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

 

Question 281

What is the primary purpose of agent orchestration?

1. Store model embeddings
2. Coordinate agents, tools, and workflow steps
3. Encrypt every user message
4. Replace all business applications

Correct Answer: 2

Explanation

Agent orchestration coordinates the different components involved in completing an AI-driven task. These components may include specialized agents, models, tools, data sources, and workflow steps. The orchestrator can determine which component should act, maintain task state, and manage dependencies between operations. Effective orchestration also needs to handle errors, permissions, and escalation paths. It is especially important in multi-step or multi-agent solutions where simply generating a response is not enough to complete the required business process.

Question 282

Which capability allows an agent to retain information about an ongoing task?

1. Task state
2. Data residency
3. Model catalog
4. API throttling

Correct Answer: 1

Explanation

Task state represents information about the current progress and status of an agent’s workflow. It can include completed steps, pending actions, retrieved information, decisions, and other temporary details required to continue the task correctly. Maintaining task state is especially important for multi-step workflows and agent handoffs. Without it, an agent may lose track of what has already happened or repeat actions unnecessarily. Task state should be managed securely and should contain only information needed to support the workflow.

Question 283

An agent repeatedly calls an API even though the requested information has not changed. Which technique can reduce unnecessary calls?

1. Fine-tuning
2. Caching
3. Prompt injection
4. Model expansion

Correct Answer: 2

Explanation

Caching can reduce unnecessary API calls by temporarily storing results that are safe to reuse. When the same information is requested repeatedly and remains valid for a defined period, the agent can retrieve the result from the cache instead of contacting the backend service every time. This can reduce latency, API consumption, and operating costs. Cache expiration and invalidation rules are important because stale information can cause incorrect responses. Sensitive data should also have appropriate access controls and retention policies when caching is used.

Question 284

Which component is responsible for converting text into a numerical representation for semantic retrieval?

1. API gateway
2. Embedding model
3. Authorization service
4. Workflow engine

Correct Answer: 2

Explanation

An embedding model converts text or other supported content into numerical vectors that capture semantic relationships. These vectors can then be stored and compared to a vector representation of a user’s query. Similar vectors generally indicate related meaning, enabling semantic retrieval even when the wording differs. Embeddings are commonly used in RAG architectures and other search scenarios. The embedding model itself does not determine whether a user is authorized to access the retrieved information, so retrieval permissions must be handled separately.

Question 285

A company wants different agents to communicate using a standardized protocol. Which technology can support this scenario?

1. MCP
2. Data masking
3. Token budgeting
4. Model compression

Correct Answer: 1

Explanation

The Model Context Protocol, or MCP, provides a standardized approach for connecting AI applications with external tools and data sources. It can help establish consistent interfaces through which models or agents discover and interact with supported capabilities. Standardized protocols can reduce custom integration work and make tool ecosystems easier to maintain. Security controls remain essential because protocol support does not automatically grant authorization. Organizations should still control which tools are available, what data can be accessed, and which actions an agent is permitted to perform.

Question 286

What should an agent do if a required tool returns an invalid response?

1. Use the invalid result without checking
2. Validate the result and apply an error-handling path
3. Grant the tool additional permissions
4. Delete the user’s request

Correct Answer: 2

Explanation

An agent should validate tool results before using them in subsequent reasoning or actions. If a required result is malformed, incomplete, or inconsistent with expected values, the workflow should use an appropriate error-handling path. Depending on the situation, the system may retry, use a fallback source, request clarification, or escalate to a human. Blindly trusting invalid tool output can lead to incorrect decisions or unintended actions. Validation and controlled error handling are therefore important parts of reliable agent architecture.

Question 287

Which approach helps prevent a model from receiving more retrieved information than necessary?

1. Targeted retrieval
2. Unrestricted indexing
3. Maximum context expansion
4. Random document selection

Correct Answer: 1

Explanation

Targeted retrieval limits the information supplied to the model to content that is relevant to the current request. Retrieval filters, metadata, query rewriting, ranking, and access controls can help identify appropriate evidence. Sending excessive information can increase token consumption and introduce irrelevant or conflicting content. Targeted retrieval also supports better response quality because the model can focus on a smaller set of useful evidence. The retrieval process should balance relevance, completeness, permissions, and freshness when selecting context.

