Career+: Design Human Escalation for AI Workflows

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-operations-ai-human-escalation-creators

1. What is the core philosophy about escalation in AI workflows presented in this material?

   1. Escalation is a feature that helps the system handle exceptions safely
   2. Escalation should only happen after the AI makes three failed attempts
   3. Escalation is unnecessary if the AI is properly trained
   4. Escalation indicates the AI has failed and should be retired

2. Which of the following is NOT listed as a situation where AI must not decide?

   1. Customer requests a minor invoice correction
   2. Sensitive data involving health or legal matters
   3. Refund request exceeding a monetary threshold
   4. Low confidence in classification

3. When an AI system's confidence falls below the defined threshold, what should it do?

   1. Automatically retry the task five more times
   2. Ask the user to decide whether to continue
   3. Decline to decide and route the work to an appropriate human
   4. Make its best guess to avoid delaying the workflow

4. Why is it important to route escalations to a specific role rather than any available person?

   1. AI systems cannot route to individuals, only departments
   2. Roles ensure the right expertise and accountability for the decision
   3. Random routing causes legal liability issues
   4. Anyone can handle any escalation efficiently

5. What does it mean to 'carry context forward' when escalating to a human?

   1. The human should restart the work from the beginning
   2. The human receives only a one-word summary of the issue
   3. The AI should provide all relevant background so the human doesn't start over
   4. Context cannot be preserved in automated systems

6. Why should escalations be reviewed on a weekly basis?

   1. Escalations should actually be reviewed monthly, not weekly
   2. To identify patterns and improve prompts, policies, or training data
   3. The AI automatically improves itself without human review
   4. Weekly reviews are required by data protection regulations

7. According to the concepts presented, what makes 'silent guesses' dangerous in AI automation?

   1. AI never guesses—it always provides accurate outputs
   2. Guesses are computationally expensive for AI systems
   3. Silent guesses are the preferred behavior for mature systems
   4. They create the illusion of autonomy while introducing hidden risks

8. A classifier cannot confidently choose between two categories for an incoming document. Which escalation trigger applies?

   1. High stakes
   2. Sensitive data
   3. Policy conflict
   4. Low confidence

9. What type of escalation is appropriate when two policy rules appear to conflict with each other?

   1. Route to the process owner
   2. Route to the queue manager
   3. Route to finance or support lead
   4. Route to IT support

10. When should an AI workflow route to an approved specialist rather than a general supervisor?

    1. When sensitive data involving health, legal, personnel, or financial details is involved
    2. When policy rules conflict
    3. When confidence is below threshold
    4. When the customer has complained about service

11. What is the relationship between uncertainty signals and confidence thresholds?

    1. Uncertainty signals replace the need for human oversight
    2. They are unrelated concepts in AI systems
    3. Thresholds only apply to high-stakes decisions, not uncertainty
    4. Confidence thresholds define the uncertainty signals that trigger escalation

12. A customer disputes an AI-generated response and insists on speaking with a manager. Which escalation trigger applies?

    1. Low confidence
    2. User complaint
    3. Policy conflict
    4. Sensitive data

13. What should happen when an AI workflow encounters a situation not covered by its existing rules?

    1. The AI should make a reasonable assumption and proceed
    2. The system should be shut down until new rules are written
    3. The AI should recognize the gap and escalate to a human who can decide
    4. The AI should wait for a human to manually review all future cases

14. Why is it insufficient for an AI to simply output 'I don't know' without routing to a human?

    1. Humans should not be bothered with AI failures
    2. The system must also route the work to ensure the task gets completed
    3. AI systems are never permitted to express uncertainty
    4. Saying 'I don't know' violates privacy regulations

15. What is the primary purpose of establishing explicit confidence thresholds in an AI workflow?

    1. To satisfy customer demands for transparency
    2. To make the system run faster by skipping complex decisions
    3. To replace the need for human review entirely
    4. To create objective criteria for when the AI should escalate versus proceed