Salesforce Trusted Enterprise AI Harness: What Ops and Security Should Verify

Verdict: Salesforce’s Trusted Enterprise AI Harness (announced around 11 September 2026) is a serious architecture pitch — six trusted capabilities plus a new AI Control Plane — but ops and security should treat it as a roadmap and RFP checklist, not a self-serve SKU you can buy tomorrow. Unified experience timing points to early fiscal FY28; pricing and packaging remain TBA.

Best for: Salesforce-centric enterprises evaluating how agents, models, and third-party AI will be governed across CRM and integration estates.
Not for: Teams expecting a consumer AI toggle next week, or buyers who need fixed public list prices before a pilot conversation.

Researched 2026 news overview based on Salesforce’s public announcement — not a vendor security audit.

What launched

Salesforce introduced a composable Trusted Enterprise AI Harness spanning six capabilities: context, agency, action, governance, security, and models. Alongside them, an AI Control Plane is positioned as the place to discover/register agents, set identity and policy, manage lifecycle, evaluate performance, observe outcomes, and control cost — across Salesforce and third-party AI. Foundational pieces exist today; the unified experience is slated to begin rolling out in early FY28. Pricing, packaging, and upgrade paths are to be detailed nearer GA.

Why it matters for buyers

The category shift is from “ship another chatbot” to “harness agents with business context and controls.” That raises classic enterprise questions: which systems supply trusted context, who approves agent actions, how model routing picks cost vs accuracy, and whether third-party agents inherit the same identity rules as first-party ones.

What ops / security should verify

• Timeline realism: What is available now vs early FY28 unified experience — get a dated capability matrix.
• Six-capability mapping: Ask which products (Data 360, MuleSoft, Agentforce, Guardian, etc.) implement context, agency, action, governance, security, and models in your org today.
• AI Control Plane scope: Does it cover third-party LLMs/agents in practice, or only Salesforce-signed components?
• Identity & least privilege: How agent identities, permissions, and human escalation are enforced at runtime.
• Action guardrails: Which write actions can agents take in CRM/ERP, and how are they dual-controlled?
• Observability & cost: Log export to your SIEM; unit economics for model routing and agent runs.
• Training & data use: Contract language for customer data in model improvement — especially mixed Salesforce + external models.
• Pricing opacity: Packaging TBA — demand SKU clarity, overages, and upgrade paths from existing Agentforce/Einstein entitlements before budgeting.

Procurement checklist (short)

Request: a now-vs-FY28 matrix; Control Plane admin demo on a third-party agent; residency/subprocessor list; retention and eDiscovery exports; failure-mode stories for incorrect agent actions; and commercial terms that do not hide model/platform fees inside vague “AI success” bundles. Run a bounded pilot with read-mostly permissions first.

Applorable take

Treat the Harness as a signal that enterprise AI buyers will keep demanding control planes, not more demos. For now, inventory agents you already have, map them to the six capability headings, and press Salesforce (and rivals) for dated delivery and clear packaging. Until FY28 unified bits and price cards firm up, keep production write-actions supervised — enthusiasm is not a control. Pair any Salesforce briefing with the same questions for Microsoft, ServiceNow, or niche agent platforms you already run, so you are comparing architectures rather than keynote slides. Document who owns agent incident response before the first autonomous update touches a customer record.