Free AB-100 practice test

Correct: AI Governance is the CoE element responsible for ethical oversight, regulatory compliance, and responsible AI enforcement. It establishes review boards, standards, and accountability frameworks that all AI projects must adhere to. | Wrong - AI Enablement: Enablement focuses on democratizing AI capability through training and shared tooling, not on enforcing compliance or ethical guardrails. | Wrong - AI Innovation Lab: The Innovation Lab experiments with new AI capabilities but operates in sandboxed contexts; it is not the accountability body for production compliance. | Wrong - AI Platform Engineering: Platform Engineering maintains infrastructure and pipelines to run AI workloads efficiently; it is a technical function, not a governance or ethics body.

Correct: The Ready phase of the CAF establishes the Azure landing zone, security baselines, and governance guardrails that must be in place before deploying workloads — including AI workloads — to production. | Wrong - Manage: Manage is a post-deployment phase focused on operations and monitoring of running workloads, not preparation before deployment. | Wrong - Innovate: While Innovate involves building AI solutions, deploying directly to production without completing Ready skips critical governance and security controls. | Wrong - Retire: Retiring legacy systems is a separate workload-lifecycle activity that may run in parallel but is not a prerequisite gate in the adoption sequence.

Correct: Enabling generative answers and connecting the firm's SharePoint repository as a knowledge source is the correct Copilot Studio pattern for grounding agent responses in proprietary content, ensuring accuracy and relevance specific to the firm's case law. | Wrong - Classic topic flows only: Pure topic/keyword flows cannot dynamically synthesize answers from large unstructured document repositories and would require manually authoring responses for every possible query. | Wrong - Generative AI without knowledge sources: Without grounding in firm-specific documents, the agent would rely on the LLM's general training data, which does not contain the firm's proprietary case law and would produce hallucinated or irrelevant answers. | Wrong - Power Automate email flow: Emailing documents per query is not a supported or practical knowledge retrieval pattern; it introduces latency, security risks, and does not integrate with the agent's response generation.

Correct: A custom SLM fine-tuned on domain-specific sensor telemetry and deployed on edge hardware meets all three requirements: domain specificity (CNC anomalies), ultra-low latency (sub-10ms on-device inference), and air-gapped operation (no internet required). | Wrong - GPT-4o via cloud API: Cloud API calls introduce network latency far exceeding 10ms and require internet connectivity, making this incompatible with the stated constraints. | Wrong - Prebuilt M365 Copilot agent: Microsoft 365 Copilot is designed for knowledge-worker productivity tasks, not industrial sensor anomaly detection, and operates via cloud with significant latency. | Wrong - Prompt library with cloud-hosted model: A prompt library still requires round-trips to a cloud-hosted model, violating both the latency and connectivity constraints.

Correct: Designing a unified data layer using Microsoft Dataverse (the common data platform for Dynamics 365) and Fabric-connected pipelines ensures cross-app data is harmonized into a single customer context, enabling the AI to surface holistic insights without data silos. | Wrong - Three separate AI models with no sharing: Siloed models cannot correlate data across apps, making cross-domain insights (e.g., linking a delivery delay in Supply Chain to a support ticket in Customer Service) impossible. | Wrong - Limiting to one app: Restricting AI to one app defeats the purpose of a cross-application AI solution and leaves critical context from the other apps unavailable. | Wrong - Manual copy-paste by agents: Manual aggregation is slow, error-prone, and not scalable, directly contradicting the goal of AI-powered automation and real-time insight delivery.

Correct: A centralized, indexed knowledge store with standardized metadata and API endpoints follows data-sharing best practices, enabling other AI systems to discover, query, and consume the content programmatically. | Wrong - Password-protected PDFs in mailboxes: This siloes data inside individual inboxes, making it inaccessible to AI systems that lack per-user credentials and cannot parse unstructured attachments at scale. | Wrong - Ephemeral in-memory summaries: Ephemeral data cannot be retrieved by other systems after the session ends, defeating the purpose of organizational knowledge sharing. | Wrong - Proprietary desktop databases: Proprietary, connectivity-limited databases create integration barriers that prevent other AI systems from accessing the data without expensive custom connectors.

