Salesforce and Anthropic announced Claudeforce on 26 August 2026. The first product from the partnership is Salesforce in Claude, a plugin carrying thirty seven prebuilt sales skills, available to selected pilot customers now with an open beta planned for September. A seller asks a question in Claude and the answer comes back from live pipeline, live accounts, live activity, without opening Lightning at all.
The connection runs over Headless 360, the Salesforce architecture that exposes data, workflows, business logic and permissions as capabilities an external system can call through MCP. Salesforce describes it as requiring no custom integration work. An administrator connects the plugin once and the whole team has it.
Two things travel with that connection. Both were configured years before anyone considered pointing a model at them.
The sharing model becomes the reading model
Salesforce in Claude carries no permissions of its own. Every answer and every action runs through the permissions and business rules already sitting in the org. Claude sees what the person driving it is authorised to see, and can do what that person is authorised to do. Nothing new to stand up, nothing to re-audit, nothing to configure account by account.
That removes the largest objection to putting a model near a system of record, and it introduces a quieter consequence. Sharing rules in most orgs were not written as a statement of principle. They accumulated. A role hierarchy opened up so the national sales manager could see everything, an org wide default relaxed so a report would run, a permission set granted for a project that finished in 2023. None of that mattered much while reaching the data required knowing which report to open and how to filter it. Effort was doing quiet work as a control.
Whatever the sharing rules already permit, they will now permit at the speed of a sentence.
On the write side the controls are specific and adjustable. Claude can be set to confirm before it emails anyone outside the company, or left to send once that feels comfortable. When it updates a record it changes only the field it said it would change. Salesforce states zero data retention on Sonnet, Opus and Haiku, and describes further controls over where data is stored and who can reach it as still being built with Anthropic.
A skill is a file
The thirty seven skills arrive with their names visible. Daily briefing, pipeline review, forecast narrative, stakeholder map, close plan, handle objection, account plan, lead triage, customer health, call prep, log activity, win loss review, hygiene check. Salesforce built them with its own engineers and its own account executives, piloted against its own pipeline, and calls itself customer zero on the product.
That origin is legible in the set. These skills describe a sales motion with stages, a close date and a signature at the end of it. Plenty of revenue does not arrive that way. Money that lands through a platform, an approved list, a ratings decision or a relationship measured in years has a shape that no close plan quite fits, and the skills that assume otherwise will be the ones that quietly go unused.
The interesting part is that a skill is written instruction rather than compiled product. Salesforce has said the same capabilities are open to any developer who wants to build their own. A firm that knows precisely how its own revenue arrives can describe that, in words, and have it work the same way the shipped skills do. The plugin sets a starting position, not a ceiling.
What is actually available
Sales first. Service, Marketing, Commerce, Revenue, Field Service, Tableau, MuleSoft, Informatica, Data 360, Headless 360 and Industries are all listed as coming soon, with no dates attached. Pilot access is by selection. The open beta is planned for September and the product will be demonstrated at Dreamforce between 15 and 17 September. The announcement makes no statement about Australian availability or data residency.
Capability has arrived ahead of the conditions for using it, which is the ordinary sequence and not a complaint. A firm that connects this will find out what its records actually contain within a week, rather than in the third year of wondering why the forecast never reconciles. Every system is eventually read back to the people who built it.
Q: Does Salesforce in Claude require Data 360 or a separate data project first?
No. The plugin reads the Salesforce org directly through Headless 360 over an MCP connection. Data 360 appears on the roadmap as a way to bring in sources beyond Salesforce. An org with populated standard objects and consistently logged activity already has what the sales skills draw on.
Q: What happens when sales activity has not been logged consistently?
The skills reason over what the record contains. A meeting that was never logged is absent from the answer, and the answer will not say so. Reasoning quality tracks record quality exactly, which tends to make logging gaps visible faster than any adoption report has managed.
Q: Does an administrator have to build anything before switching it on?
No new permissions model is required. An administrator connects Salesforce in Claude once and it applies across the team, inheriting existing permissions and business rules. The preparation worth doing is a review of what those existing sharing rules currently allow, since they now govern what can be read conversationally.
Q: Can the thirty seven skills be changed or added to?
Salesforce has said the underlying capabilities are open to any developer building their own, so the shipped set functions as a starting position. Firms whose revenue arrives through platforms, approved lists or ratings decisions rather than signed contracts will find some skills fit closely and others need describing again.