How Salesforce Is Using AI to Power the Enterprise

The Chief Scientist role began when AI research was founded about 10 years ago. My goal is to supervise and observe AI's evolution within the company. I focus on how AI can drive transformations and innovations for Salesforce's business and products. We concentrate on core innovation and research that can eventually be deployed in Salesforce's key products.

Our research concepts formed the foundation for key products. For instance, we developed large language models for developers, specifically for language-to-code translation. This was four years ago, well before ChatGPT. We built these models entirely in-house to boost developer efficiency and accelerate code creation. This eventually became a generally available product a couple of years ago.

Agents are more than just large language models; they are complex systems. They include a memory component that retrieves information, data, documents, and past conversations. Techniques like Retrieval Augmented Generation (RAG) or other embeddings enable this memory. Another key component is the brain, a reasoning engine that transforms a query or task into a series of steps. This represents a significant innovation.

For instance, a sales division might manage data one way, while the marketing division uses a different approach. Connecting these dots is crucial. Without proper data harmonization, AI cannot function effectively. That's one key aspect. The second important aspect is trust. Customers are very concerned about the quality and accuracy of AI. This includes whether AI is safe to use from a trust perspective.

Simulation environments are crucial for building trust and ensuring the quality I mentioned earlier. Unfortunately, we're seeing a saturation in the scaling law of learning. Even with GPT-5, we haven't witnessed the breakthrough some anticipated. We still observe inconsistencies in some results. Therefore, we must adopt new learning paradigms based on experience and observing how these agents behave in the real world. The most effective way to achieve this is by integrating them into simulation environments.

These are essentially simulation organizations. For example, you might simulate a customer service scenario where agents interact with a simulated human. The environment contains various components that provide data to answer common customer service questions about returns, product availability, or policies. These simulations are built using real customer information. This setup allows for a gameplay-like experience where an agent interacts with a simulated human, enabling us to identify potential areas of failure.

The main innovation will occur when agents are deployed at scale. Soon, we will each have personal agents that perform tasks on our behalf. For instance, if I need a loan, I won't call the bank myself; I'll ask my agent to do it. This will lead to interactions between my agent and the bank's agents. We need to envision a future where agents communicate with each other, requiring us to redefine protocols, communications, and guardrails. It will be like imagining the internet before its existence. Society will completely transform, and it will be very exciting.