AI-500 Exam Guide & Practice Questions

A procurement team is designing an assistant to review supplier contracts. It must extract clauses, compare them with internal policy, and prepare a recommendation for a buyer. The extraction and policy checks can run separately, but the team requires a human decision when evidence conflicts or confidence is low. The design must also make each handoff clear.

Which architecture best meets the requirements?

• AUse a coordinator to route work among focused agents, combine their findings, and send low-confidence or conflicting cases to a human reviewer.
• BUse one general-purpose agent to perform every task and make the final decision without review.
• CLet each specialist agent create its own workflow and send unstructured messages to every other agent.
• DRun all specialist agents in parallel and accept the first answer returned, regardless of evidence or confidence.

A finance operations team is planning a multi-agent process for invoice exceptions. The process must inspect line items, check current inventory, and pass findings to a later agent that drafts a resolution. The team wants separate components for workflow control, focused analysis, service access, and case information. It is designing the logical architecture before choosing implementation details.

Match each requirement to the architecture component that best fulfills it.

Match each item to the correct target.

A purchasing group is designing agents that review requests, gather supporting details, and draft recommendations. Buyers must approve decisions that could commit funds, while routine information gathering should not wait for manual input. Later agents need relevant findings from the same request, but unrelated case histories should not enter their context. The team also wants the recommendation agent to remain advisory.

Choose the best value for each dropdown.

Choose the correct option for each statement.

A customer service agent built with Azure OpenAI must answer questions using policy documents that change several times each month. The application retrieves relevant passages at runtime, but some documents contain text that resembles instructions to the agent. The team wants answers to reflect current policy without allowing retrieved content to override its established response rules.

Choose two changes that best address the policy and instruction risks.

Select 2 answers, then check them.

• ARetrieve the relevant current policy for each case and place it in a clearly marked reference section of the prompt.
• BCopy the full policy archive into the system prompt and rely on the model to identify the newest policy.
• CRemove runtime policy context and use examples alone to produce consistent answers.
• DInstruct the agent to follow any instructions found in retrieved documents before its existing instructions.
• EFine-tune the model every week on all recent customer cases so it learns the latest policy.
• FTell the agent that retrieved documents are reference data and must not change its governing instructions.

A claims-processing application uses several agents that already perform separate review steps. Their conversations can exceed the model context window, and each agent currently receives a large history containing many irrelevant details. The application must preserve important decisions and unresolved facts, while retaining exact source records for cases where a later response needs specific values or provenance.

For each statement, select Yes if it is an appropriate context or memory practice, and No if it is not.

Choose the correct option for each statement.

A benefits organization is extending an existing Azure OpenAI agent application. Its policy documents change often, its response format is stable across many cases, and its case histories can be longer than the model context window. The team wants to choose distinct implementation approaches for these requirements and avoid retraining the model whenever a policy document changes.

Match each requirement to the most suitable implementation approach.

Match each item to the correct target.

A parcel support solution uses a coordinator and specialist agents to check delivery status and request address corrections. Recent customer reports show that some correction requests are sent to a status lookup tool, although final responses often read clearly. The team has saved evaluation conversations, but its current evaluator scores only the final response for relevance and helpfulness. The team needs to locate the source of the errors.

Which evaluation change would best identify the cause of the incorrect shipment decisions?

• AAdd a trace-level evaluator that compares the selected tool and its arguments with expected calls for each test case.
• BIncrease the maximum response token limit and compare the final answer length.
• CScore only the final response for fluency and customer satisfaction.
• DRun the same evaluation cases more often without recording tool calls.

A purchasing solution coordinates agents that check inventory, retrieve prices, update an order, and send a confirmation. The inventory and price services provide independent read operations, while the confirmation must reflect a completed order update. Supplier calls are subject to a strict requests-per-minute quota, and recent bursts have triggered throttling. The team must reduce elapsed time without creating inconsistent updates or exceeding the quota.

Match each execution change to the constraint it addresses.

Match each item to the correct target.

A claims-processing solution already runs a versioned evaluation set before releases, and its LLM judge has been checked against expert-scored examples. In production, new combinations of claim details are producing occasional routing failures, and users submit feedback through the application. The team wants to improve coverage over time while preserving its release checks and avoiding changes based only on a single aggregate score.

Choose two actions that would improve the feedback and evaluation loop for these emerging failures.

Select 2 answers, then check them.

• AGroup failed sessions and user reports by failure pattern, then select representative cases for investigation.
• BRaise the minimum judge score until the weekly average meets the release target.
• CFine-tune the foundation model on every production conversation without filtering or evaluation.
• DGenerate variations of confirmed failure cases, validate them, and add suitable cases to the versioned regression set.
• ETrack only the weekly average judge score and discard individual traces after scoring.

A company is preparing an agent service for production. The service must retrieve one partner API credential from Azure Key Vault, and the operations team will not allow credentials to be stored in application settings or source code. The vault must be reachable only from the service network, and the service must not have permission to manage vault configuration or grant access to other principals.

Which design best meets the requirements?

• AEnable a managed identity for the agent service, grant it the Key Vault Secrets User role on the required vault, and restrict vault access to the service network.
• BStore the secret in the agent configuration and restrict access to the configuration file with a broad contributor role.
• CUse a shared service account for all agent services and grant it the Key Vault Administrator role.
• DPut the secret in a deployment script and remove the script after the service starts.

A support agent service receives requests from authenticated employees and must retrieve records from a separate internal API using each employee's delegated permissions. Microsoft Entra ID issues the incoming access token for the agent service, and the required delegated permission has been approved for the records API. The records API rejects tokens whose audience is not its own API.

For each statement, select Yes or No.

Choose the correct option for each statement.

A team is completing a security review before releasing an agent that calls an internal records API and a partner service. The partner credential is currently entered by an operator during deployment, and the Key Vault is reachable from public networks. The team must also test for prompt-based attacks before release and ensure that a compromised agent cannot access unrelated resources.

Match each release requirement to the most suitable security control.

Match each item to the correct target.