You are a program manager at a 180-person logistics firm piloting a Copilot agent to draft replies to customer delivery complaints using past tickets. Legal requires masking customer addresses and phone numbers, support leads want consistent tone for all regions, and no reply may go out without human approval. Which Responsible AI-aligned agent setup should you require to meet fairness, privacy, and accountability together?
Select an answer to reveal the explanation.
Short Explanation
You're covering all three bases at once, and that's the win. Redaction protects privacy, standard prompts keep things fair, and logged approval keeps humans in charge — your guardrails work like seatbelts for every trip.
Full Explanation
The required setup jointly satisfies fairness, privacy, and accountability because PII redaction removes customer addresses and phone numbers from model input, drafts, and telemetry before past tickets are used, standardized prompts enforce a single tone across all regions rather than reproducing regional variation, and mandatory human approval with logged reviews ensures no reply is sent autonomously while creating an accountability record for each draft. The role-based ticket access approach fails on privacy because access control restricts who can open tickets but leaves sensitive fields in prompts, outputs, and logs for processing. The feedback-tuned prompts approach fails on fairness because continuously updating prompts from ratings allows tone to drift and personalize by queue and region, violating the explicit consistent-tone requirement. The batched audit logging with sampled approval approach fails on accountability because post-release sampling is detective governance for low-risk automation, not preventive control, violating the requirement that every reply receive human approval before sending. As an exam caveat, prioritize control strength implied by the stem over generally good practices, so preventive pre-send approval outweighs detective audit trails and explicit masking outweighs encryption or access control when removal is required. As a practical pilot check, submit test complaints containing addresses and phone numbers, verify redaction in the trace, confirm one pinned prompt version executes in every region, and confirm sending remains blocked until an approval event is logged.