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

 

Question 121

What is the main purpose of defining an agent’s goal?

1. To specify the outcome the agent is expected to achieve
2. To increase storage capacity
3. To remove security controls
4. To disable tool usage

Correct Answer: 1

Explanation

An agent goal defines the outcome that the AI solution is intended to achieve. A clear goal helps determine which tasks, tools, knowledge sources, and decision steps are necessary. It also provides a basis for testing whether the agent is performing as expected. For example, an agent goal might be to resolve routine customer requests or summarize approved business information. Well-defined goals help prevent unnecessary agent behavior and keep the solution aligned with its intended business purpose.

Question 122

Which approach helps an AI agent maintain information about the current task?

1. Increasing screen resolution
2. Managing task state
3. Removing context
4. Disabling conversation history

Correct Answer: 2

Explanation

Task state allows an agent to maintain information about what is currently happening during a workflow. This can include completed steps, pending actions, selected records, or information needed for the next operation. State management is particularly useful for multi-step processes where the agent must continue from an earlier point. Proper state handling also helps with recovery when an operation fails. The design should determine what information needs to persist and how long it should be retained.

Question 123

A customer-service agent needs to transfer complex billing questions to a specialized billing agent. Which architecture supports this requirement?

1. Data deletion
2. Static prompting
3. Agent handoff
4. Model compression

Correct Answer: 3

Explanation

Agent handoff allows one agent to transfer responsibility for a task to another specialized agent. In this scenario, the customer-service agent can identify that the request requires billing expertise and provide the relevant context to the billing agent. The receiving agent can then continue the interaction without forcing the customer to restart the process. Handoffs should define when they occur, what information is transferred, and which permissions the receiving agent has. This helps maintain continuity and clear responsibility across specialized agents.

Question 124

Which technique is useful when an agent needs to search documents by meaning rather than exact keywords?

1. File compression
2. Semantic search
3. Data deletion
4. Password rotation

Correct Answer: 2

Explanation

Semantic search retrieves information based on the meaning and intent of a query rather than requiring an exact keyword match. It commonly uses embeddings to represent queries and documents in a way that allows semantically related content to be identified. This can improve retrieval when users phrase questions differently from the wording used in source documents. Semantic search is often used as part of RAG architectures. Retrieval quality should still be evaluated to ensure that the returned information is relevant and sufficiently complete.

Question 125

What is a key purpose of prompt versioning?

1. To track changes to prompts over time
2. To increase API latency
3. To eliminate model evaluation
4. To disable rollback

Correct Answer: 1

Explanation

Prompt versioning allows teams to track changes made to system instructions, templates, or other prompts used by an AI application. This makes it easier to determine which prompt version produced a particular result and compare performance between revisions. Versioning also supports controlled testing and rollback when a change causes unexpected behavior. Prompts can significantly affect model outputs, so treating them as managed application assets helps improve consistency, troubleshooting, and lifecycle management across development, testing, and production environments.

Question 126

An organization wants to prevent an AI agent from accessing confidential files that are unrelated to its task. What should be applied?

1. Larger context windows
2. Broad administrator permissions
3. Access controls and data boundaries
4. More frequent prompting

Correct Answer: 3

Explanation

Access controls and data boundaries help ensure that an AI agent can access only the information required for its assigned tasks. Confidential files unrelated to the agent’s purpose should not automatically become available simply because the agent can technically connect to a storage system. Permissions should follow least-privilege principles and respect existing organizational access policies. Proper identity, authorization, data classification, and auditing should be combined to reduce the risk of unauthorized data exposure through agent interactions.

Question 127

What does an AI model’s context window determine?

1. The number of users in a database
2. The amount of information the model can process within a request
3. The number of network cables required
4. The physical size of the server

Correct Answer: 2

Explanation

A model’s context window represents the amount of input and contextual information that can be considered within a request, subject to the model’s supported limits. This can include user messages, system instructions, retrieved documents, conversation history, and other information supplied to the model. Context limits are important when designing RAG and conversational applications because excessive information can increase cost or exceed model constraints. Architecture teams should therefore manage context carefully and prioritize relevant information.