Question 288

Which capability is most useful for automatically triggering a business workflow after an agent completes a task?

1. Power Automate flow
2. Vector embedding
3. Model catalog
4. Prompt template

Correct Answer: 1

Explanation

A Power Automate flow can connect an AI-driven process with business workflows and applications. For example, after an agent completes a support classification, a flow could create a task, notify an employee, update a record, or start an approval process. The workflow should still enforce authentication, authorization, validation, and appropriate business rules. Automation should not assume that every agent output is trustworthy. Where actions have significant impact, approval or additional validation may be required before the workflow executes.

Question 289

Why is evaluation needed after changing an AI model?

1. To verify that the new model still meets required quality and safety criteria
2. To increase user passwords
3. To remove monitoring
4. To guarantee zero latency

Correct Answer: 1

Explanation

Changing an AI model can affect accuracy, groundedness, latency, cost, safety, and other aspects of application behavior. Evaluation provides evidence about whether the new model continues to meet defined requirements. A representative evaluation dataset should include common scenarios, edge cases, and important business workflows. Comparing results against a previous baseline can reveal regressions or improvements. Production deployment should follow established evaluation gates rather than assuming that a newer or different model will automatically provide better results for the specific workload.

Question 290

What is a benefit of using a smaller language model for simple tasks?

1. It always provides better reasoning
2. It can reduce cost and latency when quality is sufficient
3. It eliminates the need for testing
4. It removes all security risks

Correct Answer: 2

Explanation

Smaller language models can be appropriate for straightforward workloads such as classification, extraction, routing, or simple summarization when they meet the required quality level. They may consume fewer resources and provide lower latency or cost than larger models. However, the decision should be based on evaluation rather than model size alone. Teams should compare representative workloads for accuracy, reliability, safety, latency, and cost. More complex requests can still be routed to more capable models when necessary.

Question 291

Which design helps an agent recover when a downstream service is unavailable?

1. Fallback handling
2. Unlimited retries
3. Removing timeouts
4. Ignoring service errors

Correct Answer: 1

Explanation

Fallback handling provides an alternative path when a required service is unavailable or fails repeatedly. Depending on the scenario, the fallback could use another approved service, a cached result, a simpler workflow, or human escalation. Retry policies should be limited and combined with appropriate backoff so they do not overload an unhealthy service. Timeouts are also important because waiting indefinitely can reduce overall system availability. A well-designed fallback strategy improves resilience while maintaining security and business requirements.

Question 292

An agent must update a record only if a business rule is satisfied. Where should the rule be enforced?

1. Only in the user’s message
2. At the business logic or service boundary
3. Only in the model’s training data
4. Only in the system prompt

Correct Answer: 2

Explanation

Important business rules should be enforced through application or service-level controls rather than relying solely on model instructions. A prompt can tell an agent to follow a rule, but a model may still misunderstand or incorrectly apply it. Enforcing the rule at the business logic or service boundary provides a technical safeguard before the transaction occurs. The agent can assist with determining the intended operation, while the backend verifies eligibility and authorization. This layered approach improves reliability and reduces the impact of model errors.

Question 293

What is the purpose of a system prompt in an agent application?

1. Define high-level behavior, instructions, and boundaries
2. Store API passwords
3. Replace access-control systems
4. Increase database capacity

Correct Answer: 1

Explanation

A system prompt provides high-level instructions that guide an agent’s behavior, role, response requirements, and operational boundaries. It can establish how the agent should interpret requests, use tools, handle uncertainty, and respond to specific situations. However, prompts are not substitutes for technical security controls. Sensitive credentials should never be stored in prompts, and permissions should be enforced by the underlying services. System prompts should be versioned and evaluated because changes can significantly affect agent behavior.

Question 294

Which capability can help an agent understand the relationship between a user request and relevant enterprise documents?

