Some actions cannot be taken back. A payment is settled, a contract is signed, a record is deleted, and no apology or follow-up message reverses them. When an agent can trigger these outcomes, the last line of defense cannot be a probabilistic model decision.
Agent instructions and topic classifications guide behavior, but they are not guarantees. A misclassified utterance, an injected prompt, or a hallucinated variable can push an agent past a boundary it was only told to respect. Irreversible decisions need enforcement that does not depend on model judgment.

Separate the reversible from the irreversible
- Classify every action the agent can invoke by blast radius: fully reversible, costly to reverse, or irreversible.
- Treat money movement, contract changes, external communications, and deletions as irreversible by default.
- Enforce irreversible paths in deterministic automation such as Flow, Apex, or validation rules, not in prompt text.
- Route high-impact decisions through a deliberate human gate when automation alone cannot carry the accountability.
Build gates into the execution path
- Put the enforcement logic inside the Flow or Apex the action calls, so the gate holds no matter which topic or instruction triggered it.
- Use variables and action filters to block execution when inputs fall outside approved thresholds, and return a clear refusal to the agent.
- Require explicit confirmation or human approval for irreversible actions, and log who approved, what was approved, and when.
- Test the gates adversarially: attempt deletions, refunds, and sends with out-of-policy inputs and confirm the automation refuses them.
Resources
This article is based on the following Salesforce learning and product documentation.