Salesforce, one of the world’s most valuable enterprise software firms, is reassessing its heavy dependence on large language models (LLMs) after encountering reliability issues that have weakened executive confidence in generative AI.
According to a report by The Information, Sanjna Parulekar, Senior Vice President of Product Marketing at Salesforce, acknowledged that trust in AI models has declined over the past year. “All of us were more confident about large language models a year ago,” she said, pointing to a strategic shift toward more predictable, deterministic automation in the company’s flagship AI product, Agentforce.
This recalibration follows significant workforce reductions linked to AI deployment. Salesforce CEO Marc Benioff recently revealed that the company reduced its support staff from about 9,000 to 5,000 employees after introducing AI agents. Speaking on a podcast, Benioff said the company required fewer personnel as automation took over routine tasks.
Salesforce is now positioning Agentforce as a solution that can “eliminate the inherent randomness of large models,” marking a notable departure from the AI-first messaging that dominated the tech industry until recently.
Reliability concerns surface in real-world use
Executives admitted that large language models struggled with consistency in practical business applications. Muralidhar Krishnaprasad, Chief Technology Officer of Agentforce, noted that when AI models are given more than eight instructions, they often begin omitting steps — a serious issue for enterprise workflows that demand precision.
Such problems were observed by customers as well. Vivint, a home security company serving around 2.5 million customers, uses Agentforce for customer support. Despite clear instructions to send satisfaction surveys after every interaction, the system occasionally failed to do so. To resolve this, Salesforce and Vivint introduced deterministic triggers to ensure surveys were sent reliably.
Another challenge involved what Salesforce executives described as AI “drift.” In an October blog post, executive Phil Mui explained that AI agents can lose focus on their primary tasks when users ask unrelated questions. For example, a chatbot designed to guide form completion may become distracted if customers deviate from the intended flow.
Benioff’s AI vision meets practical limits
The company’s pullback from large language models is a notable shift for CEO Marc Benioff, who has been a strong advocate of AI-led transformation. In a recent interview with Business Insider, Benioff said Salesforce is prioritising strong data foundations over AI models in its strategic planning, citing concerns over AI “hallucinations” when systems lack proper data context.
Benioff has also floated the idea of rebranding the company as “Agentforce,” noting that customers are increasingly less interested in cloud-focused messaging. However, this ambition contrasts with the technical and trust-related challenges now being acknowledged internally.
Salesforce shares have fallen about 34% from their December 2024 peak, though Agentforce is still projected to generate more than $500 million in annual revenue. As the company recalibrates its AI strategy, its shift away from large language models could have implications for thousands of enterprises that rely on generative AI for business operations.
