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

 

Question 301

Which factor should be defined first when designing an AI agent?

1. Model temperature
2. Number of embeddings
3. Database size
4. Business objective

Correct Answer: 4

Explanation

A clear business objective should guide the design of an AI agent. The objective defines what problem the agent must solve, which users it serves, what actions it should perform, and how success will be measured. Once the objective is understood, architects can select suitable models, tools, data sources, and workflows. Starting with technology alone can result in unnecessary complexity or an unsuitable solution. Business objectives should therefore remain connected to technical requirements throughout development, testing, deployment, and ongoing evaluation.

Question 302

What is a key advantage of using structured output from an AI model?

1. It makes responses easier for applications to process consistently
2. It eliminates model errors
3. It removes authentication requirements
4. It guarantees factual accuracy

Correct Answer: 1

Explanation

Structured output requires an AI model to return information in a defined format, such as JSON with specific fields. This makes responses easier for applications and workflows to parse and process consistently. Validation can also verify that required fields, data types, and allowed values are present before the output is used. Structured output does not guarantee that the model’s content is correct. It should therefore be combined with validation, business rules, authorization, and evaluation when the output will drive automated actions.

Question 303

An agent receives customer text and must extract an order number, date, and product name. Which approach is suitable?

1. Disaster recovery
2. Data extraction
3. Rate limiting
4. Model routing

Correct Answer: 2

Explanation

Data extraction is designed to identify and return specific information from unstructured or semi-structured input. In this scenario, the agent needs to identify defined fields such as an order number, date, and product name. A structured output schema can make the extracted information easier for downstream applications to process. The extracted values should still be validated before they are used in business operations. For important workflows, evaluation should measure whether the system consistently extracts the required fields accurately.

Question 304

Which control can prevent users from sending sensitive information to an AI application when such data is not permitted?

1. Larger context windows
2. Data loss prevention controls
3. Higher model temperature
4. More embeddings

Correct Answer: 2

Explanation

Data loss prevention controls can help identify and restrict the movement or use of sensitive information according to organizational policies. In an AI application, these controls may help prevent prohibited data from being submitted to models, connectors, or external services. DLP should be combined with access controls, data classification, monitoring, and appropriate application design. The exact implementation depends on the platform and organizational requirements. DLP is particularly important when AI solutions interact with multiple business systems or external services.

Question 305

What is the purpose of data classification in an AI solution?

1. Identify data according to sensitivity and handling requirements
2. Increase model context size
3. Replace authentication
4. Generate embeddings

Correct Answer: 1

Explanation

Data classification categorizes information according to characteristics such as sensitivity, confidentiality, or regulatory requirements. Classification can help determine which data an AI system may process, where it can be stored, which users can access it, and whether additional protections are necessary. For example, confidential customer information may require stricter controls than publicly available content. Classification should be considered before connecting enterprise data sources to agents. It provides an important foundation for privacy, security, access control, and data-governance decisions.

Question 306

A model produces acceptable answers but takes too long to respond. Which factor should the team investigate?

1. User password length
2. Application logo
3. Model and workflow latency
4. Document naming conventions

Correct Answer: 3

Explanation

When response quality is acceptable but response time is too high, the team should investigate latency across the entire AI workflow. Possible contributors include model processing time, retrieval, tool calls, network communication, orchestration, and sequential operations. A smaller model, caching, parallel execution, optimized retrieval, or fewer unnecessary tool calls may improve performance. Teams should measure individual components rather than assuming the model is the only source of delay. Performance changes should still be validated to ensure they do not reduce required quality.

Question 307

Which practice helps an organization maintain consistent AI development standards across teams?

1. Shared architecture and governance standards
2. Independent undocumented configurations
3. Unlimited administrator permissions
4. Different security policies for every agent

Correct Answer: 1

Explanation

Shared architecture and governance standards provide consistent expectations for how AI solutions are designed, secured, tested, deployed, and monitored. Standards can cover approved models, data handling, evaluation methods, identity controls, logging, prompt management, and deployment processes. Consistency reduces duplicated effort and makes solutions easier to review and maintain. Teams can still adapt implementations to specific business needs while following common organizational requirements. Documented standards also make it easier to scale successful AI practices across multiple projects.

