Salesforce Unveils ‘Trusted Enterprise AI Harness’ to Give AI Agents Business Context

Salesforce built its business on human users, but over the past couple of years, it’s been building for the digital ones. Since launching Agentforce in 2024, the company has invested in helping customers become what it calls “agentic enterprises”—firms where autonomous agents work alongside human employees. On Thursday, Salesforce introduced the latest step in its plan, unveiling the Trusted Enterprise AI Harness. It’s an infrastructure platform that gives agents business context assembled from across a company’s data and gives IT a way to manage them at scale.

The Trusted Enterprise AI Harness is one of several major announcements coming from Salesforce ahead of its annual Dreamforce conference. Last month, the company debuted Slack Code, a teams-agent collaborative coding feature, and a partnership with Anthropic that embeds Salesforce CRM inside Claude. These developments show that the digital labor force has arrived.

AI has indeed shifted workplace dynamics: humans are moving away from doing everything themselves, instead orchestrating agents to handle at least the routine, mundane tasks. This frees humans to pursue creative and strategic work—at least that’s what tech vendors promised. Still, while companies may be experiencing an influx of agents, executives have become concerned about the ROI. They want to know how AI will transform their business.

For Rohan Kumar, president and chief platform and engineering officer at Salesforce, customers have been asking him: How can they grow their organization faster? How will AI make their employees and operations “dramatically” more efficient? How can they improve conversions and deepen their relationship with customers? “These are all business questions, not AI feature questions,” he wrote in a LinkedIn post.

Answering them requires knowing the business in the first place, but that won’t come from the models. “Models don’t understand the business,” Kumar said in a press briefing on Wednesday. Typically, models are trained on the open internet or specialized around a skill. They’re not designed around knowing a specific organization. “When you put yourself in the shoes of the customer, the actual context that makes the business uniquely their business is their customers, products, employees, agents that run there, [and] the interactions that they have,” he said. “A lot of that context is very unique to an enterprise, and no matter how intelligent a model is, they basically don’t inherently understand that.”

Kumar argued that the layers around the context are becoming interchangeable. Every company will have access to models that are practically as good as everyone else’s, he said, and the cost of running them keeps falling. The agents built on top are heading the same way, with the code telling a model what to do and when becoming a commodity.

It’s why Salesforce built the Trusted Enterprise AI Harness. But it isn’t the only company arguing that context is what separates a useful agent from a generic one. Glean has made it a product line, and both Snowflake and Databricks spent 2026 making versions of the same case. What distinguishes Salesforce is its position: few companies sit on as much operational business data, or connect to as many of the systems where the rest of it lives.

About the Trusted Enterprise AI Harness

Although Salesforce calls its new platform a harness, it admits it doesn’t match the classic definition—the software scaffolding around a model that runs its tools, manages its memory and context, and turns it into an agent. Amid internal debate over the term, Kumar said that definition is “very limited” and looks at only part of what an agent can do. It also has to secure agents, govern them, and create the context the agent relies on. This is Salesforce’s definition, and one Kumar described as “the most comprehensive that I have seen.”

The Trusted Enterprise AI Harness has six components, from pulling context together and helping with decision-making to governance, security, and management. Salesforce isn’t offering it solely as a packaged suite. It’s an “open and composable” platform, meaning businesses can pick and choose the layers they need rather than buying the whole package.

“They’ll only pay for what they use, and it’s going to be a very composable, very open ecosystem,” Kumar said.

Trusted Context

Perhaps the central layer of the Trusted Enterprise AI Harness, it gives AI an understanding of an organization. It’s the knowledge layer, bringing together information from Informatica, Salesforce’s Data 360, and Tableau.

Enterprises have long dealt with basic data foundations, Kumar said. “Their data is locked into a lot of systems—SaaS applications, warehouses, lakehouses—and the volume is going up. The number one challenge that I see customers have is: Do you understand the entire data state? Where does all your data live? Is it clean? Do you understand the semantics of where things are being kept? What are the relationships between them?”

Answering these questions can help companies determine if they have AI-ready data. But instead of manually parsing databases, the Trusted Enterprise AI Harness will handle everything.