Microsoft 365 Copilot for Sales is specifically designed to surface CRM insights (including Dynamics 365 Sales data) in Outlook and Teams. The correct setup involves installing the Copilot for Sales app from the Microsoft 365 admin center and establishing the CRM connection so it can pull real-time deal context. Option B (custom Copilot Studio agent) is a custom development path that duplicates existing Copilot for Sales functionality unnecessarily. Option C (Sales Insights dashboard as Teams tab) provides static reporting access, not the embedded, contextual AI-driven insights that Copilot for Sales delivers within Outlook/Teams. Option D (Power BI in Outlook) is a reporting surface, not an AI-driven conversation and insight experience.

CLU with a custom-trained model is best suited for this scenario. When there are hundreds of fine-grained, domain-specific intents that require precise discrimination (e.g., firmware reset vs. remote reboot), a CLU model trained on domain-specific utterances provides superior accuracy over standard NLP trigger phrases and more deterministic routing than pure generative AI. Option A (standard NLP trigger phrases) does not scale reliably to hundreds of similar technical intents and is prone to conflicts. Option C (generative AI orchestration exclusively) introduces non-determinism and latency for a use case requiring precise, auditable routing to specific backend systems. Option D (hybrid without CLU) foregoes the precision of a purpose-trained CLU model, which is specifically designed to handle dense, similar intent sets in enterprise contact center scenarios.

An autonomous agent with a structured agent flow is the correct design. Autonomous agents in Copilot Studio are purpose-built for multi-step, goal-directed workflows that operate without continuous human input, and agent flows support chaining retrieval, AI evaluation, and system actions like Dynamics 365 connector calls. Option A (prompt-and-response) is too simplistic for a multi-step process and cannot reliably execute system writes. Option C (task agent) is explicitly limited to single-topic deterministic routing and is not suited for autonomous multi-system orchestration. Option D (Microsoft 365 Copilot scoped to Teams) is a collaboration surface, not an appropriate platform for back-office financial workflow automation.

Azure AI Document Intelligence with custom classification, managed through Azure AI Foundry, is designed precisely for classifying unstructured documents with mixed text and tabular content. It supports PDF ingestion and can be fine-tuned on labeled invoice samples. Option B (Power Automate desktop OCR) is a manual, brittle process that does not produce a scalable AI classification model. Option C (Copilot Studio generative answers) is a knowledge retrieval feature, not a document classification training pipeline. Option D (AI Builder object detection) is designed for identifying objects in images, not classifying document content into spend categories.

Dynamics 365 Supply Chain Management's Demand planning module includes native AI-generated forecasting with Copilot integration that provides anomaly detection explanations and recommended actions directly within the planner's workflow—this is the purpose-built, lowest-friction solution. Option A (AI Builder prediction model + canvas app) requires custom model training, data pipeline work, and a separate app surface—significantly more effort for something the platform provides natively. Option C (Azure ML + Dataverse + model-driven app) is a valid pattern for highly custom scenarios but introduces unnecessary infrastructure overhead when a native capability exists. Option D (Copilot Studio + Power BI) is a reporting-layer workaround that does not integrate with Supply Chain Management's planning workflows and lacks the deterministic anomaly detection logic of the demand planning module.

Dataverse auditing captures who created, modified, or deleted records at both the entity and field levels, providing the user-attributed audit trail required by the compliance team. DLP policies control which connectors and data flows are permitted but do not log individual data record changes to specific users. Azure Purview catalog scans classify and discover data assets for governance metadata purposes — they do not produce change-level audit logs. Managed identity assignment governs how the agent authenticates to downstream services but does not itself create an audit trail of data changes.

A systematic analytical approach — exporting transcripts, applying NLP tools to identify misclassification patterns, correlating with built-in analytics, then applying targeted remediations — is the recommended methodology for AI-based issue analysis. It ensures the fix addresses root causes rather than symptoms. Immediately republishing with all misclassified examples as few-shot prompts risks overfitting to historical errors and bypasses proper testing. Disabling the feature entirely is operationally disruptive and unnecessary before root cause analysis is complete. Increasing session timeout and adding a fallback topic addresses unrecognized intent routing, not urgency misclassification — it is a mismatch between the observed symptom and the proposed fix.

Full model versioning — linking each artifact to its training dataset snapshot and evaluation metrics — is the cornerstone of MLOps ALM. It enables auditability (regulators or internal reviews can inspect what data produced which model), reproducibility, and safe rollback if a new version underperforms. Using the same compute cluster for training and inference risks resource contention that could degrade production predictions. Storing model weights in Dataverse is inappropriate; model artifacts belong in blob-backed model registries (Azure ML or AI Foundry) designed for large binary assets. Deploying without a staging validation phase bypasses quality gates, risking production degradation with every quarterly cycle.

