Salesforce co-founder Parker Harris asked a fairly extraordinary question earlier this year: “Why should you ever log into Salesforce again?” He helped build the interface people have spent decades using, but in April Salesforce introduced Headless 360 as a way for AI agents to reach Salesforce through APIs, Model Context Protocol tools and command-line interfaces instead of navigating the software like a person would. Claudeforce gives that approach a much more familiar front end: Claude.
Salesforce and Anthropic announced Claudeforce on 26 August, beginning with Salesforce in Claude, a plugin containing 37 prebuilt sales skills. A salesperson can use Claude to prepare for meetings, review deals and work with live pipeline information, while actions such as updating Salesforce records are routed back through Salesforce. The companies say this allows existing business rules and permissions to remain in force even though the employee is working from Claude rather than Salesforce. Salesforce in Claude is currently available to selected pilot customers, with an open beta expected in September 2026.
Salesforce is effectively separating the interface from the system behind it. For most of the company’s history those things came packaged together: Salesforce stored the customer information and gave employees the screens through which they could work with it. Agentic AI allows the interface to move somewhere else while the records, company processes and controls remain where they are. Losing some ownership of the screen doesn’t necessarily weaken Salesforce if Claude still needs Salesforce every time it wants to retrieve the right company information or do something with it.
Salesforce has been building towards this
Headless 360 made the direction explicit before Claudeforce arrived. Salesforce has been turning functions previously buried inside individual applications into capabilities an authorised AI agent can call directly, while retaining the identity, permissions and governance already configured inside the organisation. The company describes the architecture in unusually direct terms: everything on Salesforce can increasingly be exposed as an API, MCP tool or CLI command.
A sales rep asking Claude to update an opportunity therefore doesn’t require Claude to imitate somebody navigating through the Salesforce interface. Claude can call the relevant capability and Salesforce handles what happens inside the organisation’s existing configuration. This is closer to infrastructure than the software-as-a-destination model that made Salesforce one of the biggest enterprise software companies in the world, and it gives the company somewhere to go if people increasingly start their work inside general AI assistants instead.
That possibility is uncomfortable for conventional enterprise software because an employee who can ask Claude for a customer briefing, tell it to change a sales forecast and have it prepare the follow-up work may have fewer reasons to visit several specialised applications along the way. Salesforce doesn’t appear to be trying to stop that behaviour. It’s arranging things so that even when the employee doesn’t visit Salesforce, the agent still does.
Claude gets the context while Salesforce stays in the path
Anthropic benefits because enterprise AI becomes much more useful when Claude can reach the information companies have accumulated over years rather than relying primarily on whatever somebody places in a prompt. Salesforce says its Claude integration can read a seller’s enterprise context, including Salesforce and Slack information available to that user, and use it to generate views around accounts, pipeline information and other live business data.
Salesforce keeps control over the point at which Claude’s reasoning becomes an action inside the business. Claude might conclude that an opportunity needs updating, for example, but Salesforce says the change still passes through the organisation’s existing rules and permissions. An administrator can connect Salesforce in Claude centrally instead of constructing another permissions system around every salesperson, which could become increasingly useful if companies end up running several different agents against the same underlying systems.
The Anthropic relationship also predates Claudeforce. Salesforce customers gained access to Claude models through Amazon Bedrock in 2024, initially allowing Claude to be used inside Salesforce applications while interactions continued to pass through Salesforce’s AI security systems. The companies have steadily moved beyond putting Claude inside Salesforce towards allowing Salesforce itself to appear inside Claude.
Slack has been moving in the same direction. Anthropic introduced Claude Tag in June as a way for teams to bring Claude directly into Slack channels, where it can build context from selected conversations and use connected tools rather than requiring everyone to move into a separate Claude window. Under Claudeforce, Salesforce says Claude is now the default model for Slack and powers Slackbot, while Claude is also used across Salesforce in Claude, Headless 360 and other Salesforce AI products.
The relationship now runs heavily in both directions. Salesforce plans to make Claude Code and Claude Enterprise available to its developers and knowledge workers, while Anthropic uses Salesforce as its preferred CRM and Slack as its preferred internal work platform. Salesforce says Slackbot is generating 8.1 million annualised hours of productivity gains internally, although that remains Salesforce’s own measurement rather than an independently verified assessment of what those saved hours are ultimately worth.
