Share with your CIO
Salesforce is betting that the competitive moat in enterprise AI won’t be model quality but contextual depth, and it’s restructuring its entire platform stack around that thesis. The company has packaged capabilities from Data 360, MuleSoft, Tableau, and Agentforce into what it calls the Enterprise AI Harness, a six-layer architecture covering context, agency, action, governance, security, and model routing. A new AI Control Plane sits above all of it, giving enterprises a single dashboard to register, monitor, and govern AI agents across both Salesforce and third-party systems. New unified capabilities begin rolling out in early fiscal FY28.
What this means for your business
The story this announcement tells most clearly is about switching costs, not features. If your organization already runs Salesforce for CRM, MuleSoft for integration, and Tableau for analytics, Salesforce is arguing that your existing data, permissions, and business logic are now the scaffolding for your AI layer, and migrating away becomes considerably more expensive once agents are trained on that proprietary context. CIOs who are mid-evaluation on their AI platform strategy need to decide whether that depth of integration is an asset worth doubling down on or a lock-in risk worth pricing carefully.
The AI Control Plane is the most operationally significant piece here, and it’s the one that deserves the most scrutiny. The recurring failure mode in enterprise AI deployments isn’t bad models, it’s unmanaged proliferation: agents spun up by separate teams, each with different permissions, unknown costs, and no unified audit trail. A single control plane that can register and govern agents regardless of whether they were built on Salesforce or a third-party system addresses a real organizational pain point. The credibility of that “openness” claim depends entirely on whether third-party agent registration is genuinely functional or just a marketing posture, and Salesforce hasn’t shown that in production yet.
The headless architecture and MCP (Model Context Protocol, a standard for letting AI models call external tools and data sources) support signals that Salesforce is serious about appearing in AI surfaces it doesn’t own, including Claude and Microsoft Teams. That’s a defensive move as much as an offensive one. If enterprise AI increasingly happens inside general-purpose assistants rather than inside Salesforce’s own UI, Salesforce needs its context and action layers to be callable from anywhere. I’d revisit this assessment if the FY28 rollout shows MCP integrations working reliably in non-Salesforce surfaces at scale, because that’s when the “bring Salesforce to any AI” claim either earns its credibility or collapses into a slide deck promise.
Concept deep-dive: AI Control Plane
A control plane is the management layer that sits above a system’s operational components and coordinates what they’re allowed to do, how they behave, and how much they cost. In networking, the control plane routes traffic rules; the data plane moves the actual packets. Salesforce is applying the same logic to AI agents: one place to see which agents exist across the enterprise, set their permissions, track their performance, and cap their spend, regardless of which vendor built them.
Based on reporting from Salesforce Rolls Out the Trusted Enterprise AI Harness, originally published 2026-09-11 07:10:00.