Question 128

Which approach can improve reliability when an AI service becomes temporarily unavailable?

1. Removing all monitoring
2. Using a fallback service or workflow
3. Increasing prompt complexity
4. Disabling error handling

Correct Answer: 2

Explanation

A fallback service or workflow provides an alternative when the primary AI service is unavailable or unable to complete a request. Depending on the business scenario, the fallback might use another model, a simpler process, cached information, or human escalation. The fallback should have clearly defined conditions and should not bypass required security controls. Testing should verify that users receive an appropriate response during failures. This approach can improve resilience and reduce disruption when individual AI components experience temporary availability problems.

Question 129

Which practice helps reduce hallucinations when an agent answers questions about company policies?

1. Increasing unrelated context
2. Removing knowledge sources
3. Grounding responses in approved policy information
4. Allowing unrestricted tool access

Correct Answer: 3

Explanation

Grounding an agent with approved policy information gives the model relevant evidence to use when generating responses. Instead of relying solely on general model knowledge, the agent can retrieve current organizational policies and use them as context. This can reduce unsupported or outdated responses, particularly for information that changes over time. Grounding should use authoritative sources and appropriate access controls. Retrieval quality, source freshness, and response accuracy should also be evaluated regularly to maintain dependable policy-related answers.

Question 130

Why might an organization use a small language model for a simple classification task?

1. It may provide lower cost and latency
2. It always produces better reasoning
3. It removes the need for evaluation
4. It guarantees perfect accuracy

Correct Answer: 1

Explanation

A smaller language model may be appropriate for simple tasks such as classification, routing, or straightforward text processing. Smaller models can sometimes provide lower inference costs and faster responses than larger models. However, the choice should be based on measured business requirements rather than model size alone. The organization should evaluate accuracy, latency, cost, reliability, and other relevant factors. If the smaller model cannot meet the required quality level, a more capable model may be necessary for that workload.

Question 131

What is the purpose of defining an agent’s scope?

1. To specify what the agent should and should not handle
2. To increase database size
3. To remove human oversight
4. To allow unrestricted actions

Correct Answer: 1

Explanation

An agent’s scope defines the tasks, users, data, and actions that fall within its intended responsibilities. A clearly defined scope helps prevent an agent from attempting tasks outside its purpose or accessing unnecessary resources. For example, a customer-support agent may be allowed to answer product questions but not approve refunds above a specified threshold. Scope should be reflected in instructions, permissions, tools, and governance controls. Clearly defined boundaries make agent behavior easier to test, monitor, and manage.

Question 132

An agent must process information from a database and then generate a customer response. Which sequence is appropriate?

1. Generate the answer before retrieving data
2. Retrieve the required data before generating the grounded response
3. Delete the database record first
4. Disable authorization before retrieval

Correct Answer: 2

Explanation

When an agent needs current structured information, it should generally retrieve the required data before generating the final response. The retrieved information can then provide factual context for the model. For example, an order-support agent could retrieve the customer’s current order status and then generate a response based on that result. The database operation must still use proper authentication and authorization. This approach helps reduce responses based on outdated assumptions and keeps the answer connected to current business information.

Question 133

Which feature is important when an AI application handles personally identifiable information?

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

Correct Answer: 1

Explanation

Personally identifiable information requires appropriate protection throughout the AI application’s lifecycle. Access should be limited to authorized users and services, and data should be handled according to applicable privacy and organizational requirements. Logging should also avoid unnecessarily exposing sensitive information. Other controls may include encryption, data classification, retention policies, and appropriate processing boundaries. AI architecture should identify where personal information enters, is stored, retrieved, processed, and returned so that privacy risks can be addressed at each stage.

Question 134

What can hybrid search combine in a retrieval system?

1. Authentication and encryption
2. Keyword and semantic search
3. Storage and networking
4. Logging and billing

Correct Answer: 2

Explanation

Hybrid search can combine traditional keyword-based retrieval with semantic search techniques. Keyword search is useful when exact terms, names, identifiers, or phrases matter, while semantic search can identify content based on conceptual similarity. Combining both approaches can improve retrieval across different types of business information. The exact implementation depends on the search platform and application requirements. Retrieval results should still be evaluated using representative queries to confirm that the system consistently returns relevant information for the intended use cases.

