“I’ve reduced it from 9,000 heads to about 5,000, because I need less heads.”
Marc Benioff, Salesforce’s chief executive, said it in an interview on The Logan Bartlett Show, a podcast published at the end of August 2025. He was talking about customer support. In a statement a few days later, the company gave the credit to Agentforce, its own AI agents and the product it sells to other businesses.
Salesforce had numbers behind the bet. It put Agentforce on its help site in early 2025, and by the company’s later account, within a few months the agents were handling 2.6 million customer conversations, resolving 63 percent of them, with satisfaction scores that matched human staff.
About six weeks after the podcast, Salesforce’s developer blog introduced something new for those agents. Agent Script is a scripting language for pinning down what an agent does. Large language models are good at reading human language, the post explained, but “many essential tasks, like order refunds or loan applications, must follow strict, sequential, and predictable steps.” It spoke to developers whose agents might be failing in production. With Agent Script, a developer writes the conditions, rules and handoffs the agent must follow exactly, and leaves the rest to the model. Salesforce called the mix hybrid reasoning.
In December, The Information reported that Salesforce executives’ trust in large language models had declined. According to The Decoder’s summary of that report, Sanjna Parulekar, a senior vice president of product marketing, said: “We all had more confidence in LLMs a year ago.” She pointed to the models’ randomness and their tendency to ignore specific instructions. The Information reported that Salesforce was moving Agentforce toward simple rule-based automation and limiting generative AI in certain uses. A spokesperson denied that the company was backing away from LLMs, saying it was simply being more intentional about how it used them.
What happened to the 4,000 people depends on who is telling it. CNBC’s headline called them layoffs. Salesforce’s statement at the time said that support cases had declined and “we no longer need to actively backfill support engineer roles.” In the same CNBC report, the analyst Ed Zitron argued that tech companies were blaming AI for cuts after hiring too many people during the pandemic.
In April 2026, a company post said Salesforce had redeployed “hundreds” of support engineers into fast-growing parts of the business, while not backfilling other roles. In its telling, people trained during work hours and were already certified when new jobs opened. One engineer, moved from support into contract renewals, was uncomfortable at first. None of these accounts says how many of the 4,000 left and how many moved.
The agents kept selling. By The Decoder’s account of the December report, Agentforce was on track for more than $500 million in annual sales, and it let customers set deterministic rules that strictly constrain what the AI can do.