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

 

Question 181

What is the main benefit of defining clear agent instructions?

1. To increase storage
2. To remove authorization
3. To guide the agent’s behavior and boundaries
4. To disable tool integration

Correct Answer: 3

Explanation

Clear agent instructions define how an agent should behave, what objectives it should pursue, and which boundaries it must follow. They can specify when the agent should use tools, how it should respond when information is unavailable, and when it should escalate a request. Well-defined instructions help reduce ambiguous behavior and make testing more consistent. Instructions should complement, rather than replace, technical security controls. Permissions, authentication, and authorization must still be enforced by the underlying systems and connected resources.

Question 182

Which approach can help manage long conversations without sending unnecessary history to the model?

1. Context management and summarization
2. Disabling all conversation history
3. Increasing every prompt indefinitely
4. Removing system instructions

Correct Answer: 1

Explanation

Context management can reduce unnecessary information sent to a model during long conversations. Instead of including the complete conversation every time, the application may summarize older interactions, retain important task state, and provide only relevant recent information. This can help control token usage, latency, and context-window limitations. The application should ensure that important details are not lost during summarization. Sensitive information should also be handled according to appropriate privacy and retention requirements.

Question 183

An AI application must process structured sales data from a business database. Which approach is generally appropriate?

1. Use only an unrelated public website
2. Use an approved data connection or tool
3. Ask the model to invent missing records
4. Store all records in the system prompt

Correct Answer: 2

Explanation

Structured business information should generally be accessed through an approved data connection, API, connector, or tool rather than being manually embedded into prompts. This allows the application to retrieve current information when required and maintain appropriate access controls. For example, an agent can query authorized sales records and then use the returned information to generate a response. Authentication, authorization, data validation, and auditing should remain in place. This architecture also reduces the risk of relying on stale or fabricated business information.

Question 184

What does model drift monitoring help identify?

1. Changes in user interface design
2. Changes in AI performance or input patterns over time
3. Changes in keyboard hardware
4. Changes in screen brightness

Correct Answer: 2

Explanation

Drift monitoring helps identify meaningful changes in the data, usage patterns, or AI performance over time. For example, customer questions may change after a new product launch, causing an agent’s previous evaluation results to become less representative. Changes in source data can also affect retrieval quality. Monitoring relevant metrics and periodically evaluating the system against updated datasets can reveal these changes. When significant drift is detected, teams can investigate whether prompts, models, data sources, or workflows need to be updated.

Question 185

Which feature is important when an AI workflow performs multiple dependent steps?

1. Task state and workflow coordination
2. Screen resolution
3. File compression
4. User interface themes

Correct Answer: 1

Explanation

Multi-step AI workflows often require task state so the system knows what has already happened and what remains to be completed. Workflow coordination determines how dependent steps are executed and how information is passed between them. For example, an agent may first retrieve customer information, then validate an order, and finally prepare an approved update. Maintaining state helps prevent steps from being repeated or skipped. The workflow should also define failure handling, authorization requirements, and conditions for human escalation.

Question 186

A model produces correct answers but takes too long to respond. Which area should be investigated?

1. Data classification only
2. Performance and latency optimization
3. Password complexity
4. Document ownership

Correct Answer: 2

Explanation

When response quality is acceptable but performance is too slow, the team should investigate latency across the complete AI workflow. Possible contributors include model selection, prompt size, retrieval operations, external API calls, network delays, and excessive sequential processing. Smaller models, optimized prompts, caching, parallel operations, or more efficient retrieval may improve response time where appropriate. Changes should be evaluated carefully because reducing latency should not introduce unacceptable declines in accuracy, safety, groundedness, or other required quality measures.

Question 187

Why should an AI solution have documented ownership?

1. To identify who is responsible for operation and governance
2. To increase token usage
3. To remove monitoring
4. To allow unrestricted access

Correct Answer: 1

Explanation

Documented ownership establishes who is responsible for an AI solution’s operation, maintenance, governance, and risk management. Owners may be responsible for approving changes, reviewing monitoring results, responding to incidents, and ensuring that the solution continues to meet business requirements. Clear ownership also helps organizations determine who should approve new integrations or changes in permissions. Without defined responsibility, problems can remain unresolved because teams may be uncertain about who manages the system and its associated risks.

