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

 

Question 101

What is the primary purpose of agent memory in an AI solution?

1. To increase network bandwidth
2. To replace authentication
3. To remove all prompts
4. To retain relevant information across interactions

Correct Answer: 4

Explanation

Agent memory allows an AI agent to retain relevant information that can improve future interactions. Depending on the architecture, memory may store conversation context, user preferences, task state, or other approved information. This can make the agent more consistent and useful across multiple interactions. Memory should still be designed with privacy, security, retention, and access requirements in mind. It should not automatically retain sensitive information simply because it is available during an interaction.

Question 102

Which design is most appropriate when two independent AI tasks can run at the same time?

1. Parallel execution
2. Sequential execution
3. Manual execution
4. Single-step execution

Correct Answer: 1

Explanation

Parallel execution allows independent tasks to run simultaneously rather than waiting for one task to finish before starting another. This can reduce overall processing time and improve responsiveness when the tasks do not depend on each other’s results. For example, an agent could retrieve customer information and product information concurrently. The architecture must still account for resource limits, failures, and synchronization when combining results. Parallel processing is most useful when tasks have minimal dependencies between them.

Question 103

An agent must call an external API that limits requests per minute. What should the architecture consider?

1. Increasing prompt length
2. Removing authentication
3. Rate limiting and retry handling
4. Disabling monitoring

Correct Answer: 3

Explanation

External APIs commonly enforce rate limits to control traffic and protect service availability. An AI agent integrating with such an API should understand the allowed request rate and implement appropriate controls. Retry logic should also use suitable delays, such as exponential backoff, rather than repeatedly sending requests immediately. This helps prevent additional failures caused by excessive traffic. Monitoring API responses and request volumes can further help identify throttling conditions and maintain reliable agent behavior.

Question 104

Why is idempotency important when an AI agent performs business actions?

1. It increases model creativity
2. It helps prevent duplicate effects from repeated requests
3. It removes authorization requirements
4. It guarantees perfect model output

Correct Answer: 2

Explanation

Idempotency helps ensure that repeating the same operation does not unintentionally create multiple business effects. This is especially important when an agent performs actions such as submitting orders, creating records, or initiating transactions. Network failures or retries can cause an action to be sent more than once. An idempotent design can recognize repeated requests and avoid duplicate processing where appropriate. This improves reliability and reduces the risk of unintended business changes caused by agent retries or communication failures.

Question 105

Which approach is most appropriate for storing API credentials used by an AI application?

1. A secure secrets-management service
2. The system prompt
3. A public knowledge source
4. A user-visible response

Correct Answer: 1

Explanation

API credentials should be stored using secure secrets-management mechanisms rather than being embedded in prompts, source code, documentation, or user-visible content. A secrets-management service can provide controlled access, encryption, auditing, and credential rotation capabilities. Access should also follow least-privilege principles so an application receives only the permissions it requires. Keeping credentials outside prompts and application responses reduces the chance that sensitive authentication information will be accidentally exposed through logs, outputs, or other channels.

Question 106

What does an agent handoff typically involve?

1. Deleting the original conversation
2. Disabling the receiving agent
3. Transferring a task or context to another agent
4. Replacing all business data

Correct Answer: 3

Explanation

An agent handoff occurs when responsibility for a task moves from one agent or workflow component to another. The receiving agent may need relevant context, task state, user information, and instructions to continue effectively. For example, a general customer-service agent could hand a billing issue to a specialized billing agent. A well-designed handoff defines what information is transferred and what permissions the receiving agent has. This helps maintain continuity while keeping responsibilities and security boundaries clear.

Question 107

Which component helps an agent discover and use approved external capabilities through a standardized interface?

1. A spreadsheet
2. A screen resolution setting
3. A local printer
4. A tool or protocol interface

Correct Answer: 4

Explanation

Standardized tool or protocol interfaces can allow agents to discover and interact with approved capabilities without requiring every integration to use a completely different communication pattern. Protocol-based approaches can define how tools, resources, or context are exposed to an AI application. This can improve interoperability and simplify integration across different systems. Security controls remain important because discovering a tool does not automatically mean the agent should be authorized to invoke every operation that the tool provides.

