A New Labour Model: Salesforce to Spend US$300m on Anthropic

Salesforce had announced in 2025 that the company would halt hiring software engineers, attributing the decision to productivity improvements from AI tools in development workflows.

The company now plans to invest US$300m in Anthropic tokens during 2026, with the majority allocated to coding applications, says CEO Marc Benioff, speaking on the All-In podcast. 

Marc has clarified his position since the initial announcement. He now describes AI as a tool that transforms engineering work rather than eliminating the need for human developers.

The company maintains approximately 15,000 engineers who work alongside AI systems including Anthropic models, OpenAI Codex and Cursor.

According to Marc, the investment in Anthropic tokens could lower software development costs whilst increasing output.

He characterises AI coding agents as "awesome" and says the investment would lower software development costs while increasing output.

AI changes engineering workflows

The hiring freeze began following productivity improvements of more than 30% attributed to Agentforce and other AI technologies within Salesforce engineering teams.

Marc first disclosed these plans during a 2024 earnings call.

"We're not adding any more software engineers next year because we have increased the productivity this year with Agentforce and with other AI technology that we're using for engineering teams by more than 30%," Marc says.

He notes that engineering teams had achieved higher velocity through AI integration.

The company expanded hiring in other departments during the same period. Marc states Salesforce planned to add between 1,000 and 2,000 salespeople to explain AI products and their business value to customers.

Engineers at Salesforce now work in what Marc describes as a supervisory capacity over AI agents. However, he maintains that current AI technology cannot operate autonomously in software development roles.

"When they start to use these models, they're now working not only with the AI, but agents to help them code – and they can even become somewhat supervisory over these agents. But still, those engineers are needed. The model still cannot operate autonomously," Marc says.

Token investment supports product expansion

The US$300m allocation to Anthropic tokens accompanies broader AI integration across Salesforce's product portfolio, including Slack and Agentforce.

Marc indicates the company intends to develop routing systems that direct AI requests between larger and smaller models based on task complexity, potentially reducing operational expenses.

Salesforce reportedly holds a 1% stake in Anthropic valued at US$1bn. The AI start-up provides the Claude models that power several new capabilities in Slack.

Agentforce has reached US$800m in annual recurring revenue according to company figures. Marc describes agentic AI as "a new labour model, new productivity model and a new economic model" in a company statement.

Salesforce reports that AI accounts for between 30% and 50% of its overall workload.

The company's engineering organisation has become approximately 30% more productive following AI integration, Marc adds.

New economic models emerge

Marc characterises the shift as a fundamental change in business operations.

"Digital labour is a new horizon for business. How we architect our businesses and run our businesses and staff our businesses and think about our businesses will never be the same," Marc says.

The approach represents a growing trend among technology companies where AI serves as an efficiency tool rather than a replacement mechanism. Large organisations including Anthropic and NVIDIA have expressed similar positions, stating that AI should enhance existing workflows.

Marc acknowledges current limitations in AI capabilities. "We're not at that level yet of AI," he says, whilst noting the productivity gains achieved by the company's engineering organisation.

The AI request routing system under development could allow Salesforce to optimise costs by matching request complexity with appropriately sized models. Marc suggests this approach could reduce operational expenses whilst maintaining performance standards.