This browser is no longer supported.
Upgrade to Microsoft Edge to take advantage of the latest features, security updates, and technical support.
An Adventure Works agent proposes approving a 350 US dollar return exception for a Gold customer. The agent's confidence score is 0.91, which is above the 0.75 escalation threshold. The action amount is above the 200 USD financial threshold requiring human approval. Should this request escalate?
No—the confidence is 0.91, which is above threshold. Confidence gates take priority, so the agent proceeds autonomously.
Yes—the 350 USD amount exceeds the 200 USD financial action threshold, which independently triggers human approval regardless of the agent's confidence score.
Yes—but only because the customer is Gold tier. Standard customers wouldn't require escalation for this amount.
Adventure Works' approval workflow is implemented with a simple in-memory state machine in the orchestrator agent. On Tuesday afternoon, the Azure Container Apps environment scales down the orchestrator instances due to low traffic. Several pending approval requests are lost. What design change would prevent this?
Increase the minimum instance count for the orchestrator Container App to prevent scale-to-zero during low traffic periods.
Replace the in-memory state machine with a durable workflow service such as Power Automate, which externalizes approval state so workflows survive service restarts and scale events.
Shorten the approval window to 15 minutes to ensure all approvals are processed before scale-down occurs.
Adventure Works reviews the monthly feedback data and finds that 78% of high-confidence rejections (agent confidence > 0.85, human rejected the proposal) cluster around requests to waive the 30-day return window for specific product categories. What is the most impactful next step?
Lower the confidence threshold for all return requests from 0.85 to 0.60 to ensure more requests escalate for human review.
Analyze the system prompt's representation of the 30-day return window policy for those product categories, update the policy language to reflect the actual exception criteria, and validate the fix against the rejection examples.
Remove the product category distinction from the return policy section of the system prompt to simplify the policy the agent needs to interpret.
You must answer all questions before checking your work.
Was this page helpful?
Need help with this topic?
Want to try using Ask Learn to clarify or guide you through this topic?