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

 

Question 161

What is the purpose of agent orchestration?

1. To coordinate tasks, tools, and agents toward a goal
2. To increase database storage
3. To remove all user authentication
4. To disable monitoring

Correct Answer: 1

Explanation

Agent orchestration coordinates different tasks, agents, tools, and workflows so they can work together toward a defined objective. An orchestrator may determine which component should act, what information should be passed between steps, and when the process is complete. This is especially useful for complex business workflows involving multiple specialized capabilities. Effective orchestration should include clear boundaries, error handling, authorization, and monitoring. The design should avoid unnecessary complexity and use multiple agents only when specialization provides meaningful value.

Question 162

Which method is commonly used to represent text for semantic similarity searches?

1. Encryption
2. Embeddings
3. Compression
4. Hashing passwords

Correct Answer: 2

Explanation

Embeddings represent text as numerical vectors that capture aspects of its meaning. These vectors can be compared to identify content that is semantically related even when different words are used. Embeddings are commonly used in retrieval systems, including RAG architectures, where a user’s query can be matched with relevant document content. The quality of embeddings and the indexing strategy affect retrieval performance. Organizations should evaluate retrieval results using representative queries to ensure that semantically related information is actually useful for the application.

Question 163

An agent must process thousands of independent documents overnight. Which approach can improve throughput?

1. Processing every document manually
2. Running independent tasks in parallel where supported
3. Disabling monitoring
4. Using only sequential processing

Correct Answer: 2

Explanation

Independent document-processing tasks can often be executed in parallel to improve throughput. Instead of waiting for one document to finish before starting the next, the system can distribute independent workloads across available processing capacity. The architecture should still account for model quotas, API limits, concurrency, failures, and resource consumption. Monitoring can help determine whether parallel execution is producing the expected improvement. If downstream services have strict limits, controlled concurrency may be required to prevent excessive traffic or throttling.

Question 164

What is a key purpose of an AI governance framework?

1. To define controls, responsibilities, and policies for AI use
2. To eliminate all testing
3. To provide unlimited model access
4. To remove human accountability

Correct Answer: 1

Explanation

An AI governance framework establishes policies, responsibilities, controls, and processes for developing and operating AI solutions. Governance can address areas such as data protection, security, responsible AI, model evaluation, access management, monitoring, documentation, and change control. It also helps organizations define who owns specific risks and decisions. Governance should be appropriate to the organization’s use cases and risk level. Clear governance supports consistent management of AI systems throughout development, deployment, operation, and eventual retirement.

Question 165

Which situation is most appropriate for using a human escalation path?

1. A request requiring judgment beyond the agent’s authorized scope
2. A routine greeting
3. A simple approved lookup
4. A standard formatting task

Correct Answer: 1

Explanation

Human escalation is appropriate when an agent encounters a situation that exceeds its defined authority, capability, or confidence. Examples can include unusual customer disputes, high-impact decisions, conflicting information, or requests requiring specialized human judgment. The escalation path should provide the human reviewer with relevant context so the user does not need to repeat the entire interaction. Clear escalation criteria should be defined during solution design and tested regularly. This creates a controlled boundary between automated processing and human decision-making.

Question 166

Why can retrieval metadata be useful in a RAG solution?

1. It can help filter and identify relevant sources
2. It removes the need for indexing
3. It guarantees perfect responses
4. It disables access control

Correct Answer: 1

Explanation

Metadata can provide additional information about retrieved content, such as document type, department, date, author, or access classification. This information can help a retrieval system filter results and identify which sources are appropriate for a particular query. For example, a system could prioritize recent documents or restrict retrieval to a specific business department. Metadata should be accurate and maintained as source information changes. It can complement semantic and keyword search while also supporting governance and access-control requirements.

Question 167

A company wants to reduce unnecessary model calls for repeated requests. Which technique may help?

1. Caching appropriate results
2. Increasing prompt size
3. Disabling authentication
4. Removing monitoring

Correct Answer: 1

Explanation

Caching can reduce unnecessary model or retrieval calls when the same or sufficiently similar information can be reused safely. This may improve response time and reduce processing costs. Caching must be designed carefully for data that changes frequently or contains sensitive information. Cache expiration, invalidation, access controls, and data freshness should be considered. It is not appropriate to reuse a response indefinitely when the underlying information may have changed. The caching strategy should therefore reflect the specific business and data requirements.