But what does this mean for the future of business intelligence—has AI ruined this field? Kumar pushed back at the question, calling apps like Tableau a “very critical part” of the Trusted Context Layer. Though tools like Claude, ChatGPT, Slack, and Agentforce make it easy to create dashboards using a prompt, “the accuracy…is going to depend on the right understanding of the semantic model,” he said. Any reports that “dashboards are dead; self-service BI is dead” are apparently greatly exaggerated. “If anything, it’s becoming a lot more important,” Kumar said.

Trusted Agency

This is where agent logic resides. Kumar compared it to a builder harness—think Agentforce or Salesforce Flow—where users can create agentic logic that uses the Trusted Context, knows which model is used, and produces decisions based on the preferred outcome. Salesforce called this the “creation of the decision.”

Trusted Action

This provides the reasoning, planning, state, memory, collaboration, and orchestration agents need to handle complex business outcomes. It brings together flexible AI reasoning with deterministic controls. Salesforce said it can help figure out whether an order can be fulfilled and what needs to happen next, all while following business rules.

Trusted Governance

According to Kumar, this layer handles cost control and management, plus data cleanliness as the foundation for the context layer. This ensures that responses are based on trusted information and that agents follow policies dictating how an order should be handled.

Trusted Security

One major concern that organizations have is around security. This layer provides the needed protection by verifying agent identities, handling data classifications, preventing data leakage, ensuring agents don’t go “rogue”, and blocking unsafe deletions.

Trusted Models

If you’re going to provide tools to help control costs, why not have one that does model routing? This layer lets enterprises securely connect to the right model for each job. This component is like what Glean, Stripe via OpenRouter, Microsoft, AWS, Palo Alto Networks via Portkey, and Snowflake offer.

Salesforce Gets an AI Control Plane

Another part of the Trusted Enterprise AI Harness is the new AI Control Plane. It’s a management and observability plane that extends across all of the above. Organizations use this to manage Model Context Protocol and Large Language Model servers. It also serves as an agent registry, defining, tracking, and querying these digital “workers.” The AI Control Plane functions similarly to ServiceNow’s AI Control Tower, Microsoft’s Agent 365, Amazon Bedrock AgentCore, and SAP’s AI Agent Hub.

Furthering Salesforce’s Headless Push

Unlike some of Salesforce’s previous product announcements, the Trusted Enterprise AI Harness doesn’t have a standard app interface. It’s built specifically for agents and is part of the company’s headless movement, which began earlier this year with the launch of Headless 360 at Salesforce’s TDX. Every capability, from building agents to securing them, managing them, and assembling the trusted context, is reachable through developer tools like APIs, MCP, or a command line rather than requiring Salesforce’s own interface. That’s what lets the harness show up wherever the work is happening, whether that’s Claude, ChatGPT, Slack, Teams, or Agentforce itself.

“We’ve really moved from AI FOMO to AI FOMU—fear of messing up,” said Rebecca Wettemann, chief executive and principal analyst at Valoir. “Salesforce’s investment in harness capabilities is a natural evolution of the trust layer as we’ve all learned more about how AI works. People worry that autonomous agents will run amok unharnessed and either do something deleterious to the business or run up a million-dollar token bill. Harnesses are designed to reduce that risk and quell those fears. Moving forward, it’s the vendors that succeed with governance and trust that will succeed in getting customers from interesting AI pilots to AI in production.”

Currently, the Trusted Enterprise AI Harness is not generally available. Salesforce said that “many of the technologies that form the foundation” are available today, though it didn’t offer specifics. It added that “new capabilities and the unified experience” are expected to roll out in Q1 FY 2028, but declined to provide details. It also didn’t provide information on packaging, pricing, or upgrade paths.

More details about the Trusted Enterprise AI Harness will undoubtedly be revealed at Dreamforce during Chief Executive Marc Benioff’s keynote. Check back here on Tuesday for a live blog of the event.

Disclosure: I’m attending Dreamforce as Salesforce’s guest, with travel and accommodation expenses covered. However, what I write reflects my own reporting and analysis. No one reviewed or approved my posts before publication.