Adding a Human-in-the-Loop Approval Step Before the Agent Acts

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-builders-agentic-ai-human-in-loop-r9a8-teen

1. What is the primary purpose of adding a human approval step before an AI agent performs an action?

   1. To prevent mistakes or irreversible actions before they happen
   2. To comply with government regulations about AI systems
   3. To make the agent run faster by offloading work to humans
   4. To give the agent time to learn from human feedback

2. Which of these actions should definitely require human approval before an AI agent executes it?

   1. Generating a summary of a long document
   2. Correcting spelling errors in a draft
   3. Formatting a spreadsheet with new colors
   4. Sending money from a bank account

3. According to the 95% automation / 5% gating principle, what portion of a workflow should be automated versus require human approval?

   1. Only dangerous actions should be automated
   2. Exactly half automated and half approved by humans
   3. Most of the repetitive work automated, with approval only for risky or irreversible steps
   4. 95% of all actions should require human approval

4. An AI agent is designed to automatically delete email accounts that haven't been used in 2 years. Why would you recommend adding an approval step?

   1. The AI might delete accounts that are still needed or have important data
   2. Deleting accounts is a simple task that doesn't need human input
   3. Approval steps always slow down agents unnecessarily
   4. Deleting accounts is reversible so no approval is needed

5. Why do AI safety experts recommend keeping humans 'in the loop' for certain AI agent tasks?

   1. Because AI agents are too slow to make decisions
   2. Because humans can approve actions faster than AI can
   3. Because it's required by most computer programming languages
   4. Because humans provide oversight that catches errors the AI might miss

6. A code-merging AI agent wants to automatically merge pull requests that pass all tests. What's the main risk?

   1. Tests might fail randomly causing merge issues
   2. Merging code is against copyright law
   3. Tests take too long to run
   4. The code might have security bugs or logical errors that tests don't catch

7. A scheduling AI agent automatically sends meeting invites. What approval improvement would make it more trustworthy?

   1. Pre-fill the meeting invite but wait for human confirmation before sending it
   2. Let the AI send invites but allow people to decline after
   3. Have the AI send invites only during business hours
   4. Have the AI check calendar availability before sending invites

8. An invoice-paying AI agent automatically pays any invoice under $100. What's the problem with this approach?

   1. The AI might pay duplicate, fraudulent, or incorrect invoices without verification
   2. Human approval is only needed for very large amounts
   3. Invoices under $100 don't need to be paid
   4. Small invoices are the most important to pay quickly

9. A student uses an AI to write email responses to teachers. What's the safest workflow?

   1. Have the AI send emails only to math teachers
   2. Let the AI send emails directly since it writes well
   3. Let the AI write emails but delete them after 24 hours
   4. Have the AI write the email and show it to the student before sending

10. Why might an AI agent that corrects spelling need human review before sending the document?

    1. Humans cannot spell as well as AI
    2. Spelling correction should never be automated
    3. The AI might change the meaning of sentences while fixing spelling
    4. Spelling correction is too complex for AI

11. What happens when AI agents perform actions without any human approval step?

    1. The actions become reversible
    2. The actions always turn out perfectly correct
    3. Errors can propagate immediately without any human to catch them
    4. The agents will refuse to work

12. How does adding approval steps change how users view an AI agent's trustworthiness?

    1. It shows the agent respects human control and can be relied on safely
    2. It makes the AI more likely to make mistakes
    3. It makes the agent seem slower and less capable
    4. It has no effect on trust

13. An AI agent suggests code improvements that always pass tests. Why should a human still review before applying them?

    1. Tests are perfect and never miss bugs
    2. The improvements might break other parts of the system or introduce new problems
    3. AI code suggestions are always wrong
    4. Human review is required by law

14. What does it mean to 'gate' an AI action?

    1. To make the AI run faster by skipping steps
    2. To require a human to pass through a checkpoint before the action proceeds
    3. To teach the AI to do a new task
    4. To block the AI from performing any actions

15. A team wants their AI agent to generate weekly reports. What's the best approval approach?

    1. Only approve if the report is longer than 10 pages
    2. No approval needed since reports aren't dangerous
    3. Generate the report automatically but require approval before sending it to recipients
    4. Have the AI generate and email the report, then ask for feedback afterward