Salesforce Unveils Proprietary AI Models, Paving the Way for Agentforce
Salesforce today announced the release of new AI models, including xGen-Sales and the xLAM (Large Action Models) family, designed to power autonomous sales tasks and complex function-calling within its Agentforce platform. These proprietary models, developed by Salesforce AI Research, aim to enhance the capabilities of AI agents, allowing for more precise and rapid responses in sales-related tasks.
The xGen-Sales model, fine-tuned for industry-specific tasks, has reportedly outperformed larger models in Salesforce’s internal evaluations. Meanwhile, the xLAM family, including the open-source xLAM-1B, promises lower costs and faster performance compared to more complex models currently available.
MaryAnn Patel, SVP of Product Management at Salesforce, emphasized the significance of these developments: “We envision a future in which sellers are augmented by AI to help them drive selling efficiency, freeing up precious time to focus on their customers. The xGen-Sales model is purpose-built to help companies build generative AI solutions that will augment the work of their sales teams with Agentforce.”
Table of contents
What is xGen-Sales?
xGen-Sales is a proprietary AI model developed by Salesforce, specifically designed to power autonomous sales tasks within the Agentforce platform. This model has been fine-tuned to increase accuracy for relevant industry tasks, enabling it to deliver more precise and rapid responses in sales-related activities.
Key capabilities of xGen-Sales include generating customer insights, enriching contact lists, summarizing calls, and tracking sales pipelines. According to Salesforce, xGen-Sales has demonstrated superior performance compared to larger models in their internal evaluations.
What are Large Action Models?
Large Action Models (LAMs) represent a new generation of AI models that specialize in function-calling capabilities. Unlike traditional Large Language Models (LLMs) that primarily generate content, LAMs are designed to execute actions within systems and applications.
The xLAM family, introduced by Salesforce, includes models of various sizes, from the compact xLAM-1B to the more robust xLAM-8x22B. These models are engineered to offer lower costs, faster performance, and greater accuracy compared to larger, more complex models.
A key feature of LAMs is their ability to trigger actions needed for AI agents to independently perform tasks. This makes them crucial components in the development of autonomous AI systems.
Implications for Agentforce
The introduction of xGen-Sales and the xLAM family appears to be laying the groundwork for a major overhaul of Salesforce’s AI strategy, with Agentforce at its core. Industry rumors suggest that a significant announcement regarding Agentforce is expected at Dreamforce 2024, scheduled to begin on September 17. This announcement may involve a comprehensive reorganization of Salesforce’s AI capabilities under the Agentforce branding.
The absence of Einstein branding in this release, apart from the Einstein Trust Layer, further fuels speculation about a shift in Salesforce’s AI strategy. By developing these proprietary models, Salesforce is positioning itself to create a more integrated and cohesive AI ecosystem within its platform. The focus on function-calling capabilities in the xLAM models suggests that Agentforce will likely emphasize the creation of autonomous agents capable of interfacing with APIs and executing complex tasks across various Salesforce functions.
This strategy could potentially allow Salesforce to offer more customized, efficient, and cost-effective AI solutions to its customers, particularly in the realm of sales automation and customer relationship management.
Agentforce Too Soon?
While Salesforce’s ambitious AI strategy with Agentforce shows promise, it raises questions about alignment with the company’s own stated principles, particularly regarding explainability and observability.
The rapid move towards autonomous agents, especially those intended to interact directly with the public in service roles, is concerning given the current state of AI technology. Without robust confidence ratings or full explainable AI (xAI) functionality, there’s a risk of unreliable or inappropriate agent responses.
Moreover, as a pioneer in low-code AI orchestration, Salesforce appears to have overlooked critical observability factors in AI application development. The lack of a comprehensive Prompt Software Development Life Cycle (SDLC) to ensure prompt reliability and reproducibility is a glaring omission. While Salesforce has announced LLM logging capabilities within Data Cloud, this is merely the first step in developing true observability.
The Salesforce community has yet to establish methods and knowledge bases for managing and interpreting LLM prompts at scale. Given that nondeterministic algorithms are still in use, deploying these systems for public interaction without robust explainability and observability measures raises serious questions about whether Salesforce is truly ready to claim it’s deploying enterprise-grade AI.
It’s Raining Agentforce
Salesforce’s decision to develop its own AI models appears to be a strategic move that could pay significant dividends in the long run. By creating proprietary models tailored to its specific use cases, Salesforce is positioning itself to offer more integrated and efficient AI solutions within its ecosystem. The company’s contribution to the open-source community with xLAM-1B is also commendable, potentially fostering innovation in the broader AI landscape.
Moreover, this strategy may help Salesforce reduce its dependence on external AI providers like OpenAI, potentially leading to cost savings and greater control over its AI capabilities. As the AI landscape continues to evolve rapidly, Salesforce’s investment in proprietary AI technology could prove to be a pivotal move in maintaining its competitive edge in the CRM and enterprise software market.
