Five Takeaways from Dreamforce 2026

Come for the giant roaming mascots and stay to learn how AIforce unlocks value in the four-tier Salesforce architecture. For three days, Dreamforce attendees packed sessions on everything from the Atlanta Hawks’ use of Data 360 to their efforts to unify fan data, all the way to a conversation with Uber co-founder and Atoms CEO Travis Kalanick. Here are the five key takeaways from the keynotes and other sessions. (For a complete live recap of the Dreamforce announcements, check out the summary on LinkedIn.)

#1 Businesses Can’t Be Run on Models Alone

AI models are powerful, but probabilistic—they understand the world, but not the rules that govern your business. As Marc Benioff, Chairman and CEO of Salesforce, said in his keynote: “They have a rough idea of what’s going on, but they aren’t anchored to a single source of truth like traditional Salesforce apps. The real question is: How do you bridge the probabilistic world of AI into the deterministic world? The deterministic world is where all that business data resides—it’s the single source of truth.” Linking model intelligence with your workflows, policies, permissions, and context is the way to empower agents to act with confidence.

#2 We’re in the midst of an interface revolution

There was the shift from DOS to the GUI, then from the web to mobile. With AIforce, all permissions, business logic, security, and analytics contained in Salesforce metadata can be consolidated into live interfaces in Agentforce Coworker, Slack, Lightning, Claude, and eventually more. As Patrick Stokes, Salesforce President of Applications & Marketing, said during the keynote: “Other software companies think their product is the user interface…. At Salesforce, our product is the trust that all of you—our customers—place in us to safeguard your data, your workflows, your business processes, your permissions, and your security rules—that is the Salesforce product.”

#3 Slack Is Your New Front Door

As Rob Seaman, EVP and General Manager of Slack, said during the Slack keynote: “There’s a new operating system for work in the AI era, and that’s Slack. It’s the only place where AI becomes multiplayer.” What does that mean? While most AI tools are “single-player”—one person, one terminal, one chat window—Slack brings agents together with human teams in an open environment, drawing on existing conversations, decisions, and context. The result, Seaman said, is a flywheel: “Intelligence amplifies itself for your employees and your AI. Your company gets smarter, and your company moves faster.”

#4 Introduction to Koa, our custom-built inference model

Salesforce collaborated with NVIDIA on its first Frontier CRM inference model, Koa. It is based on NVIDIA’s Nemotron 3 superarchitecture and was fully fine-tuned on synthetic (non-customer) data. Koa isn’t trained on customer data, but it draws on nearly three decades of Salesforce business context. As Jayesh Govindarajan, EVP of Salesforce AI and Agentforce, and Silvio Savarese, EVP and Chief Scientist at Salesforce AI Research, wrote: “We know what a service escalation looks like from the inside, what qualifying a lead actually entails, and what ‘resolved’ means when a customer is on the other end. That’s exactly what we could give Koa.”

#5 Agents Need Agents

The moment you make one step of a process agent-driven, everything downstream must also be agent-driven—otherwise, humans become the bottleneck. Joe Inzerillo, President of Enterprise & AI Technology at Salesforce, described this in “Salesforce on Salesforce: Scaling the Agentic Enterprise”: “If you take something and scale it agentically, what comes next must also be agentic scaling. If not—and that’s human scaling—then the ROI dies. So when you hear, ‘Oh, 95% of companies aren’t getting the value,’ they aren’t getting the value because they aren’t making the entire process agentic.”