The planned Enterprise AI Harness joins business data, actions, security, models, and governance. Some parts exist now, while the unified product is due to start arriving in early fiscal 2028.
New to this? Read it in simple words
- Salesforce introduced a planned system to control many AI agents across a company. It calls the design a Trusted Enterprise AI Harness.
- It joins business data, actions, security, AI models and rules in one system. Salesforce also plans an AI Control Plane, one place to register and watch agents.
- Shared controls can make basic rules more consistent. But wrong permissions or bad data could affect many agents at once.
- The unified product is planned to start arriving in early fiscal 2028. Pricing and exact regional availability will come later.
- AI agent
- An AI that takes steps on its own to finish a task.
- Harness
- In AI, the system around a model that gives it tools, information and rules.
- Fiscal year
- The year a company uses for its accounts. It may not match the calendar year.
The harness joins six jobs around an AI agent
Salesforce calls the design a Trusted Enterprise AI Harness. A harness is the system around an AI model that gives it information, tools, rules, and checks. It matters when an agent can do work, not only write an answer.
The six parts cover context, agency, action, governance, security, and models. Context means approved business information. Agency means planning and memory. Action connects the agent to software and work steps. Governance and security set rules. Model choice lets a company use different AI systems for different tasks.
Salesforce also plans an AI Control Plane. This would be one place to register agents, give them identities, set policy, watch their behaviour, test performance, and follow costs. It is designed to include Salesforce tools and outside AI systems.
Some base technology exists now. Salesforce plans to start rolling out the unified experience in early fiscal 2028.
Shared controls can reduce repeated work
A company may have many agents in sales, service, finance, and operations. If every team builds its own login rules, data links, and safety checks, small differences can create weak points. A shared control layer can make basic rules more consistent.
The design also tries to keep business context separate from one model. Models change quickly, while customer records, contracts, and company rules stay important. A company could change the model without rebuilding every connection.
There is a trade-off. One control layer can make management easier, but it also becomes an important point of failure. Wrong permissions or bad data could affect many agents at once. Companies will need strong change records, separate approvals, and a quick way to stop actions.
Buyers should separate today’s tools from the roadmap
Salesforce says many technologies under the harness are available today. The new capabilities and unified experience are planned to start rolling out in early fiscal 2028. Pricing, packaging, and exact regional availability will come later.
That timing changes how buyers should read the announcement. It is a product direction, not one finished package that every customer can switch on now. Teams should base buying decisions on features they can test today, as Salesforce itself advises.
A useful control plane should answer simple questions: Which agents are running? What can each one read or change? Which person approved the action? What did it cost? Can the company stop it now? Salesforce has described the right areas. Real customer tests will show whether the planned system makes those answers clear.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · September 10, 2026Salesforce introduces the Trusted Enterprise AI Harness Salesforce
- 2Research · September 10, 2026Salesforce’s new Enterprise AI Harness seeks to ground agents in shared business context VentureBeat
We used Salesforce’s announcement for the six parts, control-plane functions, availability, and product limits. We used VentureBeat to confirm the launch and add outside product context. We separate available base tools from planned unified features.