With AI agents no longer confined to pilot programmes but part of the daily business fabric, enterprises are confronted with an issue that is more than a matter of model selection. The real question is how to scale these systems in a secure manner while ensuring they operate within set boundaries and have a grasp on the business at hand.
Salesforce has put forward two elements of its AI strategy to meet this head on: an AI Control Plane and the Trusted Enterprise AI Harness. In effect, they provide a unified base for the deployment and oversight of AI agents on both third-party and Salesforce systems.
Building a More Trustworthy Foundation for Enterprise AI
The Harness is built around six composable capabilities: models, security, governance, action, agency and context.
There is the matter of trusted context, which pools everything from metadata and knowledge to real-time signals and customer data so agents have a common view of the organisation. Then there is trusted agency, which handles the reasoning and orchestration but also enforces deterministic controls for compliance and approvals.
When it comes to execution, trusted action will have agents interfacing with APIs and workflows to reserve inventory or update orders, or ceding the task to a human if the situation calls for it. The other layers are concerned with the finer points of policy and security, as well as the flexibility to direct workloads to different models depending on cost, performance or accuracy. Underpinning all of this is a suite of Salesforce tech from Data 360 and MuleSoft to the Salesforce Platform and Guardian.
One Control Plane to Manage an Expanding AI Workforce
For an organisation with agents spread across various vendors and departments, keeping track of where AI is active can be difficult. The AI Control Plane is meant to be the central point for that. CIOs and platform teams will have the visibility to see what agents are running, the data they are touching and the models in use. From there they can register and identify agents, set policies, evaluate the results and rein in costs.
Keeping Enterprise AI Open and Model-Agnostic
It is an open and headless setup. Access is via MCP, APIs and plug-ins, and it works with Agentforce as well as Microsoft Teams, Slack and Claude. Being model-agnostic is key for an enterprise that does not wish to have its AI plans dictated by a single vendor as the market changes.
Why Enterprise AI Needs More Than a Powerful Model
“The value of enterprise AI is found in the proprietary context around the model, not the model itself,” says Rohan Kumar, Salesforce’s President and Chief Platform and Engineering Officer.
In short, for any business looking to take an AI pilot to production, Salesforce is offering the means for agents to understand their environment, act with the proper controls and adapt over time.
