Salesforce, a famous enterprise software company, is withdrawing from its heavy dependence on large language models (LLMs) after facing reliability issues that the executive didn't like. The company believes that trust in AI LLMs has declined in the past year, according to The Information.  Parulekar, senior VP of product marketing said, “All of us were more confident about large language models a year ago.” This means the company has shifted away from GenAI towards more “deterministic” automation in its flagship product Agentforce. In its official statement, the company said, “While LLMs are amazing, they can’t run your business by themselves. Companies need to connect AI to accurate data, business logic, and governance to turn the raw intelligence that LLMs provide into trusted, predictable outcomes.” Salesforce cut down its staff from 9,000 to 5,000 employees due to AI agent deployment. The company emphasizes that Agentforce can help "eliminate the inherent randomness of large models.”  Failing models, missing surveys Salesforce experienced various technical issues with LLMs during real-world applications. According to CTO Muralidhar Krishnaprasad, when given more than eight prompts, the LLMs started missing commands. This was a serious flaw for precision-dependent tasks.  Home security company Vivint used Agentforce for handling its customer support for 2.5 million customers and faced reliability issues. Even after giving clear instructions to send satisfaction surveys after each customer conversation, Agentforce sometimes failed to send surveys for unknown reasons.  Another challenge was the AI drift, according to executive Phil Mui. This happens when users ask irrelevant questions causing AI agents to lose focus on their main goals.  AI expectations vs reality hit Salesforce  The withdrawal from LLMs shows an ironic twist for CEO Marc Benioff, who often advocates for AI transformation. In his conversation with Business Insider, Benioff talked about drafting the company's annually strategic document, prioritizing data foundations, not AI models due to “hallucinations” issues. He also suggests rebranding the company as Agentforce.  Although Agentforce is expected to earn over $500 million in sales annually, the company's stock has dropped about 34% from its peak in December 2024. Thousands of businesses that presently rely on this technology may be impacted by Salesforce's partial pullback from large models as the company attempts to bridge the gap between AI innovation and useful business application.

Origin

Discover how Salesforce's new Agentforce Sales App is transforming CRM with cutting-edge AI integration, enhancing sales processes, and empowering teams to achieve remarkable results.

Build RAG-powered agents with Salesforce Data Cloud and Agentforce by using web crawlers to keep AI updated with secure, real-time data.

Salesforce moves agentic AI deeper into marketing, aiming to turn website engagement into autonomous pipeline generation. Continue reading...

Salesforce, a titan in the customer relationship management (CRM) space, is undergoing a significant recalibration of its enterprise Artificial Intelligence (AI) strategy, particularly concerning its Agentforce offering. This strategic pivot, while aimed at making AI agents viable in production, introduces substantial new costs and complexities for Chief Information Officers (CIOs) and their organizations. Originally marketed as a self-directed, end-to-end resolution agent, Agentforce now requires deterministic controls through a new scripting layer called Agent Script, shifting critical responsibilities and financial burdens back to customers.

The Shift From Cloud Hype to Agent Obsession For more than a decade, the word “cloud” defined the technology industry. It was the universal promise, the safe bet, the future everyone agreed on. That era is quietly ending. In its place, a new obsession is taking over boardrooms, keynotes, and product roadmaps: AI agents. These autonomous or semi-autonomous systems are being framed as the next great productivity revolution, capable of acting, reasoning, and executing tasks with minimal human input.