Salesforce has introduced what it calls a Trusted Enterprise AI Harness: an architecture meant to give AI agents a shared understanding of a company’s customers, data, workflows, and rules before they take action. The announcement is less about a new foundation model than about the layer around models—the part that connects reasoning to enterprise context, permissions, business processes, and accountability.
An operating layer for enterprise AI
Salesforce’s example is a simple question with a complicated answer: whether an order can be fulfilled today. Customer relationship data, inventory, contracts, analytics definitions, policies, and prior interactions may all live in different systems. The proposed harness is designed to give an agent the context to answer the question, decide what should happen next, and invoke the right workflow while following the business’s controls.
Six capabilities, one control plane
The architecture groups its trust model into six capabilities. Trusted Context brings together data, semantics, knowledge, signals, and memory. Trusted Agency covers reasoning, planning, state, and orchestration. Trusted Action connects agents to applications, APIs, workflows, and tools. Trusted Governance manages lineage, quality, policies, and guardrails. Trusted Security applies identity, permissions, privacy, and runtime protection. Trusted Models handles model choice and routing based on accuracy, performance, cost, and the task at hand.
Salesforce says the same capabilities can be used together or adopted selectively. That modularity matters because enterprise AI rarely arrives as one clean replacement: teams have existing CRM, ERP, data, identity, and workflow systems, and they may want to change the underlying model without rebuilding every agent around it.
The control plane’s job
The new AI Control Plane is intended to provide a common place to discover and register agents, establish identity and policy, manage lifecycles, evaluate performance, observe behavior and outcomes, and control cost. In other words, Salesforce is proposing an inventory and governance surface for an organization’s growing collection of agents, not just a chat interface for individual users.
Designed to be composable
Salesforce says the harness is being built headlessly, with capabilities accessible through MCP, APIs, Skills, and Plug-ins. The company describes a system that can extend beyond Salesforce applications to other models, agents, and surfaces, including Claude, Slack, Microsoft Teams, and Agentforce. That openness is a strategic claim about where enterprise AI will run: across a company’s existing stack rather than inside a single vendor’s walled garden.
Availability is staged
Many of the technologies that form the foundation are available today, Salesforce says, while the unified experience is planned to begin rolling out in early fiscal FY28. Existing customers will be able to upgrade eligible Salesforce investments as capabilities arrive. The company says additional pricing, packaging, and upgrade details will be announced closer to general availability, so the announcement is an architecture and roadmap statement rather than a complete product offer.
What the announcement signals
Salesforce is betting that enterprise differentiation will come less from choosing one permanent model and more from controlling the context and actions around whichever models are available. That is a credible direction for agent-heavy businesses, but a control plane does not by itself guarantee accurate decisions. The hard work will be proving that context stays current, permissions are enforced, and evaluations catch failures before agents turn them into business actions.