Question 135

Why should AI agents have clearly defined tool permissions?

1. To ensure tools can be used only for authorized purposes
2. To make every tool publicly accessible
3. To remove auditing requirements
4. To increase unnecessary actions

Correct Answer: 1

Explanation

Tool permissions determine which operations an agent is allowed to perform. Clearly defined permissions help ensure that an agent can use a tool only for approved purposes and within its intended scope. For example, an agent may be permitted to read customer order information but not delete orders. Restricting tool permissions reduces the potential impact of incorrect instructions, compromised inputs, or unexpected model behavior. Permissions should follow least privilege and should be reviewed whenever the agent’s responsibilities or integrations change.

Question 136

What is a major benefit of using automated evaluation in an AI lifecycle?

1. It removes the need for test data
2. It allows consistent measurement after changes
3. It guarantees that models never fail
4. It prevents all production monitoring

Correct Answer: 2

Explanation

Automated evaluation allows an AI solution to be tested consistently against defined datasets and metrics. When prompts, models, retrieval configurations, or tools change, automated tests can help determine whether performance improved, remained stable, or declined. This supports repeatable quality checks throughout the AI lifecycle. Automated evaluation does not guarantee perfect behavior and should be combined with other testing methods, including security, safety, and real-world monitoring. It is particularly useful as part of controlled deployment and release processes.

Question 137

An AI agent needs to perform a high-impact business action. Which design provides additional control?

1. Human approval before execution
2. Unlimited autonomous access
3. Removal of audit logs
4. Automatic execution without validation

Correct Answer: 1

Explanation

Human approval can provide an additional control for high-impact or sensitive actions. Instead of allowing an agent to execute the operation automatically, the system can present the proposed action to an authorized person for review. This approach is useful when an incorrect action could have significant financial, legal, operational, or customer consequences. The approval process should clearly identify what the agent intends to do and provide enough information for an authorized reviewer to make an informed decision before execution.

Question 138

What should an AI architecture consider when many users may access an agent simultaneously?

1. Prompt wording only
2. Scalability and capacity
3. Screen brightness
4. Document formatting

Correct Answer: 2

Explanation

Scalability and capacity planning are important when an AI application may receive many simultaneous requests. The architecture should consider model capacity, API quotas, concurrency, latency, storage, retrieval services, and other dependencies. Load testing can help identify bottlenecks before production deployment. The solution may also require throttling, queuing, caching, or workload distribution. Scalability planning should balance performance with cost and reliability so that increased demand does not cause unacceptable delays or service failures.

Question 139

Which approach helps protect an agent from prompt injection?

1. Trusting every retrieved instruction
2. Treating external content as automatically authorized commands
3. Separating trusted instructions from untrusted content
4. Giving retrieved documents administrator permissions

Correct Answer: 3

Explanation

Prompt injection can occur when untrusted content attempts to influence an AI system’s instructions or behavior. A defensive architecture should clearly separate trusted system instructions from external or user-provided content and avoid treating retrieved text as automatically authorized commands. Additional controls can include input validation, tool authorization, prompt-shielding mechanisms, output checks, and limiting high-impact actions. Security testing should include adversarial inputs to identify weaknesses. The goal is to ensure that external content cannot freely override the agent’s intended instructions or security boundaries.

Question 140

Which activity is part of AI production readiness?

1. Removing monitoring
2. Skipping security review
3. Deploying without testing
4. Validating reliability, security, and operational controls

Correct Answer: 4

Explanation

Production readiness requires more than confirming that an AI model can generate useful responses. Teams should validate reliability, security, access controls, monitoring, error handling, scalability, data protection, and operational procedures before production deployment. Testing should cover expected and unexpected inputs as well as important failure scenarios. Appropriate ownership and escalation processes should also be established. A production-ready AI solution should have controls that support safe operation, ongoing monitoring, and controlled changes after deployment.