An automotive manufacturer's Amazon Bedrock-based chatbot handles customer service for dealerships nationwide. The AI governance steering committee wants to strengthen the production risk controls around it. Select TWO actions the committee should prioritize. (Select TWO.)
Select all correct answers, then click Submit.
Short Explanation
You need both a tripwire and a spot-check to keep this chatbot honest at scale: escalation thresholds catch a weird answer the moment it happens, and a recurring audit of sampled conversations catches the slow drift a single flagged response would never reveal. Neither control does the whole job alone, which is why your committee should prioritize them together.
Full Explanation
Establishing a recurring cadence for auditing a sample of chatbot conversations against quality and compliance standards, paired with defined escalation thresholds that trigger human review when outputs deviate from expected patterns, are the two proven production controls here: the thresholds catch risky individual responses in real time, and the recurring audits catch quality or compliance drift that a one-time launch review would miss. Requiring every dealership to independently retrain the underlying model on local sales data is out of scope for a business governance committee and would create inconsistent, ungoverned versions of the same system across locations rather than strengthening oversight of the one system in production. Publishing the chatbot's system prompt to all dealership staff sounds like a transparency measure, but broadly distributing it doesn't monitor or control anything the chatbot actually does, and it could create a new risk if the prompt contains sensitive instructions or business logic. Increasing the chatbot's response length addresses a customer-experience preference, not a risk control, and longer responses could just as easily increase the chance of an error or a non-compliant statement slipping through. The operational check: strong production controls combine a real-time trigger for unusual output with a recurring sample-based review, not a one-off technical tweak or a transparency gesture that doesn't actually monitor anything.