Question 168

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

1. Required capability, cost, latency, and quality
2. Only the model’s logo
3. Only the number of available buttons
4. Only the application color scheme

Correct Answer: 1

Explanation

Model selection should be based on the requirements of the workload rather than on model size or popularity alone. Important considerations can include task capability, response quality, latency, cost, context requirements, throughput, availability, and safety characteristics. A model that performs well for one workload may not be appropriate for another. Teams should evaluate candidate models using representative data and measurable criteria. This allows the organization to balance technical requirements with business constraints before selecting a model for production use.

Question 169

What should an AI application do with conflicting information from two approved sources?

1. Automatically choose one without validation
2. Ignore the conflict
3. Apply defined source-priority or escalation rules
4. Invent a combined fact

Correct Answer: 3

Explanation

Conflicting information should be handled through predefined rules rather than allowing the model to arbitrarily choose a result. The architecture may assign priority to authoritative sources, compare document dates, retrieve additional evidence, or escalate the issue when the conflict cannot be resolved automatically. The chosen approach should reflect the business requirements and source ownership. Evaluation should include conflicting-data scenarios so teams can verify that the agent behaves predictably. Unsupported assumptions should be avoided when reliable evidence is unavailable.

Question 170

Why is data classification useful in an AI architecture?

1. It helps determine appropriate handling and protection requirements
2. It guarantees model accuracy
3. It removes authorization requirements
4. It increases token limits

Correct Answer: 1

Explanation

Data classification identifies the sensitivity and handling requirements of information used by an AI solution. Data may be categorized according to organizational policies, such as public, internal, confidential, or highly sensitive. Classification can influence access controls, storage, processing locations, logging, retention, and whether information is appropriate for a particular AI service. Understanding data sensitivity before integration helps prevent inappropriate exposure. Classification should be combined with authorization and governance controls throughout the data and AI lifecycle.

Question 171

What is one reason to use a pilot before a full AI deployment?

1. To validate the solution against real business requirements on a limited scale
2. To avoid defining requirements
3. To guarantee production success
4. To remove monitoring

Correct Answer: 1

Explanation

A pilot provides an opportunity to test an AI solution against real business requirements before expanding it broadly. A controlled pilot can reveal issues involving usability, accuracy, data quality, cost, integration, security, and operational performance. Teams can use predefined success criteria to determine whether the solution is meeting its objectives. A pilot does not guarantee that a full deployment will succeed, but it provides evidence that can inform further development and risk management. Lessons from the pilot can then guide production readiness activities.

Question 172

Which approach helps maintain consistency when multiple teams develop AI solutions?

1. Removing shared standards
2. Establishing reusable architecture and governance standards
3. Allowing every team to bypass security
4. Disabling documentation

Correct Answer: 2

Explanation

Reusable architecture and governance standards can help organizations maintain consistency across multiple AI solutions. Shared standards may cover identity, data access, prompt management, evaluation, monitoring, security, deployment, and responsible AI requirements. Teams can still adapt implementations to their specific business needs while following common controls. Standardization can reduce duplicated work and simplify support. Governance should remain flexible enough to account for differences in risk and functionality rather than forcing every AI workload into exactly the same implementation.

Question 173

An AI agent uses a third-party API to retrieve customer information. What should be verified before integration?

1. Authentication, authorization, data handling, and API limits
2. Only the API’s website design
3. Only the response font
4. Only the number of documentation pages

Correct Answer: 1

Explanation

Third-party API integration requires evaluation of authentication, authorization, data handling, rate limits, reliability, and contractual or organizational requirements. The team should determine what information is transmitted to the API and whether that processing is permitted. Credentials must be protected, and the agent should receive only the permissions necessary for its tasks. Error handling and monitoring should also be implemented. Reviewing these factors before integration reduces the risk of security problems, unexpected costs, service interruptions, or inappropriate data sharing.

Question 174

What is the purpose of a model evaluation baseline?

