Salesforce pulls back from large language models after reliability concerns

Salesforce

Salesforce, one of the world’s biggest enterprise software companies, is quietly rethinking how much it should rely on large language models, after running into reliability issues that have dented internal confidence in generative AI.

According to a report by The Information, Salesforce executives have become more cautious over the past year as AI models showed unpredictable behaviour in real-world business use. Sanjna Parulekar, Senior Vice President of Product Marketing at Salesforce, summed it up bluntly. “All of us were more confident about large language models a year ago,” she said, acknowledging that trust in these systems has slipped.

As a result, Salesforce is now shifting focus away from open-ended generative AI and toward what it calls “deterministic” automation in its flagship AI product, Agentforce. The idea is to reduce randomness and make outcomes more predictable, especially for tasks that demand accuracy and consistency.

This change in approach comes alongside major internal changes. CEO Marc Benioff recently revealed on a podcast that Salesforce cut its support staff from around 9,000 to 5,000 employees after deploying AI agents. While automation helped reduce headcount, the limits of AI became clear once customers started using these systems at scale.

Executives say the models struggle when instructions become too complex. Muralidhar Krishnaprasad, the CTO of Agentforce, noted that when AI agents receive more than eight instructions, they sometimes start skipping steps. For enterprise customers, that kind of failure can be costly.

One example came from Vivint, a home security firm that uses Agentforce to handle customer support for millions of users. Despite clear instructions to send a satisfaction survey after every interaction, surveys were sometimes missed with no obvious explanation. To fix this, Vivint and Salesforce added deterministic triggers to ensure surveys were sent every time.

Salesforce has also flagged another issue known internally as AI “drift”, where chatbots lose focus when users ask unrelated questions. A bot meant to guide users through a form, for instance, can get sidetracked instead of completing its main task.

The shift is a notable moment for Benioff, who has been one of the loudest advocates for AI-led transformation. He has since stressed that strong data foundations matter more than flashy models, warning that AI can hallucinate without the right context.

Salesforce shares are down about 34 percent from their December 2024 peak, even as Agentforce is expected to bring in over $500 million annually. For now, the company appears to be choosing reliability over hype, as it tries to balance AI ambition with the realities of running mission-critical business software.