Salesforce introduces Enterprise AI Harness, AI Control Plane

Salesforce has previewed two products aimed at one of the messier parts of enterprise AI: turning a large language model into something that can actually do work without wandering off the rails. The company is calling the first one Enterprise AI Harness, and the second one AI Control Plane.

The harness is meant to make AI agents more reliable and more secure. Salesforce says many of the underlying pieces already exist in its cloud services, while the rest should begin rolling out early in its 2028 fiscal year, which starts next February. In other words, this is partly a packaging job and partly a future roadmap.

Salesforce says Enterprise AI Harness will bundle six main feature collections, with three centered on the business data agents rely on. Trusted Context is the name for the layer that brings together records such as customer data and product availability, plus what Salesforce calls real-time signals. The company also says it will add data-quality controls, including a data lineage tool, and give cybersecurity teams a way to control which agent can access which dataset and how.

There’s also a practical angle here. Companies often use more than one model in an agent project, sending harder tasks to a frontier model and cheaper requests to a smaller one. Salesforce wants Enterprise AI Harness to handle that with intelligent model routing, automatically sending each request to the most suitable LLM based on things like accuracy and cost. It also says the system will add deterministic reliability controls and make it easier to connect agents to outside applications.

The AI Control Plane is the cleaner, more managerial half of the pitch. It gives administrators a central place to monitor agent performance and inference costs, while also helping enforce cybersecurity policies. That’s the kind of product buyers tend to want after they’ve already spent months discovering that “AI everywhere” also means “problems everywhere.”