1. To provide a reference point for comparing future changes
2. To prevent all model updates
3. To eliminate test datasets
4. To increase model temperature

Correct Answer: 1

Explanation

An evaluation baseline establishes a reference level of performance against which future model, prompt, retrieval, or configuration changes can be compared. For example, a team can record accuracy, groundedness, latency, cost, and safety results for the current production configuration. When a new version is tested, its results can be compared with the baseline to identify improvements or regressions. Maintaining a baseline supports controlled change management and makes it easier to determine whether a proposed update meets established requirements.

Question 175

Which design can help an agent recover from a temporary tool failure?

1. Retry with controlled backoff or use a defined fallback
2. Repeat the request without limits
3. Ignore the failure
4. Remove all error handling

Correct Answer: 1

Explanation

Temporary tool failures can be handled through controlled retry logic or a predefined fallback path. A retry strategy should consider the type of error and use appropriate delays to avoid overwhelming the external service. If retries are unsuccessful, the agent can use an alternative tool, provide a suitable message, or escalate the task. Error handling should be observable through logs and telemetry. The system should avoid unlimited retries because repeated requests can increase service load and potentially create duplicate business actions.

Question 176

What does least privilege mean for an AI agent?

1. Giving the agent only the permissions required for its tasks
2. Giving the agent all administrator permissions
3. Sharing credentials across all systems
4. Removing authentication

Correct Answer: 1

Explanation

Least privilege means providing an AI agent with only the permissions necessary to perform its approved responsibilities. For example, an agent that needs to read order information should not automatically receive permission to delete customer records. Limiting permissions reduces the potential impact of incorrect outputs, compromised inputs, or unexpected behavior. Permissions should be tied to the agent’s identity and business role and reviewed periodically. Least privilege should apply to data access, tools, APIs, and other connected resources.

Question 177

Why should AI-generated outputs sometimes be validated before being sent to users?

1. To check for format, safety, accuracy, or policy requirements
2. To make responses longer
3. To disable all automation
4. To remove useful information

Correct Answer: 1

Explanation

Output validation can identify problems before AI-generated content reaches users or downstream systems. Depending on the application, validation may check required formats, prohibited content, business rules, sensitive information, or structured fields. For higher-risk workflows, validation may also include human review. Output validation does not guarantee that every response is correct, so it should complement grounding, evaluation, access controls, and monitoring. The level of validation should reflect the potential impact of incorrect or unsafe AI-generated content.

Question 178

Which capability can help identify whether an agent’s response is supported by retrieved information?

1. Groundedness evaluation
2. Screen sharing
3. Password expiration
4. File compression

Correct Answer: 1

Explanation

Groundedness evaluation examines whether an AI-generated response is supported by the information provided to the model, such as retrieved documents or other approved sources. This is useful for RAG applications where responses are expected to remain connected to business information. Teams can create evaluation datasets containing representative questions and source material and then assess whether generated answers are adequately supported. Groundedness should be considered alongside other measures such as relevance, correctness, safety, and user satisfaction.

Question 179

What should an organization do when an AI agent no longer meets its intended business requirements?

1. Review performance and modify, replace, or retire the solution as appropriate
2. Keep it unchanged indefinitely
3. Remove all monitoring
4. Expand its permissions

Correct Answer: 1

Explanation

AI solutions should be managed throughout their lifecycle rather than treated as permanent deployments. If an agent no longer meets business requirements, the organization should review performance, user feedback, costs, risks, and changing requirements. Possible actions include updating prompts, changing models, improving data sources, modifying workflows, replacing the solution, or retiring it. Lifecycle management should include appropriate testing and change control. Retiring an ineffective or outdated system can be part of responsible AI operations when continued use no longer provides sufficient value.

Question 180

Which practice supports responsible AI deployment?

1. Defining accountability, monitoring, and risk controls
2. Allowing unrestricted autonomous decisions
3. Ignoring user feedback
4. Removing documentation

Correct Answer: 1

Explanation

Responsible AI deployment requires clear accountability and controls that address the risks associated with the specific use case. These can include appropriate human oversight, monitoring, data protection, security, transparency, evaluation, and documented responsibilities. User feedback and operational evidence can help identify problems after deployment. Controls should be proportional to the potential impact of the AI system. Responsible AI is not a single configuration setting; it requires ongoing governance throughout design, deployment, monitoring, and lifecycle management.