Salesforce Enterprise AI Harness: AI Control Plane, Six Capabilities and FY28 Rollout
Salesforce Enterprise AI Harness is the company’s new architecture for making AI agents easier to govern, connect and operate across large organizations. Announced on September 10, 2026, it combines six capabilities—Trusted Context, Trusted Agency, Trusted Action, Trusted Governance, Trusted Security and Trusted Models—with a new AI Control Plane designed to give companies a common view of Salesforce and third-party AI agents.
The timing matters. Businesses are moving beyond isolated chatbots toward agents that can read company data, call APIs, update systems and work across applications. That creates a different problem from simply choosing the best language model. An enterprise may have dozens of agents, several model providers, multiple data systems and different permission rules. The hard questions become: who can see what, which agent is allowed to act, how do you audit an action, and what does the whole system cost?
Salesforce’s answer is the Enterprise AI Harness. It is not a new foundation model. It is an orchestration and control architecture around models, agents, data, workflows, identity and governance. Salesforce’s official announcement says the architecture is open and composable and is intended to work with third-party models, agents and systems as well as Salesforce technology.
What Is Salesforce Enterprise AI Harness?
Salesforce describes the Salesforce Enterprise AI Harness as a reusable foundation for enterprise AI. Instead of rebuilding context, permissions, actions, governance and model connections separately for every agent, organizations can share those capabilities across AI experiences.
A useful way to think about it is as the layer between an AI model and the systems where real business work happens. A model may be good at reasoning, but an enterprise agent also needs reliable access to CRM records, inventory, contracts, policies, APIs and workflow tools. It must know which data belongs to the current user, which actions are permitted and when human approval is required.
Salesforce is drawing on technology already spread across Data 360, Informatica, MuleSoft, Agent Fabric, Tableau, Agentforce, Salesforce Guardian and the Salesforce Platform. The company says customers can use the six capabilities together or adopt only the components they need. Third-party agents, models and systems are also part of the design.
Salesforce Enterprise AI Harness: Six Capabilities Explained
| Capability | What it does | Why it matters |
|---|---|---|
| Trusted Context | Combines data, metadata, semantics, knowledge, signals and memory | Grounds agents in business-specific information |
| Trusted Agency | Reasoning, planning, state, memory and orchestration | Lets agents pursue multi-step outcomes under controls |
| Trusted Action | Connects agents to apps, APIs, tools and workflows | Moves AI from answering to doing |
| Trusted Governance | Lineage, quality, policies, guardrails and controls | Helps keep decisions tied to governed information |
| Trusted Security | Identity, permissions, privacy and runtime security | Limits what agents can see and do |
| Trusted Models | Routes work based on accuracy, performance, cost and requirements | Supports a multi-model strategy |
Trusted Context
Trusted Context is the grounding layer. It combines customer data, metadata, semantics, knowledge, real-time signals, memory and business-process information. This matters because a useful enterprise answer often depends on several systems at once. A CRM might know the customer, an ERP may know inventory, a contract contains entitlements, and a policy determines what an employee is allowed to promise.
Trusted Agency
Trusted Agency handles reasoning, planning, state, memory, collaboration and orchestration. The important idea is that flexible AI reasoning can sit beside deterministic controls. An agent can decide how to approach a task while hard business rules still control actions that cannot be left to probabilistic judgment.
Trusted Action
Trusted Action connects agents to applications, APIs, tools, workflows and business processes. This is where an agent stops merely recommending an action and actually performs one, such as updating an order, reserving inventory or triggering a workflow.
Trusted Governance and Trusted Security
Trusted Governance focuses on the data, policies and processes AI depends on. Salesforce lists lineage, quality, guardrails and controls as core parts of this layer. Trusted Security applies identity, permissions, privacy, data protection and runtime controls. An agent should not gain broad access simply because it can technically reach a connected system.
Trusted Models
Trusted Models is the model-routing layer. Salesforce says enterprises will be able to connect the right model for a task based on accuracy, performance, cost and business requirements. That makes the architecture intentionally multi-model rather than locking every workflow to one LLM.
What Is the Salesforce AI Control Plane?
The Salesforce AI Control Plane is arguably the most important part of the announcement. Salesforce says it will provide a central place to discover and register AI agents and capabilities, establish identities and policies, manage their lifecycle, evaluate performance, observe behavior and outcomes, and control cost.
