View Full Microsoft AB-100 Exam Dumps and Practice Test Dumps.

 

Question 381

What is the primary purpose of an agent planning loop?

1. To permanently store every conversation
2. To repeatedly plan, act, observe results, and adjust as needed
3. To replace all backend systems
4. To disable tool usage

Correct Answer: 2

Explanation

An agent planning loop allows an agent to break a goal into actions, execute those actions, observe the results, and determine what should happen next. This approach is useful for tasks where the next step depends on information obtained during execution. The loop should have defined boundaries, termination conditions, and appropriate permissions. Without these controls, an agent could repeatedly execute actions or consume excessive resources. Planning behavior should therefore be tested against expected, unexpected, and failure scenarios before production use.

Question 382

Which feature helps an agent discover available tools and their capabilities?

1. Tool discovery
2. Data residency
3. Token compression
4. Model distillation

Correct Answer: 1

Explanation

Tool discovery allows an agent or orchestration system to identify available tools and understand what each tool is designed to do. Useful tool metadata can include the tool name, purpose, required inputs, output format, permissions, and limitations. Clear descriptions help the agent select an appropriate capability for a task. Discovery should still operate within governance and authorization boundaries. An available tool should not automatically be considered an authorized tool for every user, agent, or workflow.

Question 383

An agent calls an external service that sometimes takes several seconds to respond. Which control prevents the agent from waiting indefinitely?

1. Embedding refresh
2. Data classification
3. Request timeout
4. Prompt versioning

Correct Answer: 3

Explanation

A request timeout establishes the maximum period the application should wait for a service response. If the service does not respond within that period, the application can stop waiting and follow a defined error-handling or fallback process. Timeouts help prevent stalled workflows and improve overall system reliability. They should be selected based on realistic service behavior rather than arbitrary values. Timeout handling can be combined with retries, circuit breakers, and fallback services to create more resilient AI applications.

Question 384

Which practice is most appropriate for handling API credentials that must be rotated regularly?

1. Store them in source code
2. Send them through prompts
3. Use a secure secret-management solution
4. Include them in user-visible responses

Correct Answer: 3

Explanation

Secure secret management provides a controlled location for storing and managing credentials such as API keys and connection secrets. It can support access control, auditing, and credential rotation without exposing secrets in source code or prompts. Applications should retrieve secrets securely when required and should avoid logging them. Where supported, managed identities can eliminate the need for some stored credentials entirely. Secret-management practices should be incorporated into development, deployment, and operational procedures.

Question 385

Which responsible AI principle focuses on protecting personal and sensitive information?

1. Privacy
2. Latency
3. Scalability
4. Throughput

Correct Answer: 1

Explanation

Privacy focuses on protecting personal, confidential, and sensitive information throughout an AI system’s lifecycle. Privacy considerations can include data collection, storage, processing, retrieval, sharing, logging, and retention. AI architectures should collect only information necessary for the intended purpose and apply appropriate access controls. Sensitive data should not automatically be included in prompts, logs, or model inputs. Privacy requirements may also influence model selection, deployment location, retention periods, and integration design.

Question 386

A RAG application returns accurate information but exposes documents that a user is not authorized to view. What failed?

1. Model temperature
2. Access control
3. Prompt formatting
4. Token budgeting

Correct Answer: 2

Explanation

The retrieval system failed to enforce appropriate access control before providing protected documents to the application or user. Retrieval quality alone does not determine whether information should be shown. Enterprise RAG systems should consider the identity and permissions of the requesting user or service when retrieving content. Authorization should be enforced through technical controls rather than relying on the language model to decide what information is permissible. Access-aware retrieval and backend authorization provide important layers of protection.

Question 387

Which technique can reduce unnecessary context while preserving the most relevant information?

1. Context compression
2. Unlimited history
3. Duplicate retrieval
4. Larger prompts

Correct Answer: 1

Explanation

Context compression reduces the amount of information provided to a model while attempting to preserve details that are important to the current task. This can involve summarizing previous conversations, removing duplicate information, selecting relevant passages, or compressing retrieved content. Reducing unnecessary context can lower token consumption and improve response efficiency. However, compression must be evaluated carefully because removing an important detail can affect answer quality. Teams should test compression strategies with representative workloads before applying them broadly.

