# How Salesforce Is Using AI to Power the Enterprise Source: https://www.youtube.com/watch?v=K20bOHtTOZY Recap page: https://rapidrecap.app/video/K20bOHtTOZY Generated: 2025-11-11T11:08:19.506+00:00 --- ## Quick Overview Salesforce Chief Scientist Silvio Savarese explains that the main innovation in AI for the enterprise will come from creating sophisticated, specialized agents that can simulate complex, real-world interactions and data, rather than relying solely on large language models (LLMs). These specialized agents, which Salesforce has been developing for about a decade, must be trustworthy and context-aware to handle business-critical tasks like loan applications or customer service interactions effectively, bridging the gap between simulation and reality. **Key Points:** - Salesforce Chief Scientist Silvio Savarese states that the future of enterprise AI lies in specialized agents, not just LLMs, which the company has been developing for about ten years. - These specialized agents must be capable of simulating complex, real-world scenarios, such as loan applications or customer service interactions, with high accuracy. - The key challenge is ensuring trust and context, as current LLMs often produce inconsistent results when dealing with siloed or specific data. - Salesforce develops agents that can perform specific, well-defined tasks (like sending an email or opening a website) by connecting to various components and data sources. - The goal is to create agents that can perform complex actions on behalf of customers, bridging the gap between simulation and real-world reality. - The complexity of simulating enterprise environments (B2B/B2C) requires specialized agents that can handle context and different interaction modalities (voice/text). ![Screenshot at 00:22: Silvio Savarese explaining the role of AI scientists at Salesforce and defining what AI means in the context of their enterprise products.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-00-22.png) **Context:** The video features an interview between Dan Shipper, CEO of Every, and Silvio Savarese, Chief Scientist at Salesforce, recorded during the Dreamforce conference. The discussion centers on the evolution of Artificial Intelligence (AI) within the enterprise, specifically focusing on the shift from general-purpose Large Language Models (LLMs) to highly specialized, trustworthy, and context-aware autonomous agents designed to execute complex business processes. ## Detailed Analysis Silvio Savarese argues that the next major wave of AI innovation for enterprises will involve deploying highly specialized agents capable of performing complex, real-world tasks, rather than relying on general LLMs like GPT-4. Salesforce has been building these agents for about ten years, focusing on creating systems that can simulate enterprise environments (B2B/B2C) accurately. A critical component is ensuring trust and quality, as Savarese notes that LLMs often struggle with the consistency required for business-critical tasks, leading to potential failures when dealing with siloed customer data or conflicting requests. The agents he describes function almost like a toolkit, invoking specific functions (like sending an email or making a reservation) based on real-world data and context. The ultimate goal is to create agents that can operate reliably on behalf of customers, bridging the gap between simulation and actual execution. Furthermore, Savarese highlights the need for these agents to reason about context and handle multiple interaction modalities (voice/text), pushing beyond simple LLM outputs to achieve complex, trustworthy actions. ### AI Innovation Focus - Future innovation centers on specialized agents, not just LLMs - Agents simulate complex enterprise scenarios (loan applications, customer service) - Trust and context are paramount for deploying agents in critical business functions. ### Agent Architecture - Agents use a toolkit of functions that interface with real data - Agents are designed to handle different data silos and ensure consistency across customer touchpoints - Simulation environments are crucial for stress-testing agents before real-world deployment. ### Salesforce's History - Salesforce has been developing these specialized agents for nearly a decade - The Atlas reasoning engine, introduced last year, is a key component for enabling agents to reason and act. ### The Trust Gap - Current LLMs often fail on consistency when faced with real-world data or conflicting requests (e.g., loan applications) - The challenge is building agents that guarantee accuracy and reliability. ### Future of Agents - Future agents will be designed to communicate with each other via defined protocols - This will lead to complex, interconnected workflows that transform the entire enterprise landscape. ![Screenshot at 00:15: Dan Shipper, CEO of Every, introduces the guest, Silvio Savarese, Chief Scientist at Salesforce.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-00-15.png) ![Screenshot at 00:24: Silvio Savarese begins to explain the role of the Chief Scientist and the evolution of AI.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-00-24.png) ![Screenshot at 00:50: Savarese describes the focus on core innovation and research for deploying AI products.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-00-50.png) ![Screenshot at 01:35: Savarese elaborates on how this foundational work on agents predates the public introduction of models like GPT-3.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-01-35.png) ![Screenshot at 02:05: Savarese details the four components of their agent framework, emphasizing memory and reasoning.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-02-05.png) ![Screenshot at 03:22: Dan Shipper asks about positioning Salesforce's approach against direct large-scale AI deployment.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-03-22.png) ![Screenshot at 05:55: Savarese emphasizes that agents must be built based on real customer data and policies for trust.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-05-55.png) ![Screenshot at 07:14: Savarese discusses the need for agents to connect to the right data and the importance of context.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-07-14.png) ![Screenshot at 08:39: Savarese discusses the need to make agents super-robust to avoid failure in high-stakes scenarios.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-08-39.png) ![Screenshot at 12:11: Dan Shipper asks for clarification on the simulation environment used for training these agents.](https://ss.rapidrecap.app/screens/K20bOHtTOZY/00-12-11.png)