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

 

Question 321

What is the primary purpose of an AI agent’s goal?

1. Define the desired outcome the agent should accomplish
2. Store authentication credentials
3. Increase model context
4. Replace the knowledge source

Correct Answer: 1

Explanation

An agent goal defines the outcome that the agent is expected to accomplish. A clear goal helps determine which tasks, tools, data sources, and decisions are relevant to the workflow. It also provides a basis for evaluating whether the agent successfully completed its assigned responsibility. Goals should be specific enough to guide behavior without unnecessarily expanding the agent’s scope. They should align with business requirements and be supported by appropriate permissions, validation, monitoring, and success criteria.

Question 322

Which capability allows an agent to decide which tool is appropriate for a request?

1. Data retention
2. Tool selection
3. Disaster recovery
4. Model encryption

Correct Answer: 2

Explanation

Tool selection allows an agent or orchestration layer to determine which available capability should be used for a particular request. For example, an agent may choose a CRM lookup tool for customer information and a workflow tool for creating a service ticket. Accurate tool descriptions, schemas, permissions, and clear instructions help improve selection. Tool selection should not provide unrestricted access. Each selected tool should still enforce authentication, authorization, validation, and business rules before executing consequential operations.

Question 323

An agent receives a request containing an unknown customer ID. What should happen before retrieving the record?

1. Validate the identifier
2. Grant administrator access
3. Ignore the identifier
4. Disable authentication

Correct Answer: 1

Explanation

Input validation should occur before an agent uses user-provided values in a business-system request. The customer ID should be checked for the expected format, validity, and potentially whether it exists and is accessible to the requesting identity. Validation reduces errors and can help prevent malicious or malformed inputs from reaching backend systems. Authorization must still be enforced separately because a valid customer ID does not automatically mean the requester is permitted to access that record.

Question 324

What is the purpose of API throttling?

1. Increase the number of unrestricted requests
2. Control request volume to protect service capacity
3. Remove authentication
4. Expand model context

Correct Answer: 2

Explanation

API throttling controls the rate at which requests can be sent to a service. It helps protect backend systems from excessive traffic and allows available capacity to be distributed more predictably. AI agents can generate many requests through repeated tool calls, retries, or concurrent workflows, making throttling particularly important. Applications should handle throttling responses gracefully using appropriate backoff and retry strategies. Throttling should be considered alongside quotas, monitoring, scalability, and service-specific limits when designing reliable integrations.

Question 325

Which security approach provides different permissions based on a user’s or service’s assigned role?

1. RBAC
2. RAG
3. Reranking
4. Tokenization

Correct Answer: 1

Explanation

Role-based access control, or RBAC, assigns permissions according to defined roles. Instead of creating separate permissions for every individual operation and user, organizations can establish roles that represent job responsibilities and assign users or workloads to those roles. RBAC can help simplify access management and support least privilege. In AI architectures, the identity used by an agent can receive a role containing only the permissions required for its approved tasks. Additional data-level controls may still be necessary.

Question 326

An agent must use an external service that has strict request quotas. Which design consideration is important?

1. Quota management
2. Unlimited retries
3. Removing monitoring
4. Maximum prompt length

Correct Answer: 1

Explanation

Quota management helps ensure that an agent stays within the usage limits imposed by an external service. Architects should understand request limits, concurrency restrictions, token quotas, and other provider-specific constraints. The application can use rate limiting, request prioritization, caching, batching, and controlled retries to manage consumption. Monitoring is important for detecting approaching limits. Ignoring quotas can cause failed requests or service disruptions. Quota management should therefore be included in both normal workflow design and failure-handling strategies.

Question 327

Why is prompt injection considered a security concern for AI agents?

1. It can attempt to manipulate the agent into ignoring intended instructions or boundaries
2. It automatically increases model accuracy
3. It improves API availability
4. It reduces all token costs

Correct Answer: 1

Explanation

Prompt injection occurs when untrusted content attempts to influence an AI system’s behavior in ways that conflict with its intended instructions or boundaries. Attackers may try to make an agent reveal sensitive information, misuse tools, or disregard restrictions. Defenses should include separating trusted instructions from untrusted content, limiting tool permissions, validating actions, filtering inputs where appropriate, and performing adversarial testing. Prompt injection cannot be addressed solely through a single sentence in a system prompt because technical controls are also required.