Question 388

What does query expansion attempt to improve in a retrieval system?

1. The number of user permissions
2. The number of model parameters
3. The range of terms used to find relevant information
4. The duration of secret rotation

Correct Answer: 3

Explanation

Query expansion adds related terms, concepts, or alternative expressions to a search request to improve the chance of finding relevant information. Users may describe a business concept differently from the terminology used in enterprise documents. Expanding the query can help bridge those differences and improve retrieval coverage. However, excessive expansion may introduce irrelevant results. Query expansion should therefore be evaluated using representative searches and measured against retrieval quality requirements such as precision and recall.

Question 389

Why should AI-generated output sometimes be validated against a predefined schema?

1. To ensure the output follows the expected structure
2. To increase employee permissions
3. To replace authentication
4. To disable model evaluation

Correct Answer: 1

Explanation

Schema validation ensures that generated output follows an expected structure, such as required fields, data types, or allowed values. This is useful when an AI response is passed to another application, API, workflow, or database. Without validation, unexpected output can cause processing failures or incorrect downstream actions. A schema does not guarantee that the content itself is factually correct, so business-rule validation may also be necessary. Sensitive or consequential actions should additionally use authorization and appropriate human oversight.

Question 390

Which architecture decision can improve throughput when many independent AI requests arrive simultaneously?

1. Sequentially processing every request
2. Scalable parallel processing
3. Disabling queues
4. Increasing prompt length

Correct Answer: 2

Explanation

Scalable parallel processing allows independent workloads to be handled concurrently rather than forcing every request through a single sequential path. This can improve throughput when demand increases, provided the underlying models, APIs, and infrastructure support the required concurrency. Queues can help absorb temporary workload spikes, while rate limits and quotas prevent excessive service consumption. The architecture should also monitor resource utilization and latency. Scaling should be based on measured workload patterns rather than simply increasing infrastructure without validating the actual bottleneck.

Question 391

What is the main purpose of an evaluation baseline?

1. To provide a reference point for measuring future changes
2. To grant administrative access
3. To store API credentials
4. To increase model context

Correct Answer: 1

Explanation

An evaluation baseline provides a known reference against which later model, prompt, retrieval, or configuration changes can be compared. A baseline can include representative test cases and measurements such as accuracy, groundedness, safety, latency, and cost. When a new version is evaluated, the team can determine whether performance improved, declined, or remained stable. Baselines are especially valuable for regression testing because AI systems can change behavior after seemingly small modifications to prompts, models, data, or retrieval configuration.

Question 392

A company needs to ensure an AI service can recover after a major regional outage. Which area should be addressed?

1. Disaster recovery
2. Prompt formatting
3. Query expansion
4. Document chunking

Correct Answer: 1

Explanation

Disaster recovery defines how an AI solution can restore or continue critical operations after a major outage or infrastructure failure. Planning may include backup strategies, recovery objectives, alternative infrastructure, replicated data, dependency mapping, and tested recovery procedures. AI applications also need to consider model endpoints, knowledge indexes, databases, connectors, and authentication dependencies. A disaster recovery plan should be tested periodically because documented procedures may fail if dependencies or configurations change.

Question 393

Which option is an example of a human-in-the-loop design?

1. An agent automatically deletes records without review
2. A model generates a recommendation that a human approves before execution
3. A system removes all authorization checks
4. An agent executes every action without limits

Correct Answer: 2

Explanation

A human-in-the-loop design includes a person at an appropriate point in the workflow to review, approve, reject, or modify an AI-generated recommendation or action. This can be valuable for high-impact, sensitive, or irreversible decisions. The review point should be clearly defined, and the human reviewer should have enough information to make an informed decision. Human involvement does not replace technical security controls. Authorization, validation, monitoring, and audit logging should still be applied where required.

Question 394

Which Microsoft platform is designed to provide a low-code environment for building AI agents and connecting them to business actions?

