A project manager presents four proposed computer-use agent scenarios to an AI architecture review board. The board must identify which scenario is most appropriate for a computer-use agent and which scenario is most problematic from a reliability and governance standpoint. Which evaluation is correct?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Computer-use agents are the janitor's skeleton key — great for opening doors that don't have proper locks (no API), but you would never hand that key to someone standing next to a bank vault. Using a computer-use agent to click financial wire approvals is dangerous because a screen misread or a visual injection attack could trigger an irreversible financial action with no deterministic safeguard.
Full explanation below image
Full Explanation
Computer-use agents are designed for scenarios where no API, connector, or programmatic interface exists — they are the integration option of last resort for legacy or third-party systems that cannot be connected any other way. The classic appropriate use case is exactly what Option B's first scenario describes: a vendor web portal with no API that requires visual navigation and data extraction, combined with entry into an internal system. This is a high-value automation that would otherwise require a human operator, and the consequences of occasional errors (mis-entered data) are detectable and correctable.
The most problematic scenario is executing financial wire transfer approvals. This case fails on multiple dimensions: (1) financial transactions are irreversible — a misread screen balance or injected instruction could trigger an incorrect transfer that cannot be undone; (2) financial regulatory frameworks (SOX, PCI-DSS) require deterministic, auditable approval mechanisms — a computer-use agent clicking a button based on visual interpretation does not meet this bar; (3) this is a high-privilege action on a high-stakes outcome, making visual prompt injection particularly dangerous.
Option A incorrectly identifies the 'most problematic' scenario. Multi-internal-web-app onboarding form submissions are a legitimate computer-use case — form submissions in internal apps are lower-risk, more reversible, and the only concern is form layout changes breaking the agent's navigation. This is manageable operational risk, not a governance failure.
Option C is incorrect on both counts. SharePoint has Power Platform connectors — using a computer-use agent to click lists is an unnecessary choice when proper connectors exist. And PDF summarization from a network drive does not require computer-use at all; it is a standard knowledge processing task. Neither is a particularly good or bad computer-use scenario — both are simply wrong tool choices.
Option D identifies regression testing of desktop applications as most appropriate — this is a valid use case (QA automation for legacy desktop apps is a genuine computer-use scenario), but competitor web scraping raises legal and policy concerns and is harder to classify as 'most problematic from governance standpoint' compared to financial wire approvals.
Exam tip: Computer-use agent appropriateness is evaluated on two axes: (1) Is there a better (API-based) alternative? If yes, don't use computer-use. (2) Is the action reversible and low-stakes enough that visual interpretation errors are acceptable? If the action is irreversible and high-stakes (financial, medical, legal), computer-use agents are inappropriate.