1. Semantic retrieval
2. Password rotation
3. API throttling
4. Disaster recovery

Correct Answer: 1

Explanation

Semantic retrieval uses representations such as embeddings to identify documents that are conceptually related to a user’s request. Unlike exact keyword matching, semantic retrieval can find relevant content even when the user and document use different wording. It is commonly used in RAG systems to provide evidence to a language model. Retrieval quality should be evaluated using representative queries, and access controls must be applied so that semantic similarity never exposes documents that the requesting user is not authorized to access.

Question 295

Why should agent actions be logged?

1. To support auditing, troubleshooting, and accountability
2. To increase model temperature
3. To eliminate authentication
4. To make every response longer

Correct Answer: 1

Explanation

Logging agent actions provides visibility into important operations performed during an AI workflow. Records can help teams investigate failures, understand tool usage, verify approvals, identify unauthorized activity, and support compliance requirements. Logs should be designed to capture useful operational information without unnecessarily storing sensitive data. Access to logs should also be restricted appropriately. Action logging is particularly important for agents that can modify records, call external services, or perform other consequential operations because it provides an audit trail of system behavior.

Question 296

What should an AI architecture consider when integrating a third-party model provider?

1. Only the provider’s logo
2. Data handling, security, compliance, cost, and availability
3. Only the model’s name
4. User interface font selection

Correct Answer: 2

Explanation

Third-party model integration requires evaluation of how data is transmitted, processed, stored, and protected by the provider. Organizations should also consider contractual terms, privacy requirements, compliance obligations, model availability, latency, pricing, and operational dependencies. The architecture should define what information can be sent to the external provider and whether sensitive data requires additional controls. Vendor dependency and fallback options may also be important. These considerations help ensure that third-party AI services fit the organization’s technical and governance requirements.

Question 297

A company wants to compare two AI models for the same customer-support workload. What should it use?

1. A representative evaluation dataset and consistent metrics
2. Random user opinions only
3. Different test cases for each model
4. Model names alone

Correct Answer: 1

Explanation

A fair model comparison requires the models to be tested against the same representative workload using consistent evaluation criteria. Metrics may include accuracy, groundedness, relevance, safety, latency, token usage, and cost, depending on the business requirements. Testing different examples for each model can produce misleading comparisons. A baseline and clearly defined acceptance criteria make results easier to interpret. The final selection should reflect the requirements of the actual application rather than assuming that one model is universally better for every workload.

Question 298

Which architecture principle helps prevent an AI agent from becoming responsible for unrelated business functions?

1. Clear agent boundaries
2. Unlimited tool access
3. Shared administrator credentials
4. Unrestricted data access

Correct Answer: 1

Explanation

Clear agent boundaries define which responsibilities, tools, data, and actions belong to a particular agent. This makes the solution easier to secure, test, monitor, and maintain. For example, an HR agent should not automatically receive access to financial transaction systems simply because both functions exist within the same organization. Narrow responsibilities also support least privilege and make human escalation easier to design. Well-defined boundaries are particularly valuable in multi-agent architectures where several specialized agents collaborate.

Question 299

What is a key purpose of continuous AI monitoring?

1. Detect operational, quality, security, and performance issues after deployment
2. Remove the need for testing
3. Guarantee perfect responses
4. Disable system alerts

Correct Answer: 1

Explanation

Continuous monitoring helps teams understand how an AI system behaves after it enters production. Monitoring can cover response quality, latency, failures, tool usage, cost, safety events, and important business metrics. It can reveal problems that were not visible during development or testing, including changes in data, user behavior, service availability, or model performance. Monitoring should be connected to defined alert thresholds and incident-response procedures. It complements pre-production evaluation rather than replacing it.

Question 300

An organization is planning to retire an AI agent. What should happen before decommissioning it?

1. Remove all documentation first
2. Ignore dependent workflows
3. Identify dependencies and complete an orderly retirement process
4. Grant the agent permanent permissions

Correct Answer: 3

Explanation

Retiring an AI agent requires more than simply disabling its model endpoint. The organization should identify dependent applications, workflows, users, data sources, tools, and integrations before decommissioning it. Access permissions and credentials should be removed or revoked, while required records and audit information should be retained according to applicable policies. Users should also be informed about replacement processes when necessary. A controlled retirement process reduces service disruption, prevents orphaned integrations, and ensures that unnecessary access does not remain active after the agent is removed.