1. Microsoft Copilot Studio
2. Microsoft Paint
3. Windows Notepad
4. Microsoft Calculator

Correct Answer: 1

Explanation

Microsoft Copilot Studio provides a low-code environment for creating and configuring AI agents and connecting them to business capabilities. It can support conversational experiences, knowledge sources, actions, connectors, and workflow integrations. Organizations can use it to build agents for specific business scenarios while applying appropriate authentication, governance, and security controls. The platform does not eliminate the need for architecture decisions. Teams still need to define agent scope, knowledge sources, permissions, testing requirements, and operational monitoring.

Question 395

What should an organization review before allowing an agent to access sensitive customer records?

1. Authorization, data protection, and business need
2. Only the agent’s name
3. The screen resolution
4. The number of prompt examples

Correct Answer: 4

Explanation

Before an agent accesses sensitive customer records, the organization should verify authorization, data-protection requirements, and the legitimate business need for that access. Permissions should follow least privilege so the agent receives only the access required for its assigned tasks. Data handling, logging, retention, and regulatory requirements may also need review. The agent should be tested to ensure it cannot expose information to unauthorized users. Sensitive record access should be treated as a controlled business capability rather than a general-purpose model feature.

Question 396

Which approach is useful for reducing cost when a simple task does not require a large model?

1. Route every request to the largest model
2. Use a smaller suitable model
3. Increase the context unnecessarily
4. Duplicate every model request

Correct Answer: 2

Explanation

A smaller suitable model can reduce inference cost and sometimes improve latency for tasks that do not require advanced reasoning or extensive capabilities. Examples may include simple classification, extraction, routing, or straightforward transformations. The smaller model should first be evaluated against the required quality, safety, and latency criteria. Model routing can then direct simple workloads to the lower-cost option while reserving more capable models for complex tasks. This approach can balance business requirements with operational cost.

Question 397

Which activity is part of AI lifecycle management?

1. Versioning, evaluation, monitoring, and retirement
2. Disabling all documentation
3. Removing access controls after deployment
4. Permanently avoiding updates

Correct Answer: 1

Explanation

AI lifecycle management covers activities from initial development through deployment, monitoring, updates, and eventual retirement. Important practices include version control, evaluation, security reviews, monitoring, change management, and documentation. Models, prompts, knowledge sources, and integrations may all require updates over time. A lifecycle approach helps organizations maintain predictable behavior and governance as business requirements change. Retirement should also be planned so unused identities, permissions, data, and infrastructure can be reviewed and removed appropriately.

Question 398

A model provider changes how customer prompts are processed. What should the organization review?

1. Data handling, privacy, security, and contractual requirements
2. Only the user interface
3. Only document font sizes
4. Only the number of employees

Correct Answer: 1

Explanation

Changes in how an external model provider processes customer prompts may affect privacy, security, compliance, data residency, retention, and contractual obligations. The organization should review the provider’s current terms and technical behavior against its requirements before continuing or expanding use. Data sent to external services should be limited to what is necessary, and sensitive information should be handled according to applicable policies. Vendor assessments should also consider service availability, transparency, incident handling, and changes to processing practices.

Question 399

What is the purpose of a pilot before a large AI deployment?

1. To validate assumptions and business value on a controlled scale
2. To avoid defining requirements
3. To eliminate security testing
4. To guarantee production success

Correct Answer: 1

Explanation

A pilot provides a controlled environment for validating whether an AI solution can meet defined technical and business requirements before broader deployment. Teams can evaluate quality, user experience, latency, cost, security, operational workload, and business outcomes using realistic scenarios. A pilot does not guarantee production success, but it can reveal weaknesses and assumptions that were not visible during development. Clear success criteria should be defined in advance so the pilot produces measurable evidence for the next implementation decision.

Question 400

What should an organization do when an AI agent repeatedly fails a defined business requirement?

1. Ignore the failures
2. Increase permissions automatically
3. Investigate the cause and improve the relevant component
4. Remove all evaluation criteria

Correct Answer: 3

Explanation

Repeated failure against a defined business requirement should trigger investigation rather than automatic expansion of the agent’s permissions or deployment. Teams should determine whether the issue originates from the model, prompt, retrieval system, knowledge quality, tool integration, business logic, data, or workflow design. Evaluation results and telemetry can help identify the source of the problem. After making a change, the team should test it against representative scenarios and acceptance criteria before promoting the updated solution into production.