{
  "id": 20630021,
  "title": "Which of Your Agent's Actions Can't Be Undone?",
  "url": "https://urgent.news/2026/11/18/which-of-your-agents-actions-cant-be-undone",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-18T17:38:38.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/rick_leffke_6541e28129e2a/which-of-your-agents-actions-cant-be-undone-4dei"
  },
  "original_language": "en",
  "account": "An AI agent with the ability to issue refunds, cancel orders, send emails, or move money poses significant risks, as the outcomes of its actions cannot always be reversed once they have been executed. To mitigate these risks, it is crucial to list every action the agent can perform and identify those that cannot be undone or require significant effort to reverse. Every action the agent can take should be explicitly documented, rather than relying on vague job titles such as \"support,\" \"booking,\" or \"following up with leads.\" Each action should be clearly defined, such as the agent's ability to answer questions, look up orders, start returns, issue refunds, cancel subscriptions, change addresses, send emails, or create credits. If the agent utilizes tools, those actions should also be included in the list. After compiling a comprehensive list, each action should be evaluated to determine whether it can be fully reversed in a timely manner without causing harm to the customer or the business. Actions that cannot be easily reversed should be marked for requiring human sign-off before execution. This includes actions that have already been processed, such as issued refunds, canceled orders, received emails, or completed payments. The rationale behind this approach is to ensure that only actions with a high risk of negative consequences require human oversight. While reversible actions can proceed without immediate human intervention, irreversible actions must undergo a review process. Designated individuals should be identified to approve such actions, and clear communication channels should be established to facilitate timely approvals. A reasonable time limit should be set for review, with a default safe choice of \"wait\" if no approval is received. Logging all irreversible actions, including who approved them, is essential for accountability and troubleshooting. Before deploying the agent, it is advisable to test its permissions extensively. Conducting tests to ensure that the agent cannot bypass the sign-off process is crucial. If the agent can circumvent the necessary approvals, the permission scope needs to be tightened further. The ultimate goal is to strike a balance between efficiency and safety, ensuring that the majority of queries can be handled without human intervention while maintaining a robust oversight mechanism for the few actions with irreversible consequences. Businesses should create a formal list of all actions their AI agents can perform and conduct thorough risk assessments to determine which actions necessitate human sign-off. If the necessary information is lacking, it indicates a gap in understanding the agent's capabilities. By following this structured approach, organizations can harness the benefits of AI agents while minimizing the potential for irreversible harm.",
  "summary": "An AI agent that only drafts a reply is one kind of risk. An agent that can issue a refund, cancel an order, send an email or move money is another. The second kind can create a result you cannot take back, or cannot take back easily, before a person has even seen it. The practical move is simple, and it should happen before the agent goes live: list every action it can take, mark the ones you…",
  "key_points": [
    "AI agents can perform actions like refunds, cancellations, emails, and money transfers.",
    "Not all actions of AI agents can be easily undone or reversed.",
    "Irreversible actions require human sign-off before execution."
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}