Question 308

An agent must retrieve information from a CRM system. What should be configured before allowing access?

1. The largest available model
2. Appropriate authentication and authorization
3. Unlimited tool permissions
4. A longer system prompt

Correct Answer: 2

Explanation

Before an agent accesses a CRM system, the organization should establish secure authentication and appropriate authorization. Authentication identifies the identity making the request, while authorization determines which resources and operations that identity can access. Permissions should follow least privilege and should reflect the agent’s defined responsibilities. Input validation, logging, and monitoring should also be considered. Simply connecting an agent to a business system without enforcing access boundaries can expose sensitive customer information or allow unauthorized changes.

Question 309

Which technique can help improve retrieval when documents contain many similar terms?

1. Query rewriting
2. Removing metadata
3. Disabling indexing
4. Increasing user permissions

Correct Answer: 1

Explanation

Query rewriting can make a user’s request more precise before it is sent to a retrieval system. When documents contain similar terminology, a clearer query can help identify the intended subject and improve the relevance of retrieved results. Query rewriting can be combined with metadata filters, semantic search, keyword search, and reranking. The rewritten query should not introduce unsupported assumptions or bypass access controls. Retrieval improvements should be validated using representative queries and measurable evaluation criteria before being adopted in production.

Question 310

Why should an AI solution maintain a baseline evaluation?

1. To compare future changes against a known level of performance
2. To prevent all model updates
3. To eliminate monitoring
4. To increase token consumption

Correct Answer: 1

Explanation

A baseline evaluation establishes a reference point for the system’s performance before significant changes are introduced. Teams can compare future model, prompt, retrieval, or configuration changes against this baseline to identify improvements or regressions. Metrics may include accuracy, relevance, groundedness, safety, latency, and cost. A baseline is especially useful when multiple components can influence output quality. Maintaining consistent evaluation datasets and criteria makes comparisons more meaningful and supports controlled decisions during the AI solution lifecycle.

Question 311

What is the main purpose of prompt versioning?

1. Track changes to prompts and support controlled testing
2. Store customer passwords
3. Replace model monitoring
4. Increase API capacity

Correct Answer: 1

Explanation

Prompt versioning records different versions of prompts or agent instructions so teams can identify what changed and evaluate its effect. This is important because even a small instruction change can alter an agent’s behavior. Versioning allows teams to reproduce earlier configurations, compare results, and roll back problematic changes when necessary. Prompts should be treated as managed application assets rather than informal text. Combining versioning with evaluation and deployment controls helps reduce unexpected production regressions.

Question 312

A company wants to process independent documents simultaneously. Which approach can improve throughput?

1. Sequential processing only
2. Parallel processing
3. Manual processing
4. Disabling queues

Correct Answer: 2

Explanation

Parallel processing allows independent tasks to execute concurrently, which can improve throughput and reduce total processing time. For example, if several documents can be summarized independently, the system does not necessarily need to wait for one document to finish before starting the next. However, parallel execution should respect service limits, concurrency restrictions, resource capacity, and failure-handling requirements. Tasks that depend on previous results should remain sequential. Proper orchestration identifies which operations can safely run concurrently without creating data or workflow conflicts.

Question 313

What should an agent do when it lacks enough information to complete a request safely?

1. Invent missing information
2. Execute the most powerful available action
3. Ask for clarification or escalate
4. Disable authorization

Correct Answer: 3

Explanation

When an agent lacks sufficient information, it should avoid inventing details or making unsupported assumptions. Depending on the situation, the appropriate response may be to ask the user for clarification, retrieve additional approved information, or escalate the request to a human or specialized workflow. This behavior improves reliability and reduces the risk of incorrect business actions. Clear agent instructions should define how uncertainty is handled. Evaluation should also include ambiguous and incomplete requests to verify that the agent responds appropriately.

Question 314

Which factor is important when determining whether an AI agent should operate autonomously?

1. The color of the interface
2. The risk and impact of its actions
3. The number of available fonts
4. The length of the company name

Correct Answer: 2

Explanation

The level of agent autonomy should reflect the risk and potential impact of the actions it can perform. Low-risk tasks may be suitable for greater automation, while financial, legal, security, or other high-impact operations may require human approval. Architects should consider authorization, reversibility, business consequences, and error tolerance when determining autonomy. The goal is not simply to maximize automation but to establish appropriate controls around consequential actions. Human oversight and approval workflows can provide additional safeguards where required.

