AI & Agentforce
Design safe agent actions
Keep permissions, validation, human review, and failure handling in the workflow.
By Vishal Verma · Reviewed October 5, 2026
Trust features help protect AI interactions, but the application still needs carefully designed access and actions.
- Identify the execution identity and its permitted records, objects, fields, and external services.
- Validate action inputs in Flow, Apex, or the receiving API.
- Require confirmation or human approval for consequential changes where the business process requires it.
- Treat retrieved documents and user text as untrusted input. They must not grant new permissions or override business rules.
- Handle timeouts and unsuccessful responses without claiming that an operation succeeded.
- Log enough information to diagnose failures while protecting sensitive data.
The Einstein Trust Layer provides protections for supported AI features. Confirm the specific capabilities and execution behavior of the features you deploy rather than assuming every integration receives identical protections.