Question 328

Which approach is appropriate when an AI solution must provide answers from frequently changing internal documents?

1. Static model training only
2. RAG with an updated knowledge source
3. Removing the knowledge repository
4. Increasing model temperature

Correct Answer: 2

Explanation

RAG can provide an AI model with information retrieved from an external knowledge source at runtime. This makes it suitable for scenarios where internal documents change frequently and the system needs access to current approved information. The knowledge pipeline should include appropriate ingestion, indexing, permissions, freshness management, and evaluation. RAG does not automatically guarantee accurate answers. Retrieved evidence should remain relevant and authorized, and generated responses should be evaluated for groundedness and correctness.

Question 329

What is a key purpose of metadata in a retrieval system?

1. Support filtering and provide additional information about retrieved content
2. Replace all embeddings
3. Remove authorization
4. Increase model temperature

Correct Answer: 1

Explanation

Metadata provides additional attributes about indexed content that can be used to improve retrieval and filtering. Examples include document type, department, author, effective date, language, security classification, or business category. Metadata can help ensure that retrieval results meet specific requirements before content is supplied to the model. It can also support freshness and access-aware filtering. Metadata should be maintained accurately because incorrect attributes can lead to irrelevant results or, in poorly designed systems, inappropriate data exposure.

Question 330

Which technique can help an agent work with very long conversations?

1. Context summarization
2. Unlimited context without controls
3. Removing all conversation history
4. Disabling memory

Correct Answer: 1

Explanation

Context summarization can condense earlier conversation information into a smaller representation that preserves important details while reducing context usage. This is useful when conversations become too long for the model’s available context window. The summarization process should preserve critical facts, decisions, constraints, and task state while avoiding unnecessary information. Sensitive information should also be handled according to privacy requirements. Summarization should be evaluated because an inaccurate summary can cause the agent to lose important context or make incorrect decisions later.

Question 331

A company wants to ensure that an agent’s output follows a predefined schema. Which technique is appropriate?

1. Structured output
2. Random sampling
3. Unrestricted generation
4. Data deletion

Correct Answer: 1

Explanation

Structured output instructs or constrains the model to return information in a predefined format. For example, an application can require fields such as customer ID, category, priority, and recommendation. The application can then validate the result before passing it to another system. Structured output improves interoperability between AI components and conventional software. However, valid formatting does not prove that the information is correct. Business rules, authorization checks, and content validation should still be applied before consequential actions.

Question 332

Which factor is important when designing an AI agent for high availability?

1. Redundancy and failure handling
2. Removing all fallback paths
3. Using one single dependency
4. Disabling health monitoring

Correct Answer: 1

Explanation

High availability requires an architecture that can continue operating despite failures in individual components. Depending on requirements, this may involve redundant services, alternative model deployments, resilient data access, health monitoring, controlled retries, and fallback paths. Critical dependencies should be identified so that a single failure does not unnecessarily bring down the entire workflow. Availability requirements should be measurable through objectives such as uptime and recovery targets. Resilience testing can help verify that the architecture behaves appropriately during service failures.

Question 333

What is the purpose of a model evaluation baseline?

1. Provide a reference for comparing future model or configuration changes
2. Guarantee perfect responses
3. Remove the need for testing
4. Store user credentials

Correct Answer: 1

Explanation

An evaluation baseline provides a reference measurement for an AI solution before changes are introduced. Teams can compare future model versions, prompts, retrieval configurations, or orchestration changes against the baseline to identify regressions or improvements. The baseline should use representative test cases and consistent evaluation criteria. Important measures may include accuracy, relevance, groundedness, safety, latency, and cost. Maintaining a baseline supports controlled AI lifecycle management and helps teams make deployment decisions based on evidence rather than assumptions.

Question 334

An AI system must summarize confidential documents. What should be considered before sending the documents to an external model provider?

