Salesforce Moves Away From LLM Use Over Enterprise AI Reliability Concerns

Salesforce is changing how it uses artificial intelligence, signaling a notable shift in enterprise AI strategy. While many companies double down on large language models, Salesforce now takes a more cautious and pragmatic path.

According to NewsBytes, Salesforce has reduced its reliance on large language models (LLMs) after encountering reliability and trust issues in real-world enterprise environments. Company executives say confidence in generative AI declined over the past year, especially when deployed at scale with paying customers.

Why Salesforce is prioritizing deterministic AI

Instead of open-ended generative workflows, Salesforce is refocusing its flagship AI product, Agentforce, on rule-based and deterministic automation. This approach favors predictable outcomes over flexible but inconsistent responses.

Salesforce executives argue that enterprise customers value reliability more than creativity. When AI handles customer support, billing, or workflow automation, even small errors can turn into major financial and reputational risks.

AI-driven layoffs revealed both gains and limits

Salesforce CEO Marc Benioff recently disclosed that support staff dropped from roughly 9,000 employees to about 5,000 after deploying AI agents. Automation delivered clear efficiency gains, but it also exposed weaknesses.

The company found that AI agents struggled with complex or nuanced instructions. In enterprise settings, these failures can directly impact revenue and customer satisfaction.

Real-world failures pushed Salesforce to adjust

One notable issue involved Vivint, where AI systems failed to consistently send customer satisfaction surveys. Salesforce replaced generative logic with deterministic triggers to guarantee execution.

Internally, Salesforce also observed AI “drift,” where chatbots gradually lost focus during conversations. This behavior reduced consistency and undermined trust in automated systems, especially during long or complex interactions.

Salesforce’s move reflects a wider reassessment across the enterprise software industry. Companies still want automation benefits, but they increasingly demand systems that behave consistently and remain auditable.

Meanwhile, Microsoft continues to expand its AI footprint. Microsoft Copilot now supports real-time text editing, and recent updates added Ask Copilot to the Windows taskbar for business users, signaling deeper integration across the ecosystem.

As Salesforce pulls back from heavy LLM dependence while Microsoft accelerates Copilot adoption, a key question remains: will other enterprise software companies follow Salesforce’s more cautious approach, or will they continue to push forward with large language models despite their growing pains?