Question 188

Which practice helps ensure an AI application remains maintainable?

1. Avoiding documentation
2. Using version control and documented changes
3. Changing prompts without tracking them
4. Removing testing

Correct Answer: 2

Explanation

Version control and documented changes help teams understand how an AI application has evolved over time. AI solutions may contain prompts, model configurations, evaluation datasets, tools, workflows, and data-source settings that can all affect behavior. Tracking these changes makes troubleshooting and rollback easier. It also supports controlled deployment between development, testing, and production environments. Teams should record meaningful changes and evaluate them before release. This provides greater consistency and makes long-term maintenance easier as the solution grows.

Question 189

Which approach can reduce the risk of unauthorized agent actions?

1. Giving the agent unrestricted access
2. Removing authentication
3. Enforcing authorization at the tool or service level
4. Relying only on natural-language instructions

Correct Answer: 3

Explanation

Authorization should be enforced by the underlying tools and services rather than relying solely on the AI model to follow instructions. Even if an agent is told not to perform a certain operation, technical access controls should prevent unauthorized actions. Tool-level permissions can restrict which operations an agent identity is allowed to perform. Combining authentication, authorization, least privilege, and auditing creates stronger protection. This is particularly important when agents can interact with systems that contain sensitive information or support high-impact business operations.

Question 190

What is a key consideration when an AI application uses external model providers?

1. Data handling, security, availability, and service requirements
2. Only the provider’s logo
3. Only the number of website pages
4. Only the user interface font

Correct Answer: 1

Explanation

External model providers should be evaluated for factors such as data handling, security, privacy, service availability, performance, cost, and contractual requirements. Organizations should understand what information is sent to the provider and how that information is processed and protected. Availability and service limits should also be considered because they can affect application reliability. Depending on the use case, data residency and compliance requirements may be important. Vendor dependencies should be documented so the organization understands the operational implications of using an external provider.

Question 191

What can a Power Automate flow provide to an AI agent?

1. A mechanism for executing an approved business workflow
2. Unlimited administrator permissions
3. A replacement for all security controls
4. Permanent storage for model weights

Correct Answer: 1

Explanation

A Power Automate flow can provide an AI solution with a structured way to initiate an approved business workflow. For example, an agent could trigger a flow that sends a notification, creates a task, or updates an authorized business record. The flow should still enforce appropriate permissions and validate inputs before performing actions. Using workflows can help separate conversational reasoning from business-process execution. Monitoring and error handling are also important because failed flows can affect downstream business operations.

Question 192

Why should an AI system distinguish trusted instructions from user-provided content?

1. To prevent untrusted input from overriding system behavior
2. To increase document size
3. To remove user interaction
4. To disable retrieval

Correct Answer: 1

Explanation

User-provided content should not automatically be treated as trusted instructions because it may contain attempts to manipulate the agent’s behavior. Separating trusted system instructions from untrusted input helps preserve the intended control structure of the application. Additional protections should include tool authorization, input handling, output validation, and security testing for prompt injection. The model should not be given unrestricted authority simply because a user requests an action. Technical controls should enforce important boundaries independently of the model’s generated text.

Question 193

Which factor should be reviewed when an AI agent accesses data from multiple departments?

1. Data ownership and access permissions
2. Screen resolution
3. Prompt font size
4. Keyboard configuration

Correct Answer: 1

Explanation

When an agent accesses information from multiple departments, the architecture should clearly define data ownership and access permissions. Different departments may have different authorization requirements, retention policies, or sensitivity classifications. The agent should not automatically gain access to all connected data simply because it can technically retrieve it. Identity-based authorization, data classification, and least-privilege controls should determine what information is available. Access should also be monitored so organizations can identify inappropriate or unexpected data usage.

Question 194

What is one purpose of adversarial testing for an AI agent?

