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

 

Question 261

Which capability allows an agent to transfer a task to another specialized agent?

1. Data indexing
2. Tokenization
3. Agent handoff
4. Model compression

Correct Answer: 3

Explanation

Agent handoff allows one agent to transfer responsibility for a task to another agent that has more appropriate capabilities or permissions. For example, a general customer-service agent could transfer a billing issue to a specialized billing agent. The handoff should preserve relevant task context while enforcing authorization and scope boundaries. Proper orchestration is important so that the receiving agent understands what work has already been completed and what remains. This approach supports specialized multi-agent architectures without requiring every agent to perform every task.

Question 262

What is the primary purpose of embeddings in a RAG solution?

1. Represent content as vectors for similarity-based retrieval
2. Authenticate users
3. Execute business transactions
4. Monitor API availability

Correct Answer: 1

Explanation

Embeddings represent text or other supported content as numerical vectors that capture semantic relationships. In a RAG solution, these vectors can be stored in a vector-capable search system and compared with an embedded user query. This allows the system to retrieve content based on meaning rather than relying only on exact keyword matches. Embeddings do not replace authorization or business-system integrations. Retrieval quality depends on appropriate embedding models, indexing, chunking, metadata, and evaluation against representative queries.

Question 263

An agent must first retrieve an account balance and then use that result to determine eligibility. Which execution pattern is required?

1. Parallel execution
2. Random execution
3. Sequential execution
4. Background execution

Correct Answer: 3

Explanation

Sequential execution is appropriate when one operation depends on the result of another. In this scenario, the agent must retrieve the account balance before it can determine eligibility. Running both operations simultaneously would not work if the eligibility decision requires the retrieved balance. Orchestration should therefore maintain the dependency between the steps. Independent operations can still be performed in parallel where appropriate. Designing workflows according to task dependencies helps improve correctness, efficiency, and predictable agent behavior.

Question 264

Which feature helps ensure an agent cannot perform actions outside its assigned permissions?

1. Larger context window
2. Authorization controls
3. Higher temperature
4. Prompt expansion

Correct Answer: 2

Explanation

Authorization controls determine whether an authenticated identity is permitted to perform a requested action or access a resource. For AI agents, authorization should be enforced at the relevant application, API, data, or tool boundary rather than relying only on natural-language instructions. An agent may understand that an operation is restricted, but technical controls should enforce the restriction. Combining authorization with least privilege, input validation, logging, and approval requirements provides stronger protection against unauthorized actions.

Question 265

What is a major benefit of using managed identity for an AI workload?

1. It removes the need for all permissions
2. It provides secure identity without embedding credentials in application code
3. It makes every API public
4. It eliminates audit requirements

Correct Answer: 2

Explanation

Managed identity provides a way for supported workloads to authenticate to services without requiring developers to store long-lived credentials directly in application code or configuration. The identity can then receive specific permissions through the platform’s access-control mechanisms. This reduces credential-management risks and can simplify authentication between cloud services. Managed identity does not automatically grant access to every resource. Appropriate authorization, least privilege, monitoring, and secret-management practices are still required for the overall AI solution.

Question 266

A tool requires a specific JSON structure as input. What should the agent architecture define?

1. A tool schema
2. A larger vector database
3. A new embedding model
4. A different user interface

Correct Answer: 1

Explanation

A tool schema defines the expected structure, fields, data types, and requirements for tool inputs and, where applicable, outputs. Clear schemas help an agent construct valid requests and reduce failures caused by missing or incorrectly formatted parameters. The application can additionally validate tool inputs before execution to enforce business rules and security requirements. Schemas are particularly important when agents interact with APIs, databases, workflows, or other automated services. Well-defined interfaces make tool integration more predictable and maintainable.

Question 267

Why should tool descriptions be clear and specific?

1. To increase database storage
2. To help the agent select and use the correct tool
3. To disable authorization
4. To prevent all tool calls

Correct Answer: 2

Explanation

Clear tool descriptions help an agent understand what each tool does, when it should be used, what inputs it requires, and what results it returns. When multiple tools are available, precise descriptions can improve tool selection and reduce inappropriate calls. Descriptions should accurately represent the tool’s capabilities and limitations rather than encouraging unrestricted behavior. Tool descriptions should also work alongside schemas, authorization controls, validation, and monitoring. Good tool interfaces make agent orchestration more reliable and easier to maintain.