Application Insights integration with the Foundry Agents service provides distributed traces that span every orchestration step, including calls to external APIs and Dynamics 365. Reviewing these end-to-end traces pinpoints exactly which step is stalling or contributing to latency, making it the correct first action for root-cause analysis. — Increasing session timeout (B) masks the symptom without diagnosing the cause and would not explain the latency increase. — Dynamics 365 audit logs (D) capture entity-level data changes, not agent orchestration execution flow; they would not reveal which workflow step is failing. — Redeploying to a new region (C) is a disruptive last-resort action that should not precede telemetry analysis, and regional issues would typically affect all workloads, not just agent completion time.

Using Copilot to accelerate test case generation while retaining human review and validation is the recommended approach — it combines AI productivity gains with human accountability for test quality. Fully autonomous test suite generation and direct import without review abdicates human oversight and risks low-quality or incomplete coverage entering the suite. Restricting Copilot only to negative cases is an arbitrary constraint that forgoes productivity benefits for positive scenarios without a sound rationale. Having Copilot autonomously execute tests and judge pass/fail conflates authoring assistance with autonomous quality assurance — Copilot is not a substitute for a proper test execution framework with deterministic assertions.

This design covers all three audit surfaces with immutability: the ML Model Registry provides versioned, metadata-rich model provenance; ADLS with immutability policies (WORM) and Entra ID access logging covers training data change history with tamper evidence; and Dataverse Audit logs forwarded to an immutable Log Analytics workspace provides non-repudiable Dataverse change records. A shared Excel workbook in SharePoint has no immutability guarantees, relies on manual updates, and is vulnerable to human error or intentional alteration. The Azure ML activity log captures API-level events but lacks training dataset versioning and does not integrate Dataverse changes; the recycle bin is a recovery tool, not an audit trail. A Power Automate email-based approach has no immutability, is dependent on flow reliability, cannot be considered non-repudiable, and provides no cryptographic or platform-enforced record integrity.

Storing the agent definition and all related artifacts as infrastructure-as-code in source control with environment-specific parameter files (D) aligns with ALM best practices for the Foundry Agents service. It provides version history, repeatability, auditability, and environment isolation through parameterized pipelines — all key ALM requirements. — Manual JSON cloning and hand-editing (A) introduces human error, lacks auditability, creates no source-of-truth, and is explicitly contrary to ALM automation principles. — Sharing a single Foundry Agents service instance across all environments (B) eliminates environment isolation; a production incident or misconfiguration could corrupt test and development agents, and Azure AI Foundry does not natively support environment-label-based isolation within a single instance for this purpose. — Packaging a Foundry Agents service agent as a Copilot Studio solution file (C) conflates two distinct technologies. Foundry Agents service agents are not packaged or deployed via Power Platform solution files; this approach applies to Copilot Studio agents, not Foundry Agents service.

The correct approach uses the platform's native authorization mechanisms: SharePoint permissions restrict document access at the site/library level, Dataverse row-level security restricts record access by business unit, and user identity pass-through ensures that the underlying data APIs enforce access policies based on who is actually calling — not a privileged service account. Using a global service account and relying on the system prompt is a critical anti-pattern: the model can be prompted to override those instructions, and it provides no real security boundary. DLP policies govern connector usage categories (e.g., whether a connector is allowed), not fine-grained data row access. Separate agent instances per BU is an operationally expensive architecture and does not use built-in security features; it also cannot prevent a user from accessing the wrong agent instance unless identity-based routing is separately implemented.

Restricting the agent's service account via Dataverse table-level and column-level security (B) enforces a data-layer access control that is independent of the model's behavior. If the agent's identity cannot read the supplier pricing table, no prompt—however crafted—can cause the agent to retrieve and return that data. This is the correct defense-in-depth approach for prompt manipulation mitigation. — Relying solely on a system prompt instruction (A) is insufficient; prompt injection attacks are specifically designed to override or bypass LLM-level instructions, so this cannot be the primary control. — Blocking all numeric values via content moderation (D) would produce extensive false positives (e.g., product SKUs, quantities) and would still not prevent the data from being fetched; it only filters the output, not the retrieval. — BYOK encryption (C) protects data at rest from infrastructure-level access but does not affect what an authenticated service account can query through the Dataverse API; the agent would still be able to decrypt and read the data at runtime.