1. To test how the system behaves under malicious or unexpected inputs
2. To increase database size
3. To eliminate evaluation
4. To improve monitor resolution

Correct Answer: 1

Explanation

Adversarial testing evaluates how an AI system behaves when exposed to malicious, misleading, unusual, or unexpected inputs. Tests may include prompt injection attempts, attempts to bypass restrictions, malformed requests, sensitive-data requests, and unusual tool-use scenarios. The goal is to identify weaknesses before they cause problems in production. Results can inform improvements to prompts, access controls, tool permissions, input handling, and monitoring. Adversarial testing should be repeated when significant changes are made to the model or application architecture.

Question 195

An AI solution must comply with strict data residency requirements. What should architects evaluate?

1. Where data is stored and processed
2. Only the model’s response style
3. Only the prompt length
4. Only the application’s theme

Correct Answer: 1

Explanation

Data residency requirements can determine where information may be stored, processed, or transferred. AI architectures should therefore identify the locations used by models, retrieval services, storage systems, logging platforms, and external integrations. The team should verify that the selected services and configurations satisfy applicable organizational and regulatory requirements. Data movement between services should also be understood because sensitive information may cross boundaries during inference or tool calls. Residency requirements should be considered before deployment rather than discovered after integration.

Question 196

Which approach is useful for comparing two AI models for the same workload?

1. Choose the model with the largest name
2. Evaluate both using the same representative criteria and dataset
3. Select randomly
4. Compare only their logos

Correct Answer: 2

Explanation

Two AI models can be compared fairly by evaluating them against the same representative dataset, workload, and predefined criteria. Relevant measures may include response quality, accuracy, groundedness, latency, cost, safety, and task completion. Using consistent test conditions makes the results more meaningful than relying on general model descriptions. The evaluation should reflect the actual business requirements of the application. A model that performs well in one type of task may not provide the same results for another workload.

Question 197

What should happen when an AI agent reaches a permission boundary?

1. It should bypass the restriction
2. It should request or use an authorized path
3. It should expose restricted data
4. It should disable authentication

Correct Answer: 2

Explanation

When an agent reaches a permission boundary, it should not attempt to bypass the restriction. The system should either use an authorized path, request appropriate human intervention, or inform the user that the requested operation cannot be completed. Permission boundaries should be enforced technically through identity and authorization controls. This prevents the language model from becoming a mechanism for circumventing security policies. Clear escalation procedures can help users complete legitimate tasks when additional authorization is genuinely required.

Question 198

Which capability can help an AI application maintain a record of significant agent actions?

1. Auditing and logging
2. Prompt expansion
3. Model temperature
4. Document formatting

Correct Answer: 1

Explanation

Auditing and logging can maintain records of significant events and actions performed by an AI application. Depending on the design, records may include tool calls, authorization decisions, workflow events, errors, and other operational information. Audit data can support troubleshooting, security investigations, compliance activities, and accountability. Logging should be carefully designed to avoid unnecessary exposure of sensitive information. Appropriate retention, access controls, and monitoring should be applied so that audit records remain useful without creating additional privacy or security risks.

Question 199

What is an important reason to define clear acceptance criteria for an AI solution?

1. To determine whether the solution meets its required outcomes
2. To prevent all future changes
3. To eliminate user testing
4. To guarantee zero failures

Correct Answer: 1

Explanation

Acceptance criteria define the measurable conditions an AI solution should satisfy before it is considered ready for a particular stage or deployment. Criteria may cover task accuracy, response time, safety, security, cost, reliability, or business outcomes. Clear criteria make testing more objective and help teams identify gaps before release. They also provide a basis for stakeholder approval. Acceptance criteria should be derived from the actual business and technical requirements rather than being based solely on the model’s general capabilities.

Question 200

Which activity should continue after an AI solution is deployed?

1. Monitoring and periodic evaluation
2. Removing all logs
3. Disabling security controls
4. Stopping all testing

Correct Answer: 1

Explanation

AI solutions require ongoing monitoring and periodic evaluation after deployment because models, data, user behavior, external services, and business requirements can change. Teams should monitor relevant indicators such as performance, errors, cost, latency, safety, retrieval quality, and user feedback. Periodic evaluation can identify regressions that were not visible during initial testing. Production monitoring also helps detect operational incidents and unexpected behavior. Continuous lifecycle management ensures that the AI solution remains aligned with its intended business and governance requirements.