Question 268

A company wants to reduce repeated database queries for frequently requested information. Which technique can help?

1. Caching
2. Fine-tuning
3. Prompt injection
4. Model retirement

Correct Answer: 1

Explanation

Caching can store previously retrieved results so that repeated requests do not always require another call to the underlying data source. This can reduce latency, API usage, and infrastructure costs for information that remains valid for an appropriate period. The cache must have suitable expiration and invalidation rules because stale information can produce incorrect answers. Sensitive information also requires careful access controls. Caching is most useful when the same data is requested frequently and can safely be reused within defined freshness requirements.

Question 269

Which approach is most appropriate for protecting secrets used by an AI application?

1. Store them in the system prompt
2. Place them in public documentation
3. Use a secure secret-management service
4. Include them in user messages

Correct Answer: 3

Explanation

Sensitive credentials such as API keys, connection strings, and service secrets should be stored using secure secret-management mechanisms rather than prompts, source code, or public files. Secret-management services can provide controlled access, encryption, auditing, and credential rotation capabilities. Applications should retrieve secrets securely when needed and grant access only to authorized workloads. Prompts and user messages are not secure credential stores. Separating secrets from agent instructions also reduces the risk of accidentally exposing credentials through model responses or logs.

Question 270

What should an AI system do when a user request is outside the agent’s defined scope?

1. Perform the action anyway
2. Escalate or clearly communicate the limitation
3. Grant additional permissions automatically
4. Disable all security checks

Correct Answer: 2

Explanation

An agent should not perform actions that fall outside its defined responsibilities simply because a user requests them. Instead, it can explain the limitation, ask for clarification when appropriate, or route the request to a suitable agent or human. Scope boundaries should be supported by technical authorization controls so that the agent cannot bypass them. Clearly handling out-of-scope requests improves user expectations and reduces operational risk. Organizations should define escalation paths for requests that require another system, role, or level of authority.

Question 271

Which technique can improve retrieval when users phrase queries differently from the wording in stored documents?

1. Query rewriting
2. Password rotation
3. Model retirement
4. API throttling

Correct Answer: 1

Explanation

Query rewriting can transform a user’s original request into a form that is more suitable for the retrieval system. The rewritten query can include clearer terminology, relevant concepts, or alternative wording that improves the chance of finding useful documents. This is especially helpful when users use informal language while enterprise documents contain specialized terminology. Query rewriting should not bypass access controls or introduce unsupported assumptions. Its effectiveness should be evaluated using representative queries and retrieval metrics before being adopted broadly.

Question 272

What is the purpose of chunking documents in a RAG pipeline?

1. To divide large content into smaller retrievable units
2. To encrypt every document
3. To remove document permissions
4. To increase API rate limits

Correct Answer: 1

Explanation

Chunking divides large documents into smaller sections that can be indexed and retrieved individually. Smaller, meaningful chunks can help the retrieval system return focused evidence rather than an entire lengthy document. Chunk size and overlap should be selected according to the type of content and expected queries. Poor chunking can separate related information or create excessive fragments. Organizations should evaluate retrieval quality after selecting chunking strategies and should preserve appropriate metadata and access-control information for each indexed item.

Question 273

A knowledge source changes every day. Which architecture is generally better for keeping responses current?

1. Rely only on model training
2. Use runtime retrieval from the updated source
3. Put all updates into a static prompt
4. Disable the knowledge source

Correct Answer: 2

Explanation

Runtime retrieval allows an agent to obtain current information from a knowledge source when responding to a request. This is generally more suitable for frequently changing information than relying solely on information embedded in model training or a static system prompt. The source should have an appropriate update and indexing process so that recent changes become available to retrieval. Access controls, freshness metadata, and evaluation are also important. This architecture can help keep responses aligned with current approved business information.

Question 274

Which metric is particularly useful for measuring whether retrieved content supports an AI-generated answer?

