Salesforce Spots Parallels with OpenClaw in Enterprise AI

The OpenClaw-style agents are on Salesforce's radar—but require guardrails. Plus, the four key pillars to help move agentic pilots into production

Salesforce CEO Marc Benioff gestures during his keynote address at the company's 2025 Dreamforce event. Credit: Ken Yeung

Salesforce is paying close attention to the development of OpenClaw, betting that technology enabling AI agents to operate computers autonomously could fundamentally change how enterprise software is used, automated, and scaled. The company believes these sophisticated bots can add significant value, but warns that it’s critical to harness this power correctly and ensure the right security guardrails are in place.

Powered by Agentforce, Salesforce is actively pitching its vision of the agentic enterprise—a future in which humans and AI agents collaborate within organizations. In a couple of years, the company has shepherded the enterprise from reactive assistants to proactive agents capable of taking initiative and giving rise to the digital labor force.

What OpenClaw (formerly Moltbot and Clawdbot) demonstrated is that truly autonomous AI assistants are no longer theoretical. People can have AI take the proverbial wheel to get things done. The burst of activity over the past week has put a renewed focus on Salesforce’s approach, revealing clear parallels.

“They give a glimpse of how agents, when running in your personal sort of context, on your device, etc., can really add a lot of value,” said Muralidhar Krishnaprasad (MK), Salesforce’s president and chief technology officer, during a briefing this week outlining the company’s AI strategy.

His colleague, Jayesh Govindarajan, executive vice president of AI, added, “It just shows what’s possible when you give unfettered access to…these large language models.” He noted it also demonstrates “the unreasonable power of being able to orchestrate multiple agents in parallel that each do a particular task.” Pointing to Moltbook, the social network for these advanced agents, Govindarajan highlighted it as a real-time example of orchestration in action.

“It shows the art of the possible,” he said. “A lot of what we are doing sets us up to bring that into the enterprise, just like an agentic mesh, which is where we’re going with orchestration, if you think about it.”

Still, despite the fanfare, both MK and Govindarajan stressed OpenClaw by itself is not ready for enterprise primetime—it lacks the necessary protections. MK notes that although it’s “awesomely powerful,” it’s also “awesomely a nightmare from a security perspective, because [OpenClaw] was opening all your secrets to the whole wide world. So part of the things you need to make sure of is that we harness this power right and put those right security guardrails so that it can be very useful from a personal perspective.”

He suggests that combining the capabilities of OpenClaw bots with the right context, auditability and governance controls, and determinism, the agents could be “quite powerful.”

Even so, Clawbot’s technology isn’t entirely new to Salesforce. MK emphasized that the company has long run agents and AI at the edge, pointing to Salesforce’s mobile SDK that ships with its field service app and Slack’s new Slackbot as examples of OpenClaw-style capabilities—delivering similar functionality but within enterprise-grade guardrails.

Turning LLMs into Enterprise Value

Salesforce isn’t the only industry player offering agentic support for the enterprise, but the company says it’s seeing strong momentum: Agentforce has closed more than 18,500 deals—a 50 percent increase quarter over quarter—with over 70 percent of customers advancing agents from pilot to full production.

The key to reaching such a high conversion rate lies in what Salesforce calls the “Last Mile.” “[Large language models] are like nuclear reactors—they’re super powerful [and] can give a lot of energy,” MK explained. “It’s very good, but you also need that Last Mile to really harness the energy correctly, transform it right….If you just let it run wild, [it will] hallucinate and bring down your business. The Last Mile is really what makes your LLMs relevant in an enterprise and gives significant value to your business.”

The Four Pillars of the Last Mile

Context

Unsurprisingly, context includes who the customer is, their history, along with the current status of orders, support tickets, entitlements, and policies. MK boasted that Salesforce is the only platform with more context than anyone else, as it’s the leader in sales, service, marketing, commerce, analytics, and Slack. “We truly have that rich data and the context to feed these agents.”

For example, when William Sonoma built its sous chef agent called Olive, it was important to have the “proprietary context to understand what a specific return and refund policy looks like…How does delivery work? How do split delivery options and escalations to humans for refunds work in the context of William Sonoma?” Govindarajan shared. “All of this information sits within the enterprise, and being able to bring that as memory and as context to then power the 150,000 monthly conversations…is the key bit that we’re trying to do with context.”

Control

To ensure models behave as expected, controls must be implemented. In this area, Salesforce leveraged what it called “deterministic guardrails,” explicit rules and logic that reliably control what a bot can and cannot do every single time, overriding how an LLM might want to respond.

To illustrate this, Govindarajan described a situation in which Adecco wanted to use agents to screen applications. The recruiting firm sees about 300 million applications and places about 1 million. Historically, Adecco’s team could only screen a small fraction of applicants, but it didn’t want to let AI loose. They needed a fixed, auditable qualification process. Ultimately, a script was produced with deterministic steps—e.g., ask certain qualifying questions first; if they are met, move on to the next set of questions; and enforce pass/fail conditions at each stage. This process enabled Adecco to hand off qualifying 51 percent of candidates to Agentforce.

Observability

If you can’t see what’s happening inside the agent and trace what happened, how will you know that the responses are accurate? Observability provides a full account of every step the bot takes, along with topic-level performance, LLM-as-judge scoring, and audit logs for compliance and debugging.

Orchestration

The final pillar goes beyond ensuring the agent can handle business processes and enterprise data. It’s about orchestrating all digital labor across the organization—whether it comes from human resources, backend systems, or other sources.

These four areas “are based on the work that we did with [the 18,500 customers] and just shaped how the agentic operating system that we built with Agentforce has come about,” Govindarajan stated.

Responding to the Skeptics

There may be those on Wall Street and others who question Agentforce’s viability. From “decision fatigue” and complaints about a lack of pricing transparency to concerns about the platform’s inability to prevent hallucinations, the road hasn’t exactly been smooth. Add to that the loss of key executives over the past couple of years—namely, Clara Shih, Tableau Chief Executive Ryan Aytay, Slack chief Denise Dresser, and Salesforce Chief Trust Officer Brad Arkin—only heightens the skepticism.

But 18,500 Agentforce customers isn’t a statistic to shy away from. So, how does Salesforce respond to the concerns from pundits, investors, analysts, and customers?

“For any business to run in this world, you need to be able to attract customers,” MK replied. “You need to be able to sell your product to that customer, to service that product, to analyze what they have done, and you need to have collaboration with your employees. Guess which company in the world can actually do that for you, being number one in all of them? That’s Salesforce.”

He goes on to add, “In many ways, it’s not just about the data, it’s also about the metadata, about the process, workflows, and the verticals that go with it…For every investor, analyst, pundit, etc., the message that I’ll be giving is that we have reimagined agents, not just as some standalone thing outside that can give you some answers, but as an integral part of your business. We, out of the box, ship over 500 different agents for every one of your verticals, which you can customize because it’s built on the same platform, and you can make it your own.”

“For us, the customer is the heart of our entire thing,” MK concluded. “CRM—customer relationship management—customers are the heart of our company’s ethos, and we make it work. That’s the message I would give.”