For much of the past two years, enterprise software companies have projected near total confidence in artificial intelligence systems, especially large language models, as the next foundation of business operations. Salesforce was one of the loudest voices in that shift. The company positioned AI agents as tools capable of handling customer service, automation, and decision workflows at scale.
That confidence is now being reassessed.
Salesforce executives have acknowledged that expectations around large language models were higher a year ago than reality has proven, as the company quietly pivots toward more controlled and predictable automation frameworks. The change comes after Salesforce reduced thousands of support roles and expanded the use of AI agents across its enterprise products.
Sanjna Parulekar, Senior Vice President of Product Marketing at Salesforce, recently admitted that trust in large language models has weakened inside the company. Speaking about internal reassessments, she said executives were far more confident in the technology last year than they are today. The admission reflects a broader industry moment where enthusiasm for generative AI is meeting operational limits.
Salesforce shifts Agentforce toward deterministic automation
Salesforce’s change in direction is most visible in Agentforce, the company’s flagship AI agent platform. Instead of relying heavily on open ended language generation, Salesforce is now emphasizing deterministic automation. The goal is to reduce unpredictable behavior that can disrupt enterprise workflows.
Marc Benioff, Salesforce’s CEO, previously revealed the operational impact of AI agents during a podcast appearance. He said the company reduced its support organization from around 9,000 employees to roughly 5,000, citing that fewer human roles were needed as AI systems took over routine tasks. That reduction amounts to about 4,000 positions.
However, real world use exposed limitations. According to Salesforce executives, large language models begin to struggle when tasks become complex. Muralidhar Krishnaprasad, Chief Technology Officer for Agentforce, explained that when AI agents are given more than eight instructions, they can start skipping steps. For enterprises that depend on accuracy and compliance, missed instructions are not a minor issue.
These concerns surfaced publicly through customer experience. Vivint, a home security company serving about 2.5 million customers, uses Agentforce for customer support. Despite clear instructions to send satisfaction surveys after every interaction, surveys were sometimes not delivered. There was no clear trigger failure or system alert explaining why.
Vivint worked directly with Salesforce engineers to correct the issue. The solution was not more AI freedom, but tighter control. Deterministic triggers were introduced to ensure surveys were sent consistently, reducing reliance on AI judgment.
Another challenge discussed internally is what Salesforce executives refer to as AI drift. Phil Mui, a senior Salesforce executive, described how AI agents can lose focus when users ask unrelated questions. A chatbot designed to help complete a form may abandon its task entirely if the conversation strays, creating confusion rather than efficiency.
Salesforce AI strategy meets market and investor reality
The shift away from heavy language model dependence represents a turning point for Salesforce leadership. Benioff has been one of the most vocal advocates for AI driven enterprise transformation. He has frequently spoken about AI reshaping software companies from the inside out.
More recently, however, Benioff acknowledged the limits of AI without strong data foundations. In an interview with Business Insider, he said the company’s strategic priorities now place data infrastructure above AI models themselves. He cited hallucinations and unreliable outputs as key risks when models operate without strict data context and business logic.
Benioff even floated the idea of rebranding Salesforce around Agentforce, noting that customer focus groups showed less interest in traditional cloud messaging. While the comment highlighted ongoing ambition, it contrasted with the technical restraint executives are now publicly acknowledging.
Salesforce stock has fallen roughly 34 percent from its December 2024 peak, reflecting broader tech market pressures and investor caution around AI monetization. Despite that decline, Salesforce projects Agentforce could generate more than $500 million in annual revenue, signaling that AI remains central to the company’s future, just in a more controlled form.
In a statement responding to questions around the shift, a Salesforce spokesperson emphasized that large language models alone cannot run enterprise operations. The company said AI must be grounded in accurate data, governance, and deterministic frameworks to deliver reliable outcomes. Salesforce positioned Agentforce as infrastructure designed to make AI trusted, secure, and predictable rather than experimental.
For thousands of enterprises using Salesforce tools, the shift is a reminder that AI adoption is moving from hype to discipline. Salesforce is not abandoning AI. It is redefining how far autonomy can go before businesses demand certainty.
Read Also: Tech giants envision future beyond smartphones as post-screen era takes shape