Salesforce is already changing what it measures
The strategy arrives while investors are still trying to work out what AI agents do to the economics of traditional software. Much of enterprise SaaS has been sold according to how many people need access to an application, but agents create the possibility that fewer humans will need to interact directly with some of those applications even if more work ultimately passes through the systems underneath them.
Salesforce’s existing business isn’t showing signs of disappearing. In its second-quarter results for the 2027 financial year, the company reported revenue of $11.3 billion, up 11% year on year. Annual recurring revenue from Agentforce and Data 360 reached nearly $3.9 billion, while Agentforce ARR exceeded $1.5 billion.
How Salesforce measures its newer AI business is already changing, though. Agentforce ARR now includes Slackbot and Headless 360, while Salesforce has introduced what it calls Agentic Work Units to count tasks completed through its agents. The company says seven billion of those units have been delivered across Agentforce and Slack so far, including 3.2 billion during the most recent quarter.
Those figures don’t mean Salesforce is about to abandon conventional per-user licences, and the company hasn’t said that it will. They do give Salesforce ways of measuring activity that aren’t exclusively tied to how many people sit in front of a Salesforce application. If AI changes the relationship between employees and software, the company is already creating commercial metrics for work that happens after the interface becomes less important.
Giving agents better access doesn’t fix what they’re accessing
Allowing an AI agent to operate directly against company systems makes the quality of those systems harder to ignore. Salesforce can make Claude inherit an organisation’s permissions and workflow rules, but it can’t guarantee that those permissions still make sense, that the underlying records are accurate or that a decade of customised processes remains sensible simply because an AI agent can now use them.
South African-founded Verascient is building its business around a closely related problem. After raising R19.5 million in pre-seed funding, the company is trying to give enterprise AI access to institutional knowledge while retaining information about provenance and permissions. Its argument is that capable models don’t automatically understand how a company actually works when useful knowledge is scattered across documents, old systems, employee accounts and years of exceptions.
Governance becomes harder once those models are allowed to act as well as retrieve information. TrendAI research found that only 38% of organisations it surveyed had comprehensive AI policies, even as companies were expanding their use of AI agents. Forty-four percent of respondents identified agents accessing sensitive information as their biggest agentic AI risk, while organisations were also trying to decide which actions should still require human approval.
Claudeforce doesn’t remove those weaknesses. It gives an agent a governed route into whatever a company already has, which can be valuable without making the underlying environment good. A poorly designed approval process remains poorly designed when Claude can trigger it faster, while old access permissions become more consequential once software is capable of acting with them rather than simply displaying information to a person.
Salesforce’s approach puts more responsibility on decisions made before the agent starts working. A company has to decide what Claude can reach, which rules should constrain it and where human intervention is still required. That’s a different problem from simply giving employees access to an AI chatbot, because mistakes in permissions or business logic can now shape what an automated system is able to do.
South African customers still need the practical details
Salesforce says Claude can operate through Amazon Bedrock inside the Salesforce Trust Boundary, which is particularly relevant to companies working with sensitive data or operating in regulated industries. The Claudeforce announcement also says availability will vary by region and customer agreement, and Salesforce hasn’t announced a specific South African rollout for Salesforce in Claude.
For local companies, Claudeforce currently gives a clearer indication of Salesforce’s platform strategy than it does of something every Salesforce customer can immediately deploy. The September open beta should provide a better idea of how broadly Salesforce in Claude will be available, how it will be packaged and what administration is actually required once companies begin connecting real sales teams and production data to it.
Salesforce has spent more than 25 years persuading companies to put customer information and increasingly complicated business processes onto its platform. Claudeforce assumes that history remains valuable even if employees gradually stop thinking about Salesforce as somewhere they go to work. An employee may eventually spend most of the day asking Claude questions and giving it instructions while rarely seeing a Salesforce screen, but Claude still needs somewhere to get the customer record, check what the employee is allowed to do and execute the resulting action. Salesforce is trying to make sure that somewhere is still Salesforce.