1. Groundedness
2. CPU temperature
3. Number of user accounts
4. Screen resolution

Correct Answer: 1

Explanation

Groundedness measures whether an AI-generated response is supported by the information supplied from trusted sources or retrieved context. A response can be fluent and relevant while still making unsupported claims, so groundedness is an important evaluation dimension for RAG systems. Teams can create representative evaluation datasets and assess whether responses are supported by retrieved evidence. Groundedness should be considered alongside other metrics such as relevance, correctness, safety, latency, and user satisfaction when evaluating an agent.

Question 275

What is the main purpose of an approval workflow for high-impact agent actions?

1. Allow unrestricted automation
2. Add human authorization before sensitive actions
3. Increase model temperature
4. Remove audit records

Correct Answer: 2

Explanation

Approval workflows provide an additional authorization step before an agent performs sensitive or high-impact actions. For example, an agent may prepare a financial transaction or account change but require an authorized employee to approve it before execution. This approach allows organizations to automate preparation and routine processing while maintaining human oversight where risk is higher. Approval requirements should be clearly defined and supported by technical controls. Audit records should capture important decisions and actions to support accountability and investigation.

Question 276

Which practice supports responsible AI by making important system decisions traceable?

1. Audit logging
2. Removing telemetry
3. Sharing administrator credentials
4. Disabling monitoring

Correct Answer: 1

Explanation

Audit logging creates records of important system events, decisions, actions, and access activities. For AI agents, useful logs may include tool calls, approvals, errors, configuration changes, and other events relevant to security or business operations. Logging supports troubleshooting, compliance, incident investigation, and accountability. Logs should be designed carefully to avoid unnecessarily storing sensitive information. Access to logs should also be controlled. Effective auditing provides visibility into what the system did and helps teams investigate unexpected behavior.

Question 277

An AI application handles personally identifiable information. Which design principle should be emphasized?

1. Data minimization and appropriate protection
2. Unlimited retention
3. Public data sharing
4. Unrestricted logging

Correct Answer: 1

Explanation

Personally identifiable information should be handled according to applicable privacy and security requirements. Data minimization reduces unnecessary exposure by ensuring that the application processes only the information required for its purpose. Appropriate access controls, encryption, retention policies, secure logging, and data-handling procedures provide additional safeguards. AI prompts and logs should not automatically contain complete customer records when only a small portion is required. Privacy should be considered throughout the solution lifecycle, including collection, processing, storage, monitoring, and deletion.

Question 278

What is the purpose of a disaster recovery strategy for an AI solution?

1. Restore critical capabilities after a major failure
2. Increase prompt length
3. Replace all evaluation datasets
4. Disable backup systems

Correct Answer: 1

Explanation

A disaster recovery strategy defines how critical AI capabilities can be restored after significant failures such as service outages, infrastructure problems, or data loss. Depending on the architecture, recovery planning may include backups, redundant services, alternative deployment locations, model availability plans, configuration recovery, and tested restoration procedures. Recovery objectives should align with business requirements. An AI solution should not rely solely on the availability of one component if downtime would significantly affect critical operations. Recovery procedures should also be tested periodically.

Question 279

Which factor should be considered when selecting an AI model for a production workload?

1. Only the model’s name
2. Accuracy, latency, cost, and business requirements
3. The color of the application interface
4. The number of employees in the office

Correct Answer: 2

Explanation

Model selection should consider multiple factors rather than relying on the model name or a single performance measurement. Important considerations can include task accuracy, latency, context capacity, supported capabilities, operating cost, availability, security, compliance, and business requirements. Organizations should evaluate candidate models using representative workloads and measurable acceptance criteria. A highly capable model may not be appropriate for every task if a smaller model provides sufficient quality at lower cost or latency. Model selection is therefore an architectural tradeoff.

Question 280

Why should AI agents be tested with adversarial inputs?

1. To increase database storage
2. To identify security, safety, and robustness weaknesses
3. To eliminate all monitoring
4. To make every response longer

Correct Answer: 2

Explanation

Adversarial testing deliberately uses challenging or malicious inputs to identify weaknesses in an AI system. Examples include prompt injection attempts, attempts to bypass restrictions, unexpected tool inputs, sensitive-data requests, and misleading instructions. Testing can reveal whether the agent follows security boundaries and handles unsafe or ambiguous situations appropriately. Findings should be addressed through improved prompts, authorization, validation, filtering, monitoring, and other controls. Adversarial testing should be repeated as models, tools, prompts, and integrations change.