With AI Foundry, Salesforce lays the groundwork for AI as the operating system of the enterprise

Salesforce AI Research Launches AI Foundry: Pioneering Enterprise AI Development

Salesforce AI Research has unveiled AI Foundry, a groundbreaking initiative aimed at uniting research teams, clients, and academic partners to innovate and validate AI solutions tailored for enterprise systems. As large language models mature, the real challenges of AI adoption within organizations lie within information systems—a critical juncture where accuracy, consistency, and reliability must operate at scale.

AI Foundry serves as a research and development infrastructure that brings together AI research teams, customers, and academic collaborators with the goal of accelerating the transition to Salesforce’s product roadmap. This initiative is grounded in a keen understanding of market dynamics: as technical advancements in large language models reach maturity, the orchestration and observability layer of agents emerges as the central challenge for enterprise adoption. The sophistication of this layer is indicative of what future enterprise operating systems will look like in the age of AI.

“The most significant issues for enterprises no longer reside at the model level but now revolve around information systems, where components must work together to ensure accuracy, consistency, and reliability at scale,” explains Silvio Savarese, chief scientist at Salesforce. This perspective resonates with many IT leaders who find that generative AI deployments often underperform in real-world settings compared to results seen in controlled tests.

The Layer Shaping the Future of Enterprise AI Operating Systems

The analogy to an operating system (OS) is not merely rhetorical; it is functional. An OS performs three fundamental functions: it allocates resources, arbitrates priorities among competing processes, and exposes standardized interfaces for applications that run on top of it. This is precisely the role the orchestration layer for agents is beginning to fulfill in enterprise architectures. When multiple autonomous agents coexist—such as billing agents, customer relationship agents, and compliance agents—an infrastructure is needed to determine which agent operates, in what order, with which data, and under what security and governance constraints.

The agent cards developed by AI Foundry embody this logic: they function like process descriptors in an OS, carrying the identity, capabilities, and rights of each agent, thus enabling the orchestration layer to recognize and coordinate them. Observability acts as the kernel, revealing the internal state of the system, logging decisions, detecting behavioral anomalies, and facilitating human intervention when escalation is necessary. Without this layer, agents operate as “black boxes,” a tolerable scenario for demonstrations but unacceptable in regulated production environments.

For a Chief Information Officer (CIO), this perspective crucially shifts the investment conversation from “which model to choose” to “which orchestration and observability layer to deploy, and on what architecture.” Companies that treat this issue as secondary to model selection risk repeating the mistakes of those in the 1980s who purchased hardware without considering the unifying operating system. In this context, AI Foundry is not merely a research lab; it represents the framework through which Salesforce is tackling this pivotal question for its entire customer ecosystem.

eVerse: Simulating Agents Before Deployment

The first focus area of AI Foundry is simulation. Salesforce AI Research has developed eVerse, a simulation environment designed to expose agents to thousands of extreme cases, state transitions, and decision-making scenarios before they are deployed in production. The underlying principle leverages learning through simulated experience. Agents are placed in realistic business situations, allowing for assessment of their performance and behavioral consistency in conditions that standard evaluations often fail to replicate.

The most documented application to date involves Agentforce Voice, Salesforce’s voice interface, which has been rigorously tested across thousands of simulated conversations. A second use case unfolded in a hospital context: eVerse was employed to manage the billing agents of the University of California, San Francisco (UCSF) call center, an area where accuracy and compliance leave virtually no margin for error. These implementations illustrate the tool’s core mission: minimizing the risk of behavioral regression when scaling autonomous agents.

Agent Cards: The Identification System for AI Agents

The second area of focus tackles what Salesforce refers to as ambient intelligence. The goal here is to seamlessly integrate AI into enterprise workflows contextually and proactively, without overwhelming users with information. This approach places the interaction between humans and AI systems at the heart of design rather than at the periphery. The objective is to create AI that is “always on but never intrusive,” a phrase that encapsulates the balancing act between constant availability and operational discretion that CIOs seek to achieve.

The third initiative revolves around agent cooperation. AI Foundry is developing interoperability protocols, including agent cards—standardized identity cards that allow distinct agents to recognize and interact with one another in a controlled manner. This effort encompasses the construction of a multi-agent semantic layer, integrating standardized protocols, safeguards, decision logs, and escalation mechanisms. Notably, Salesforce’s legal teams and ethical technology office are directly involved in this research to define frameworks applicable to autonomous negotiations between agents. This early inclusion of legal consideration in AI research initiatives is rare and worthy of attention from corporate legal and compliance departments.

Bridging Research and Commercialization

AI Foundry is also set to expand Salesforce’s academic grant program and partner ecosystem by linking external researchers and cross-sector talent to the operational challenges faced by strategic clients. The combination of cutting-edge research and real-world business constraints, conducted in short iterative cycles, distinguishes AI Foundry from conventional research labs. “Many established rules no longer apply. AI Foundry connects fundamental research to tangible business challenges, collaborating closely with our strategic clients through rapid iteration cycles,” says Silvio Savarese.

The structure of AI Foundry—with clients engaged from the research phase, short iteration cycles, and accelerated transitions to product roadmaps—describes a mutual co-development mechanism rather than a typical R&D lab. Strategic clients contribute their data, operational constraints, and real use cases, while Salesforce fast-tracks validation of its architectural choices, significantly reducing product risk before any industrial development investment. UCSF is not just a volunteer client in this narrative; it is a genuine testing ground whose outcomes directly inform the roadmap for Agentforce. The underlying mechanics resemble a use case qualification pipeline funded in part by the clients themselves.

Scroll to Top