Question 315

What is a benefit of using reusable AI components across multiple solutions?

1. Greater consistency and reduced development effort
2. Automatic elimination of security testing
3. Unlimited permissions
4. Guaranteed model accuracy

Correct Answer: 1

Explanation

Reusable components can reduce development effort and improve consistency across AI solutions. Examples include approved prompt templates, evaluation frameworks, connectors, security patterns, tool interfaces, and shared governance components. Reuse can also make maintenance easier because improvements to a standardized component can potentially benefit multiple solutions. However, reusable components still need to be evaluated for the specific context in which they are deployed. Reuse should not mean automatically sharing excessive permissions, data access, or configuration across unrelated business applications.

Question 316

Which feature helps ensure an agent only accesses approved knowledge sources?

1. Controlled knowledge-source configuration
2. Unlimited web access
3. Random retrieval
4. Administrator credentials

Correct Answer: 1

Explanation

Controlled knowledge-source configuration defines which repositories an agent is allowed to use for grounding and retrieval. This helps organizations restrict responses to approved information rather than allowing uncontrolled sources that may be inaccurate, outdated, or unauthorized. Knowledge sources should also have appropriate ownership, permissions, freshness processes, and monitoring. Restricting sources does not eliminate the need to evaluate generated answers. The agent should still be tested for groundedness, relevance, and appropriate handling of questions that cannot be answered from approved sources.

Question 317

A business process requires a manager to approve an AI-generated recommendation before execution. What architecture pattern is this?

1. Fully autonomous processing
2. Human-in-the-loop
3. Unrestricted automation
4. Model pretraining

Correct Answer: 2

Explanation

A human-in-the-loop architecture includes human review or approval at a defined stage of an automated workflow. In this scenario, the AI can analyze information and generate a recommendation, while the manager remains responsible for approving the consequential action. This pattern is useful when decisions have meaningful financial, operational, compliance, or customer impact. The approval step should be enforced technically rather than relying only on the agent’s instructions. The workflow should also maintain appropriate audit records of recommendations and approvals.

Question 318

Which approach can help protect an agent from malicious instructions contained in retrieved documents?

1. Treat all retrieved text as trusted commands
2. Apply prompt-injection defenses and separate data from instructions
3. Give retrieved documents administrator access
4. Disable authentication

Correct Answer: 2

Explanation

Retrieved documents should generally be treated as untrusted content rather than automatically trusted instructions. Prompt-injection defenses can help separate system-level instructions from information retrieved for grounding. Additional protections include restricting tool permissions, validating actions, filtering suspicious content, and testing adversarial scenarios. Even trusted repositories can contain content that should not control an agent’s behavior. Defense should therefore be layered across prompts, retrieval, tool access, authorization, and monitoring rather than depending on one protection alone.

Question 319

Which metric can help determine whether an AI deployment provides measurable business value?

1. Business KPI
2. Prompt length
3. Number of embeddings alone
4. Model parameter count alone

Correct Answer: 1

Explanation

Business KPIs connect an AI deployment to measurable organizational outcomes. Depending on the use case, KPIs may include processing time, case-resolution rate, customer satisfaction, error reduction, employee productivity, conversion rate, or operating cost. Technical metrics such as latency and model accuracy remain important, but they do not by themselves demonstrate business value. Teams should establish relevant KPIs before or during a pilot and compare results against defined baselines. This helps determine whether the AI solution is achieving its intended business objectives.

Question 320

Before deploying an AI agent broadly, what should a pilot primarily validate?

1. Whether the solution meets defined technical and business requirements
2. Whether every employee receives administrator access
3. Whether monitoring can be disabled
4. Whether all available data can be connected

Correct Answer: 1

Explanation

A pilot should validate whether the AI solution works effectively for representative users and scenarios while meeting defined technical and business requirements. Teams can assess accuracy, groundedness, usability, latency, reliability, security, cost, and relevant business KPIs. Pilot testing can also uncover workflow issues that were not identified during development. The scope should remain controlled so that risks can be managed before broader deployment. Results from the pilot should inform improvements and production-readiness decisions rather than simply demonstrating that the system can generate responses.