This addresses a growing operational problem: agent sprawl. Teams can now create agents through many platforms, coding frameworks and SaaS tools. Without a registry, security teams may not know every agent that exists. Without common policy, two agents working with the same customer data may operate under different controls. Without shared observability, organizations can struggle to answer basic questions after an automated action goes wrong.
The control-plane concept treats AI agents more like managed enterprise infrastructure. An organization should be able to inventory them, assign ownership, define permissions, evaluate them, monitor actions and retire them when they are no longer needed.
Salesforce Enterprise AI Harness vs a Normal AI Agent Framework
A developer framework typically helps a team build an agent: connect tools, manage prompts, call a model, maintain memory and run workflows. The Enterprise AI Harness is aimed at the organization-wide layer above individual agents.
| Question | Typical agent framework | Enterprise AI Harness |
|---|---|---|
| Main goal | Build and run an individual agent | Share and govern capabilities across many agents |
| Scope | Application or development team | Enterprise-wide |
| Controls | Usually implemented per application | Designed as reusable identity, policy and governance layers |
| Models | Often configured per agent | Designed for model routing and portability |
| Observability | Agent/workflow specific | Central control-plane concept |
A normal agent framework answers, “How do we build this agent?” The Enterprise AI Harness is trying to answer, “How do we let hundreds of agents safely share company context, actions, security rules, governance and model access?”
That distinction is especially relevant as cloud-agent APIs mature. AI Tech Ledger recently examined OpenAI’s Agents API, which tackles the developer side of running long-lived agents. Salesforce is aiming at a broader enterprise-management layer around agents from multiple sources.
Why Enterprises Are Looking for AI Control Planes
Agent adoption is creating governance gaps faster than many organizations can close them. Recent industry reporting has highlighted a recurring pattern: businesses are confident in AI agents but often have weaker inventory, security gates and emergency controls than they assume.
A September 2026 report on engineering agents cited survey findings that 87% of teams experienced an agent-related security incident during the previous year. Although 77% believed they had a complete inventory of agents, only 44% could verify that inventory. Only 19% reportedly had gating mechanisms capable of stopping flawed agent-generated code from deployment, while roughly one-third had rapid-response kill switches. These figures come from broader agent-security research reported by ITPro; they are not Salesforce performance figures.
The concern is also visible in recent frontier-agent incidents. Our coverage of the Anthropic AI Threat Report 2026 looks at how increasingly capable agents change the security equation. The practical lesson is straightforward: discovery, identity, policy, evaluation and lifecycle management are becoming product categories of their own.
Salesforce Enterprise AI Harness Availability: What Is Live and What Is Not?
This is where buyers need to read the announcement carefully. Salesforce says many technologies forming the foundation of the Enterprise AI Harness are already available. However, the new capabilities and unified experience are planned to begin rolling out in early fiscal FY28.
| Item | Status as of September 12, 2026 |
|---|---|
| Underlying Salesforce technologies | Many are already available |
| Unified Enterprise AI Harness experience | Planned to begin rolling out in early fiscal FY28 |
| AI Control Plane | Announced; broader rollout is ahead |
| Final pricing | Not announced |
| Packaging and upgrade details | To be announced closer to general availability |
Salesforce has not published final pricing, packaging or detailed upgrade paths. The company says those details will come closer to general availability and warns that availability can vary by region and customer agreement. In other words, organizations should not treat the September announcement as if the complete control plane is already a finished product they can purchase today.
How the Salesforce Enterprise AI Harness Could Work in a Real Business
Imagine a customer asks a retailer whether a high-value order can be delivered today. An agent needs more than a language model. It may need the customer’s account status from CRM, inventory from ERP, contract terms, delivery capacity, fraud rules and approval limits.
Trusted Context could assemble the relevant business information. Trusted Agency could reason about possible paths. Trusted Security would restrict what the agent can access. Trusted Governance would enforce policy. Trusted Action could reserve stock and trigger fulfillment. Trusted Models could route different subtasks to models chosen for cost or accuracy. The AI Control Plane would provide the organization-level view of which agent performed the work, under which policy and at what cost.
That is the value proposition: not a smarter chatbot, but controlled execution across business systems.
Five Potential Advantages for Enterprise AI Teams
1. Reuse. If context, permissions and approved actions are defined once and shared, teams do not need to rebuild them for every agent.
2. Model flexibility. AI models are changing quickly. A governance architecture independent of a single model can reduce switching friction or let teams route work to cheaper models.
3. Visibility. A central registry and lifecycle system could help security and IT teams identify shadow agents before they become an unmanaged risk.
