Microsoft AI Agent Builder Associate (AB-620) — Practice Questions and Mock Exam

Responsible AI principles only become actionable when each one maps to a concrete practice a delivery team can schedule. Fairness attaches to the data and to independent validation of outcomes. Accountability attaches to named people who answer for the system. Ethics attaches to who is in the room during design and to a governance framework that audits the result. Transparency is the one that shapes the user-facing product surface, because it is discharged through disclosure rather than through internal process. In a regulated setting such as healthcare that distinction matters, since a review board will ask what a patient sees and is told, not only what the delivery team did. Planning a responsible AI strategy means allocating each principle to an owner and an artifact rather than treating the four as one undifferentiated commitment.

The three gated extensions differ by where in a deployment they insert their custom step, and picking the wrong one usually means the check runs after the moment it was supposed to protect. Validation that has to happen before anything leaves the development environment belongs at the export boundary, which is why that extension holds the export until your logic signals completion and why it belongs on the first stage only. Running deployments under a different identity is a separate concern handled by the delegated extension, which lets makers request deployments into environments where they have no rights. A final human sign-off after approval belongs to the pre-deployment step. Once artifacts are exported they are stored in the host and cannot be tampered with, and the same artifact flows through every later stage in order, so a later check cannot undo a bad export.

Ownership boundaries decide what a delete button actually deletes, and here the boundary runs between the two products. The knowledge base is a Foundry asset: its sources, its retrieval settings and its lifecycle belong to whoever built it, and it is designed to be reused by several agents. What Copilot Studio holds is a reference plus a tool description. Removing that reference detaches one consumer and leaves the asset untouched, which is exactly what you want when several agents share it. The same asymmetry explains the tuning guidance, since changing ranking behaviour has to happen where the asset lives. Exam tip: when a resource is described as reusable and built elsewhere, expect removal in the consumer to be a detach rather than a delete.

The REST API tool builder is a thin shell over the Power Platform custom connector pipeline, which is why its input requirements look like connector requirements rather than agent requirements. The document has to arrive as JSON, and the pipeline normalises a newer contract down to the version it processes. Everything after the upload is about helping the orchestrator: you pick which endpoint and method combinations should become tools, and you write descriptions specific enough for the model to tell one from another. Input and output value shapes come straight from the specification and are read-only, but their descriptions are yours to write and cannot be left blank. Exam tip: the specification version and file format are separate requirements, and questions often change only one of them.