRM work moves from forms and list views into conversational workflows, schemas and governance carry more weight. Field definitions, permission scopes, and data completeness directly determine whether the agent produces a useful account summary, a defensible forecast, or an incorrect update.
Comments
