Salesforce Bets on Context for Enterprise AI

Salesforce has unveiled its Trusted Enterprise AI Harness, a new offering previewed ahead of its annual Dreamforce conference. This system is designed to automatically steer generative AI models and agents, enhancing task completion predictability. The Harness integrates six core capabilities: Trusted Context, Agency, Action, Governance, Security, and Models, supported by a new AI Control Plane for managing AI across an organization.

The enterprise AI market is increasingly multi-platform, with most organizations using multiple agent orchestration platforms simultaneously. Salesforce's strategy emphasizes that durable enterprise advantage will stem from proprietary context and operational controls surrounding foundation models. Rohan Kumar, Salesforce's chief platform and engineering officer, highlighted that "enterprise trusted context"—encompassing customers, employees, data, and history—remains unique.

Salesforce broadens the definition of an AI 'harness' beyond typical coding agent tools, aiming to encompass enterprise context, security, and governance. Recent research from Salesforce demonstrates that optimizing the harness around a model can significantly improve agent performance. This work shows that the system prompt, toolset, and execution scaffolding surrounding an LLM can dramatically impact agent accomplishments.

The Trusted Enterprise AI Harness is organized into six "trusted" capabilities that customers can use together or selectively. This composable architecture aims to integrate with existing third-party systems. The AI Control Plane provides a centralized hub for discovering agents, establishing policies, managing lifecycles, and monitoring performance and costs.

Technologies from Data 360, Tableau, Informatica, and MuleSoft form the foundation of this new architecture. Salesforce emphasizes an open approach, allowing integration with various AI systems and applications. While many underlying technologies are available, the full Harness is planned for a phased rollout starting in early fiscal year 2028, beginning February 2027. Pricing and licensing details are still to be announced.