The figure is not an experimental budget — it reflects an AI deployment already generating measurable returns. Salesforce announced in January 2025 that it would freeze all new software engineering hires, citing AI-driven productivity gains exceeding 30%. By mid-2025, Benioff confirmed AI was handling 30–50% of the company's total workload. By September 2025, Salesforce had reduced its customer support headcount from 9,000 to 5,000, replacing those roles with AI agents.

Against this backdrop, a $300 million token spend on Anthropic's Claude models is less a speculative investment and more a reflection of how central AI has become to Salesforce's core business operations — particularly across coding, engineering, customer service, distribution, and marketing.

Benioff described a fundamental shift in how Salesforce now operates: a three-layer model combining humans, AI agents, and headless platforms all interoperating simultaneously. This includes the company's own Agentforce platform — which has already surpassed $800 million in annual recurring revenue and processed over 3.2 trillion tokens — alongside third-party tools including Anthropic's Claude, OpenAI Codex, and Cursor.

Salesforce also recently launched Headless 360, an API-first platform with over 60 MCP tools giving agents like Claude Code direct access to Salesforce's enterprise stack. Engineers — around 15,000 in total — have not been replaced, but their roles are evolving: they now oversee and guide AI agents rather than writing all code manually. The productivity unlock has been substantial, with Benioff noting the ability to ship and sell software simultaneously for the first time in the company's history.

Not every AI task requires a frontier model. Benioff articulated a cost optimisation thesis centred on what he described as an "intermediary layer" — an intelligent routing system that directs complex tasks to powerful models like Anthropic's Claude, while simpler, higher-volume tasks are handled by smaller, cheaper models or open-source alternatives.

This tiered approach reflects a broader maturation in enterprise AI strategy — moving beyond blanket frontier model adoption toward precision deployment that balances capability and cost. Benioff also teased forthcoming AI-powered development functionality within Slack, the messaging platform Salesforce acquired for $27.7 billion in 2021, signalling the next frontier of the company's AI integration roadmap.

Key Takeaways

• →Salesforce CEO Marc Benioff revealed the company expects to spend approximately $300 million on Anthropic AI tokens in 2026, primarily to power AI coding agents and engineering workflows.
• →AI now accounts for 30–50% of Salesforce's overall workload, with the company having frozen software engineering hiring since early 2025 and cut 4,000 customer support roles replaced by AI agents.
• →Salesforce operates a three-layer architecture combining humans, AI agents, and headless platforms, with its Agentforce division surpassing $800M ARR and 3.2 trillion tokens processed.
• →Benioff advocated for an intelligent token routing layer — directing complex tasks to frontier models like Claude and simpler tasks to smaller, cheaper alternatives — to manage AI costs at enterprise scale.
• →Anthropic's annualised revenue surged from $1 billion in early 2025 to over $30 billion by April 2026, overtaking OpenAI in US business adoption to claim over 65% of combined enterprise AI subscription spend.
• →Salesforce is building AI-powered coding capabilities directly into Slack, with Benioff teasing upcoming features that signal the next phase of the company's enterprise AI integration strategy.