4. Cost governance. Agent workflows can trigger many model calls, tool calls and subagents. Central cost observation becomes more important as autonomous workflows scale.
5. Interoperability. Salesforce says the Harness is being built headlessly, with capabilities exposed through MCP, APIs, Skills and plug-ins. It names Claude, Slack, Microsoft Teams and Agentforce among the environments it intends to support.
Limitations and Unanswered Questions
Pricing is the biggest unknown. Salesforce has not announced the final cost of the new capabilities or which existing product tiers will qualify for upgrades.
The second question is how deeply third-party agents will be governable in practice. Discovering an external agent is different from enforcing granular policies over every action it takes.
The third issue is implementation complexity. Large enterprises already have identity systems, API gateways, data catalogs, observability platforms and security tools. A new AI control plane has to integrate with those investments rather than create another management silo.
The fourth issue is measurable effectiveness. Salesforce has explained the architecture, but buyers will ultimately need independent evidence about policy enforcement, latency, reliability, cross-vendor compatibility and operating cost. Until those results exist, architectural claims should not be confused with production benchmarks.
What IT Leaders Should Do Before the Salesforce AI Control Plane Rolls Out
Do not wait for a control-plane product before creating basic agent governance. Start with an inventory of every production and pilot agent, its owner, connected data, available actions, model provider and business purpose.
Next, classify actions by risk. Reading a public knowledge base is not equivalent to issuing a refund, changing customer data or approving a financial transaction. High-impact actions should have deterministic controls and, where appropriate, human approval.
Create an agent identity model. Avoid shared credentials. Each agent or workload should have traceable permissions and access that follows least-privilege principles.
Measure outcomes, not just model quality. Track task success, human escalation, errors, unauthorized action attempts, latency and cost per completed workflow. Voice and multimodal agents add another layer of operational complexity; our GPT-Live-1 API analysis shows why agent infrastructure increasingly needs to account for real-time interactions as well as text workflows.
Finally, preserve model portability. The fastest or most accurate model today may not be the right choice six months from now. Keep business rules, data access and workflow definitions as independent from a single model vendor as practical.
Salesforce Enterprise AI Harness and the Wider Agent Market
The announcement shows where enterprise AI competition is moving. Model quality still matters, but enterprise software vendors increasingly compete on the infrastructure around models: trusted data, tool access, identity, observability, workflow execution and governance.
That is relevant even to companies that never buy the Enterprise AI Harness. The same architectural questions apply to custom agents built with cloud APIs, Anthropic tools, open-source frameworks or internal platforms. Once an agent can act, organizations need to know who authorized it, what context it used, which systems it touched and how to stop it.
Salesforce’s bet is that this management layer becomes a durable part of enterprise infrastructure.
Frequently Asked Questions
Is Salesforce Enterprise AI Harness available now?
Many underlying Salesforce technologies are already available, but Salesforce says new capabilities and the unified experience will begin rolling out in early fiscal FY28. Final availability depends on region and customer agreements.
How much does Salesforce Enterprise AI Harness cost?
Salesforce has not announced final pricing or packaging. More information is expected closer to general availability.
Does the Enterprise AI Harness work only with Salesforce AI agents?
Salesforce says the architecture is designed for Salesforce and third-party models, agents and systems, with access through MCP, APIs, Skills and plug-ins.
What does the Salesforce AI Control Plane do?
It is designed to discover and register agents, manage identity and policy, control lifecycle, evaluate performance, monitor behavior and outcomes, and track costs across enterprise AI.
Is the Enterprise AI Harness an LLM?
No. It is an enterprise architecture and control layer around agents, models, data, actions, security and governance.
Bottom Line
Salesforce Enterprise AI Harness is important because it focuses on the less glamorous problem that appears after companies start deploying many agents: control. The six-capability architecture tries to connect enterprise context, reasoning, actions, governance, security and models, while the AI Control Plane aims to provide a common management layer across Salesforce and third-party AI.
The announcement is still partly forward-looking. Pricing is unknown, the unified experience has not fully rolled out, and enterprises will need evidence that cross-vendor governance works at production scale. But the direction is clear. As AI agents move from answering questions to changing real systems, the winning enterprise platforms may be determined as much by identity, policy, observability and controlled execution as by the model underneath.
Sources checked September 12, 2026: Salesforce official Enterprise AI Harness announcement and availability notes; recent industry reporting on agent security and governance.
