Agentic AI on Microsoft — Lesson 3: Evaluation, Security and Responsible AI

Learning Objectives

• Evaluate accuracy, task completion, groundedness, safety, bias, privacy, robustness, and cost
• Apply least privilege, data protection, audit, red-teaming, human review, and failure containment
• Validate evaluation, security and responsible ai decisions with documented requirements, test evidence, risk controls, and measurable acceptance criteria

Detailed Topics

• Evaluate accuracy: Evaluate accuracy, task completion, groundedness, safety, bias, privacy, robustness, and cost
• Apply least privilege: Apply least privilege, data protection, audit, red-teaming, human review, and failure containment
• Validate evaluation: Validate evaluation, security and responsible ai decisions with documented requirements, test evidence, risk controls, and measurable acceptance criteria

Guided Demonstration

1. Review the requirement, users, data, dependencies, permissions, risks, and expected result.
2. Demonstrate the relevant configuration, creation, development, administration, analysis, or support workflow step by step.
3. Capture settings, outputs, logs, screenshots, calculations, test results, and other useful evidence.
4. Compare the observed result with acceptance criteria and explain any corrective action.

Hands-On Activity

Complete an agent evaluation and risk assessment.

• Document assumptions, prerequisites, and the planned sequence before starting.
• Complete the activity and preserve evidence of normal, boundary, and failure behavior.
• Record issues, corrective actions, reusable notes, and the next improvement priority.

Scenario Analysis

• Identify the business or technical decision that must be made.
• Compare valid implementation approaches and their trade-offs.
• Consider security, privacy, accessibility, performance, support, and operational impact.

Troubleshooting Focus

• Separate symptoms from causes and collect evidence before applying a fix.
• Check configuration, data, permissions, dependencies, capacity, connectivity, and recent changes.
• Validate rollback, recovery, ownership, user impact, and business continuity.

Validation Checklist

• The requirement, scope, inputs, and expected outcome are clear.
• The selected design or configuration is documented and follows approved standards.
• Normal, boundary, failure, security, accessibility, and operational conditions are reviewed.
• The result is supported by evidence and the next improvement is prioritized.

Lesson Outcome

Learners can explain and apply Evaluation, Security and Responsible AI in a realistic scenario and validate the result with documented evidence.