1. Data handling, privacy, and provider security requirements
2. Only the model’s response length
3. The application’s font size
4. The number of prompt templates

Correct Answer: 3

Explanation

Confidential documents may contain information subject to privacy, contractual, regulatory, or organizational restrictions. Before sending such information to an external model provider, the organization should understand how the provider processes, stores, protects, and potentially retains the data. Data residency, encryption, contractual terms, access controls, and approved-use policies may also matter. Where appropriate, the architecture can minimize or transform sensitive information before processing. External AI services should be assessed against the organization’s security and compliance requirements before production use.

Question 335

Which practice can reduce unnecessary sensitive information in application logs?

1. Data minimization
2. Full-content logging of every request
3. Unlimited retention
4. Public log access

Correct Answer: 1

Explanation

Data minimization in logging means capturing only the information necessary for operational, security, auditing, or troubleshooting purposes. AI applications may process sensitive prompts, retrieved documents, and tool results, so recording all content without restrictions can create additional privacy and security risks. Teams can use redaction, masking, selective logging, access controls, and appropriate retention periods. Logs should still contain enough information to diagnose important problems. The goal is to balance observability with responsible handling of sensitive information.

Question 336

Why should AI-generated recommendations sometimes require human review?

1. Some decisions can have significant consequences
2. Models always require manual approval
3. Human review increases token limits
4. It removes the need for testing

Correct Answer: 1

Explanation

Human review can provide additional oversight when AI-generated recommendations may lead to significant financial, legal, operational, safety, or customer consequences. The AI can assist by analyzing information and preparing a recommendation, while an authorized person reviews the result before an action is taken. The required level of oversight should reflect the risk of the workflow. Automated actions can remain appropriate for lower-risk scenarios. Approval requirements should be clearly defined and technically enforced where necessary.

Question 337

Which component can connect an AI solution to data stored in Dataverse or another business application?

1. Connector
2. Embedding
3. Temperature setting
4. Context window

Correct Answer: 4

Explanation

Connectors provide standardized ways for applications, agents, and workflows to interact with supported business data and services. In Microsoft environments, connectors can help integrate AI solutions with systems such as Dataverse and other business applications. Access should still be governed by authentication, authorization, and data policies. A connector does not automatically make all connected data available to an agent. Architects should define the required operations and restrict access to only the resources necessary for the intended business scenario.

Question 338

What is an important reason to separate development, test, and production environments?

1. To reduce the risk of untested changes affecting production
2. To provide everyone administrator access
3. To eliminate version control
4. To prevent evaluation

Correct Answer: 1

Explanation

Separating development, test, and production environments allows teams to build and validate changes without immediately affecting real users or business operations. Developers can experiment in development, while controlled evaluation and integration testing can occur in test environments before approved changes are deployed to production. This separation supports access control, change management, versioning, and safer releases. Production should contain only approved configurations and components. Environment separation is especially important for AI solutions because small changes can produce unexpected behavioral differences.

Question 339

Which practice helps ensure that an AI solution remains aligned with organizational policies after deployment?

1. Ongoing governance and monitoring
2. One-time testing only
3. Unlimited agent permissions
4. Removing audit records

Correct Answer: 4

Explanation

Ongoing governance and monitoring help organizations verify that an AI solution continues to operate according to established policies after deployment. Models, prompts, data sources, user behavior, and integrations can change over time, creating new risks or performance issues. Governance can include periodic reviews, access checks, evaluation, incident management, policy updates, and documentation. Monitoring provides operational evidence that supports these activities. A one-time approval is not sufficient for systems that continue to evolve and interact with changing data and users.

Question 340

What should an organization do when an AI agent consistently fails an important business requirement?

1. Deploy it more broadly
2. Ignore the failures
3. Investigate the cause and improve or redesign the solution
4. Remove all evaluation criteria

Correct Answer: 3

Explanation

Consistent failure against an important business requirement indicates that the solution needs investigation before broader deployment. Teams should identify whether the problem comes from the model, prompts, retrieval, tools, data quality, workflow design, permissions, or another component. Evaluation results and telemetry can help isolate the cause. The team may then improve, retrain, reconfigure, or redesign the solution and test it again. If the requirement cannot be met reliably, the solution should not be treated as production-ready simply because it performs well in other areas.