Microsoft AB-100: Design AI-powered business solutions Practice Questions

Options A, B, and D are correct. Computer Use agents require clear goal and instruction authoring (A) so the underlying vision-language model can navigate UI reliably. A configured target environment—a virtual machine or browser session—(B) is a prerequisite for the agent to have a screen context to operate on. Robust error-handling and confirmation steps (D) are critical because UI-based automation is inherently fragile; the agent must detect unexpected states and handle them gracefully. Option C is wrong—Computer Use does not require modern SPAs and is specifically designed to work with legacy applications. Option E, while AI Builder could assist with document processing, the question describes structured emails, making this optional rather than critical, and it is not a Computer Use-specific design consideration. Option F is wrong—generative answers from SharePoint knowledge sources cannot teach an agent runtime navigation patterns; Computer Use relies on vision-based interaction, not document-grounded Q&A.

Options C and E are correct. Configuring a system fallback topic ensures the agent handles unmatched utterances gracefully and offers escalation rather than failing silently—this is a standard Copilot Studio best practice for Contact Center channels. Enabling generative answers with grounded knowledge sources allows the agent to handle a broader range of queries dynamically, reducing hard fallback frequency. Option A (deleting low-confidence topics) would increase unmatched utterances and degrade user experience. Option B (silently ending sessions) is a poor customer experience anti-pattern. Option D, while useful, is a follow-on implementation detail inside the fallback topic rather than a top-level design practice distinct from Option E—Option E already encompasses the escalation design.

A custom connector with OAuth 2.0 authentication is the correct approach. Custom connectors allow secure, credential-managed connectivity to external REST APIs and can be surfaced in Copilot Studio as plugin actions, enabling the agent to invoke them at runtime without exposing secrets. Option B (Power BI dataflow) introduces unnecessary latency and doesn't provide real-time inventory data. Option C (hardcoded bearer tokens) is a security anti-pattern and tokens would expire, breaking the integration. Option D (adaptive cards redirecting users) does not actually integrate the API data into the agent's response flow.

The on-premises data gateway is the supported Microsoft mechanism for allowing cloud-based services (Power Platform, Power Automate) to securely reach on-premises systems without exposing them to the internet. Wrapping the SOAP call in a Power Automate cloud flow or custom connector that uses the gateway, then surfacing it as a plugin action, is the architecturally sound pattern. Option B is invalid because Copilot Studio agents do not have a built-in HTTP action that can reach non-internet-addressable endpoints, and managed identity does not resolve network routing to on-premises SOAP services. Option C (virtual tables via OData) requires an OData endpoint, which a SOAP-based ERP does not natively expose. Option D introduces unnecessary complexity (Logic Apps + Azure Functions + Blob polling), violates simplicity principles of the Power Platform Well-Architected Framework, and polling a Blob queue is inappropriate for synchronous order status lookups.

Circuit-breaker logic prevents cascading failures by stopping repeated calls to a failing dependency, which is a core Reliability pattern—ensuring the workload continues to function or degrades gracefully under component failure. Security (C) concerns authentication, authorization, and data protection, not failure isolation. Operational Excellence (A) relates to monitoring, deployment practices, and process improvement—not runtime failure handling. Performance Efficiency (D) addresses scaling and resource optimization, not failure resilience patterns.