In a major strategic reversal, Salesforce, a global leader in enterprise software, is reducing its heavy dependence on large language models (LLMs). This move follows significant reliability problems that have eroded executive confidence in the technology's current state.
From AI Enthusiasm to Cautious Retreat
Sanjna Parulekar, Senior Vice President of Product Marketing at Salesforce, openly admitted that trust in AI models has waned over the past year. "All of us were more confident about large language models a year ago," Parulekar stated, highlighting the company's new direction. Instead of generative AI, Salesforce is now focusing on more predictable "deterministic" automation for its flagship product, Agentforce.
This pivot is particularly notable given the company's previous aggressive AI push. CEO Marc Benioff had earlier revealed that AI agent deployment allowed Salesforce to cut its support staff from 9,000 to 5,000 employees—a reduction of roughly 4,000 roles. Now, the messaging emphasizes that Agentforce can help "eliminate the inherent randomness of large models."
Real-World Failures Expose LLM Limitations
The shift in strategy was driven by concrete technical challenges encountered during customer deployments. Muralidhar Krishnaprasad, Chief Technology Officer of Agentforce, identified a critical flaw: when given more than eight instructions, the models began omitting directives. This is a severe issue for business processes that require precision.
A case study involving home security provider Vivint, which uses Agentforce to manage support for 2.5 million customers, underscored these problems. Despite clear programming to send customer satisfaction surveys after each interaction, Agentforce sometimes failed to do so for no apparent reason. To fix this, Vivint and Salesforce had to implement "deterministic triggers" to guarantee consistent survey delivery.
Another issue, described by executive Phil Mui as AI "drift," occurs when AI agents lose focus on their primary task if users ask irrelevant questions. For example, a chatbot designed to assist with form completion could become distracted and ineffective.
Benioff's Vision Meets Market Realities
The retreat from LLMs marks an ironic turn for CEO Marc Benioff, a vocal proponent of AI transformation. In a recent interview, Benioff indicated that the company's annual strategic document now prioritizes data foundations over AI models, explicitly citing concerns about "hallucinations" when models lack proper data context.
Benioff had even floated the idea of rebranding the entire company as "Agentforce," noting that customer focus groups showed dwindling interest in cloud computing messaging. However, this rebranding enthusiasm now clashes with the publicly acknowledged technical hurdles.
Financially, Salesforce's stock has fallen about 34% from its peak in December 2024, even as Agentforce is projected to bring in over $500 million in annual revenue. The company's partial step back from large models could send ripples across the thousands of enterprises that depend on its technology, highlighting the ongoing struggle to bridge the gap between AI innovation and reliable, practical business implementation.
