How Salesforce Is Using AI to Power the Enterprise

Salesforce built large language models for code generation nearly four years before ChatGPT's release, focusing on developer efficiency. The research team also developed the trust layer for AgentForce and the Atlas Reasoning Engine through customer partnerships, demonstrating early enterprise AI innovation.

• •Built language-to-code models 4 years before ChatGPT for internal developer productivity
• •Developed trust layer components in-house to verify AI output safety for customers
• •Created Atlas Reasoning Engine through customer collaboration, launched with AgentForce
• •Research organization focuses on core innovation that deploys into key Salesforce products

AgentForce is positioned as a comprehensive agentic platform, not just a chatbot builder. Savarese breaks down the four critical components of enterprise agents: memory (RAG), reasoning engine (brain), actuators (API/function calls), and interface (communication modalities), emphasizing that agents require much more infrastructure than standalone LLMs.

• •Memory component uses RAG and vector databases for retrieving documents and conversation history
• •Reasoning engine (Atlas) breaks down queries into executable step-by-step plans
• •Actuators enable function calling for actions like sending emails or making reservations
• •Interface layer handles multi-modal communication (voice, text) with external world
• •Enterprise agents require full infrastructure stack, not just LLM access

Large enterprises face critical challenges with siloed data across divisions (sales, marketing, etc.). Salesforce's key value proposition is connecting disparate data sources and providing harmonization, which is essential for AI agents to operate effectively across organizational boundaries.

• •Sales and marketing divisions typically manage data differently, creating silos
• •AI cannot operate effectively without proper data harmonization across touchpoints
• •Connecting enterprise data is a prerequisite for successful agent deployment
• •AgentForce's agent builder provides infrastructure to connect data with agent functionality

Enterprise customers prioritize trust, accuracy, and safety over raw capabilities. Savarese identifies 'jagged intelligence' as a critical issue—LLMs can prove theorems but fail at common-sense business tasks. This inconsistency is particularly problematic for high-stakes applications like loan processing.

• •LLMs exhibit 'jagged intelligence': excel at complex tasks but fail at common-sense operations
• •Common-sense failures are often the most business-critical scenarios
• •Trust layer ensures consistency and quality alongside pushing capability boundaries
• •Example: loan agents must avoid wrong computations that damage customer and bank relationships

Salesforce builds enterprise-specific simulation environments to stress-test agents, moving beyond scaling laws to experience-based learning. These environments simulate CRM systems with real organizational structures, testing agents against corner cases like accented voices, background noise, incomplete requests, and conflicting instructions.

• •Simulation enables experience-based learning as scaling laws show diminishing returns
• •Environments simulate real CRM/enterprise systems with actual fields and objects, not game environments
• •Voice agents tested against multiple accents (including Italian), background noise, incomplete/conflicting requests
• •Uses synthetic data validated by experts to minimize reality gap between simulation and production
• •Agents play against simulated humans to identify failure modes in customer service scenarios
• •Testing covers product availability, returns, policies in realistic organizational contexts

Savarese envisions a near-future where personal agents interact with enterprise agents at scale, requiring new communication protocols and guardrails. This agent-to-agent interaction paradigm will transform society similarly to how the Internet revolutionized communication.

• •Personal agents will perform tasks on behalf of users (e.g., calling banks for loans)
• •Agent-to-agent communication will require standardized protocols and guardrails
• •Transformation comparable to pre-Internet era—fundamental societal shift expected
• •Scale deployment of interacting agents represents the next major innovation wave