Microsoft Certified: Agentic AI Business Solutions Architect (beta) Advanced Practice Exam: Hard Questions 2025

You are designing an agentic AI platform for a regulated bank. The agent must autonomously plan multi-step workflows (loan servicing) across CRM, core banking APIs, and document systems. Requirements: (1) strict separation between planning and execution, (2) all tool invocations must be policy-checked and logged, (3) the agent must gracefully degrade to human approval when confidence is low or policy is ambiguous, and (4) the design must support multiple specialized sub-agents (documents, payments, customer comms) without letting them call tools directly. Which architecture best meets these requirements?

A retail enterprise is building an agent that can update inventory and issue refunds. An incident occurs: the agent repeatedly calls the refund API in a loop after receiving a transient 502 error, creating duplicate refunds. You must redesign to prevent this class of failure while preserving autonomy and resiliency. Which approach is best?

Your organization must support an agent that performs long-running, stateful investigations (hours) across multiple systems. Requirements: (1) the agent must resume after failures without repeating completed tool calls, (2) you need a full audit trail of decisions and actions, and (3) you must support parallel sub-tasks (e.g., gather evidence from email, tickets, and logs) while maintaining consistent state. Which design pattern best fits?

You are implementing a customer-support agent that can propose and execute account changes. The business requires that the agent never performs high-impact actions (e.g., account closure) unless it has verified the user’s authorization using enterprise identity and has captured an explicit confirmation step in the conversation. The agent must still handle low-impact actions autonomously. What is the best implementation strategy?

A sales ops agent must generate quotes and sometimes apply negotiated discounts. Policy: discounts above 10% require manager approval; discounts above 20% are prohibited. The agent often receives ambiguous customer emails. You want to minimize unnecessary approvals while preventing policy violations. Which approach best balances autonomy and control?

A procurement agent extracts supplier details from PDFs and creates vendor records. A recurring issue appears: the agent intermittently swaps bank account numbers between two vendors when processing batches in parallel. You must fix the root cause without removing parallelism. What change is most likely to resolve the issue?

You are integrating an agent with multiple AI services: (1) a foundation model for reasoning, (2) a search index for enterprise content, and (3) a classifier for routing requests. Users report that the agent answers with outdated policy even though the newest policy document is in the content repository. Investigation shows the search index is eventually consistent and sometimes returns an older version as top result. What is the best mitigation at orchestration time?

A multi-agent solution uses tools that call internal REST APIs. During peak usage, the system experiences cascading failures: tool timeouts cause the agent to re-plan, generating more tool calls, which increases load further. You need an orchestration strategy that stabilizes the system while preserving successful throughput. Which is the best approach?

Your agent handles HR requests and must use enterprise documents that include sensitive PII. Requirement: prevent accidental disclosure of PII in responses while still allowing the agent to use PII internally to complete tasks (e.g., verifying identity). You also need to prove to auditors that safeguards are enforced consistently. Which design is best?

A company deploys an agent that can modify configuration in production. Security asks for defenses against prompt injection from retrieved content and from user inputs. The agent currently uses RAG and tool calling. Which set of controls is the most effective defense-in-depth approach?