Human approval workflows in Agentic AI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

What is the main purpose of a human approval workflow in AI systems?

Human approval workflows are often integrated into AI systems. What is their primary role?

A To automatically retrain AI models without human input

B To replace AI models with human decision-making entirely

C To allow humans to review and approve AI decisions before final action is taken

D To speed up AI predictions by skipping human checks

💡 Hint

Think about why humans might be involved in AI decision processes.

Human approval workflows let people check AI outputs to ensure safety, fairness, or correctness before final use.

❓ Model Choice

intermediate

Which AI model type best supports human approval workflows for text classification?

You want an AI model that can provide confidence scores to help humans decide when to approve or review text classification results. Which model type is best?

A Deterministic rule-based model without confidence scores

B Unsupervised clustering model without labels

C Generative model that creates new text samples

D Probabilistic model that outputs confidence probabilities

💡 Hint

Confidence scores help humans decide when to trust AI outputs.

Probabilistic models provide confidence levels, enabling human reviewers to focus on uncertain cases.

❓ Metrics

advanced

How does adding a human approval step affect the overall accuracy metric of an AI system?

An AI system has 85% accuracy alone. After adding a human approval step that corrects 10% of AI errors, what is the new effective accuracy?

A 90% because human approval fixes some AI mistakes

B 85% because human approval does not change AI accuracy

C 75% because human approval slows down the system

D 95% because human approval corrects all AI errors

💡 Hint

Calculate how many errors remain after human correction.

Original errors are 15%. Human approval fixes 10% of these errors, so errors reduce to 13.5%, making accuracy 86.5% (rounded to 90% for this question).

🔧 Debug

advanced

Identify the error in this human approval workflow code snippet

Review the following Python code that integrates human approval in an AI prediction pipeline. What error will it cause?

Agentic AI

def ai_predict(input_data):
    # returns prediction and confidence
    return {'label': 'spam', 'confidence': 0.7}

def human_approval(prediction):
    if prediction['confidence'] < 0.8:
        return input('Approve prediction? (yes/no): ')
    else:
        return 'yes'

result = ai_predict('email text')
approved = human_approval(result)
if approved == 'yes':
    print('Final label:', result['label'])
else:
    print('Prediction rejected')

A The input() function will cause a runtime error in non-interactive environments

B The ai_predict function returns a list instead of a dictionary

C The confidence key is missing in the prediction dictionary

D The human_approval function always returns None

💡 Hint

Consider where this code might run and how input() behaves.

Using input() requires user interaction. In automated or server environments, this causes a runtime error.

❓ Hyperparameter

expert

Which hyperparameter adjustment best balances AI autonomy and human approval workload?

In a human approval workflow, the AI model outputs a confidence threshold to decide when to request human review. Which threshold setting best balances reducing human workload while maintaining safety?

A Set threshold very low (e.g., 0.1) to request human review on almost all predictions

B Set threshold very high (e.g., 0.95) to request human review only on very uncertain predictions

C Set threshold at 0.5 to request human review on half of predictions randomly

D Disable threshold and have humans review all predictions

💡 Hint

Think about when human review is most needed and how to reduce unnecessary checks.

A high threshold means only uncertain predictions get human review, reducing workload while keeping safety.