How Salesforce is trying to get companies to trust agents

A year ago, Salesforce CEO Marc Benioff announced a huge bet on AI agents amid the cartoonish sprawl of his company’s annual Dreamforce conference, which resembles something of a Disneyland for B2B SaaS enthusiasts.

This year, the sprawling San Francisco convention was remade in the image of Agentforce; the term “agent” was mentioned in around three-quarters of the 1,443 sessions. The company says 12,000 customers are now using the autonomous AI builder platform.

That may sound like a lot, but Salesforce has more than 150,000 customers, meaning 92% have yet to adopt Agentforce.

Agents have been billed—particularly by Salesforce—as the next big chapter in AI adoption in the office. But one of the biggest obstacles for companies is trust: Handing over business processes to free-acting AI is a big ask, reports have shown.

Salesforce VP of AI and Agentforce Nancy Xu told Tech Brew that building agents that are predictable at scale is a major focus for her employer. It’s also one of the biggest barriers for companies that may have tried agents as a proof of concept, but encountered snags around reliably integrating business processes when fully deployed, she said.

“We have customers that go to the big LLM companies, like, ‘Oh, this is great. I just can plug in. It’s so easy, I get an agent working off the bat,’” Xu said. “But if it doesn’t verify your user’s identity first and follow your business processes, it’s not an agent you can use at scale…Yes, they can get the initial implementation. The agent can say things that conversationally make sense, but then if they don’t follow the processes, it doesn’t scale up.”

Salesforce’s latest version of its agent program, Agentforce 360, has features designed to improve predictability, like a new programming language called Agent Script that allows for “more determinism and more control,” according to Xu. They’re also supposed to be easier to build so nontechnical subject matter experts can hone them.

When it comes to hallucinations, Salesforce President and Chief Legal Officer Sabastian Niles said during Dreamforce last week that the company thinks about three types of shortcomings: inaccuracy, irrelevance, and a lack of creativity. Sometimes, there can be trade-offs between them, but Agentforce is designed to let customers combine the more creative aspects of generative AI with more deterministic actions, he said.

“If you try to completely eliminate hallucinations, maybe you lose the creativity,” Niles said. “As you bridge those two together—determinism, non-determinism in agentic deployments—you really also should be reducing the set of hallucinations in a way that is effective across accuracy, relevancy, but also bringing forward the right sense of creativity.”

Security concerns are also becoming more complex as companies build networks of multiple agents, according to Brad Arkin, Salesforce’s chief trust officer. Arkin said that as agents interact with multiple other agents and humans in different roles, security concerns multiply.

“How on earth do you capture all of that, and how do we consider unexpected outcomes that may arise from the composition of these very complex capabilities?” Arkin said. “And that composability analysis is something that, as an industry, everyone is anticipating is going to get complicated really fast.”

Salesforce’s Futures team has considered possible bigger-picture scenarios around an agentic future, according to Niles, from “AI trade wars” to “digital sovereignty issues.”

“We’ve talked about, what if we really do end up manifesting the [agent-to-agent] economy,” Niles said. “There’s a lot of different implications about that, and not just on the upside opportunities, but areas that we ourselves take very seriously around trust and ethics.”