Question 108

A company wants an agent to answer questions using frequently changing inventory information. What should the solution prioritize?

1. Current data retrieval
2. Static prompts only
3. Manual transcription
4. Larger fonts

Correct Answer: 1

Explanation

Frequently changing inventory information should generally be retrieved from an authoritative and current data source rather than being permanently embedded in a prompt. A retrieval or tool-based architecture can allow the agent to access current inventory information when responding to users. This reduces the risk of outdated answers and helps align responses with operational data. The solution should also enforce appropriate authorization and validate the retrieved information. Data refresh frequency and source reliability should be included in the architecture.

Question 109

What is a key benefit of using an evaluation dataset for an AI agent?

1. It replaces production monitoring
2. It provides repeatable test cases for measuring performance
3. It eliminates the need for security testing
4. It guarantees zero hallucinations

Correct Answer: 2

Explanation

An evaluation dataset provides representative inputs and expected outcomes that can be used to measure an AI solution consistently. Teams can use it before and after changes to compare accuracy, groundedness, relevance, safety, or other defined metrics. This makes evaluation more repeatable than relying only on occasional manual testing. An evaluation dataset does not guarantee perfect behavior and should not replace production monitoring or security testing. Its value comes from providing structured evidence about how the system performs against known scenarios.

Question 110

Which strategy can help reduce AI processing costs when a less capable model is sufficient?

1. Always selecting the largest model
2. Increasing context unnecessarily
3. Routing suitable requests to a smaller model
4. Repeating every request multiple times

Correct Answer: 3

Explanation

Model routing can reduce costs by sending suitable requests to smaller or less expensive models while reserving larger models for tasks that genuinely require additional capability. For example, a simple classification or short summarization task may not require the same model used for complex reasoning. Routing decisions should consider accuracy, latency, complexity, and business requirements rather than cost alone. The architecture can also monitor performance to confirm that the selected lower-cost model continues to meet the required quality standards.

Question 111

What is the main purpose of a fallback model or workflow?

1. To provide an alternative when the primary path fails
2. To remove all monitoring
3. To prevent users from receiving responses
4. To permanently disable the primary model

Correct Answer: 1

Explanation

A fallback model or workflow provides an alternative path when the primary AI service cannot complete a request successfully. Failure may result from service availability issues, capacity limitations, timeouts, unsupported requests, or other operational conditions. A fallback could use another model, a simpler workflow, or escalation to a human. The fallback should be tested and monitored like the primary path. Good fallback design improves resilience while ensuring that users receive an appropriate response instead of an unexplained system failure.

Question 112

Which approach is most suitable for protecting an agent from excessive permissions?

1. Give the agent administrator access
2. Apply least-privilege permissions
3. Share one account across all agents
4. Remove authentication

Correct Answer: 2

Explanation

Least privilege means giving an agent only the permissions required to perform its defined tasks. If an agent only needs to read customer orders, it should not automatically receive permission to delete records or modify unrelated systems. Limiting permissions reduces the potential impact of accidental or unauthorized actions. Permissions should be reviewed as the agent evolves and should align with identity, authentication, authorization, and governance requirements. This principle is particularly important for autonomous agents that can perform actions without constant user intervention.

Question 113

An AI agent must update a customer record after receiving a request. What should happen before the update?

1. The model should invent the record ID
2. The agent should skip authorization
3. The request should be validated and authorized
4. The agent should publish the record publicly

Correct Answer: 3

Explanation

Before changing a customer record, the system should validate the request and confirm that the requesting user or agent has permission to perform the specific operation. The solution should also verify important input details, such as the correct customer record and requested changes. Authorization prevents an AI agent from becoming an uncontrolled pathway into business systems. Combining authentication, authorization, validation, and auditing provides stronger protection for sensitive operations than relying on the language model alone.

Question 114

What is one reason to separate development, testing, and production environments for AI solutions?

