Salesforce is retreating from its heavy reliance on large language models (LLMs) after executives acknowledged a significant decline in trust regarding the technology’s reliability. Despite recently laying off approximately 4,000 employees to pivot toward AI-driven automation, the enterprise software giant is now shifting its strategy toward “deterministic” automation to ensure more predictable outcomes for its clients.
Technical Failures and Declining Trust
According to Salesforce executives, the company encountered critical issues where AI agents failed to follow instructions during real-world applications. Specifically, the Chief Technology Officer of Agentforce noted that models began omitting directives when provided with more than eight instructions. A notable example involved home security company Vivint, where AI agents failed to send customer satisfaction surveys despite explicit programming. Furthermore, the company observed “AI drift,” where chatbots lost focus on primary tasks when users asked irrelevant questions.
Strategic Pivot Amidst Workforce Changes
This technical reassessment comes shortly after CEO Marc Benioff confirmed the reduction of the company’s support staff from 9,000 to 5,000, explicitly attributing the cuts to the efficiencies expected from AI agents. However, facing the reality of “hallucinations” and inconsistency, Benioff is now prioritizing data foundations over generative models. The company aims to eliminate the inherent randomness of LLMs by implementing deterministic triggers, marking a sharp contrast to the industry’s previous “AI-first” hype.
Mentoring question
Given Salesforce’s realization that AI models require more deterministic controls than anticipated, how should leaders evaluate the maturity of automation technologies before making irreversible workforce reductions?