1. To increase prompt length
2. To avoid all evaluation
3. To prevent versioning
4. To reduce the risk of untested changes affecting users

Correct Answer: 4

Explanation

Separate development, testing, and production environments allow teams to develop and validate AI changes before exposing them to real users and business processes. Prompts, models, tools, integrations, and agent instructions can be tested in controlled environments. This reduces the risk that an untested configuration or workflow will cause production problems. Environment separation also supports change management, version control, approval processes, and deployment governance. Production access should be restricted and changes should follow an appropriate lifecycle.

Question 115

Which technique can improve retrieval results for a RAG solution?

1. Using relevant document chunks
2. Removing all source metadata
3. Ignoring query meaning
4. Indexing unrelated information

Correct Answer: 1

Explanation

Effective retrieval depends on providing the generation model with relevant information. Proper document chunking can divide large sources into meaningful sections that are easier to retrieve. Other factors, such as embeddings, metadata, semantic search, hybrid search, and indexing quality, can also influence retrieval performance. Poorly chosen chunks may contain too little context or too much unrelated information. Retrieval should therefore be tested against representative questions to determine whether the returned content is relevant, complete, and suitable for generating grounded responses.

Question 116

A tool call fails because an external service temporarily returns an error. What should the agent architecture support?

1. Immediate permanent shutdown
2. Unlimited repeated requests
3. Appropriate retry and error handling
4. Ignoring the error completely

Correct Answer: 3

Explanation

Temporary external-service failures should be handled through controlled error-handling mechanisms. Depending on the type of error, the agent may retry the operation using an appropriate backoff strategy, use a fallback service, or notify the user that the action could not be completed. Unlimited retries can increase load and worsen an outage, while ignoring errors can produce misleading results. The architecture should distinguish temporary failures from permanent errors and record relevant information through monitoring and logging.

Question 117

What is a primary purpose of content safety controls in generative AI applications?

1. To increase database storage
2. To reduce harmful or inappropriate outputs
3. To eliminate authentication
4. To increase API request rates

Correct Answer: 2

Explanation

Content safety controls help reduce the likelihood that an AI application will generate or process harmful, unsafe, or inappropriate content. Depending on the solution, controls may include input filtering, output filtering, safety classifiers, policy enforcement, and human review for higher-risk scenarios. These controls should be evaluated against realistic use cases and potential misuse. Content safety is one part of a broader responsible AI architecture and should work alongside security, privacy, monitoring, access control, and governance measures.

Question 118

Which factor is especially important when choosing between fine-tuning and RAG?

1. Whether the application needs access to changing external knowledge
2. Whether the monitor has a high resolution
3. Whether users prefer dark mode
4. Whether the application has a large screen

Correct Answer: 1

Explanation

The choice between fine-tuning and retrieval-augmented generation depends partly on the type of knowledge the application needs. If information changes frequently and must come from current business sources, RAG can provide a way to retrieve updated information at request time. Fine-tuning is generally used for adapting model behavior or patterns rather than serving as a continuously updated knowledge repository. Architecture decisions should consider data freshness, cost, maintenance, security, model behavior, and the specific business requirement.

Question 119

Why should AI agent actions be logged?

1. To make prompts longer
2. To remove the need for testing
3. To support auditing and troubleshooting
4. To prevent every tool call

Correct Answer: 3

Explanation

Logging agent actions provides useful information for auditing, troubleshooting, security investigations, and operational monitoring. Logs can help teams understand which tools were invoked, when actions occurred, whether requests failed, and what workflow path was followed. Logging should be designed carefully so that sensitive information is not unnecessarily captured. Appropriate retention and access controls are also important. Well-designed audit information can help organizations investigate unexpected behavior and verify whether agents are operating according to approved business processes.

Question 120

What should be defined before measuring whether an AI pilot is successful?

1. Success criteria and measurable KPIs
2. The largest available model
3. Unlimited user permissions
4. A permanent production deployment

Correct Answer: 1

Explanation

An AI pilot should have clearly defined success criteria before performance is measured. Relevant KPIs might include task completion rate, response quality, processing time, user satisfaction, cost per interaction, or reduction in manual work. The selected measures should directly relate to the business objective being tested. Clear criteria make it easier to determine whether the pilot achieved its intended purpose and whether further investment is justified. They also provide a consistent basis for comparing results during later testing and production deployment.