Free AB-100: Agentic AI Business Solutions Architect (beta) Exam Questions and Answers

Contoso uses Microsoft Dynamics 365 Contact Center with a voice channel. You need to recommend a design for an agent that answers incoming calls, collects the caller's account number and reason for calling, and then hands the call to a human representative who must see the collected details. The solution must minimize development effort. What should you recommend?

• AA. Build a Bot Framework bot on Azure Communication Services and route calls to it by using a custom Azure function that copies the details.
• BB. Create a Foundry Tools speech workflow that transcribes each call and sends the transcript to the representative by email.
• CC. Publish a Copilot Studio agent to a Teams channel and forward calls from a Teams auto attendant to the representatives' queue.
• DD. Set a Copilot Studio agent as the first responder on the voice workstream and pass the collected values as context variables on escalation.

Contoso has a Microsoft Copilot Studio agent that answers employee questions about HR policies. The agent uses generative answers over a Microsoft SharePoint library of policy documents and has two custom topics for leave requests. Analytics for the last 30 days show the following: Sessions: 14,000 Resolved: 3,500 Escalated: 1,400 Abandoned: 9,100 Most abandoned sessions end immediately after a generative answer, and 38 percent of the thumbs-down feedback is attached to those answers. You need to identify the most likely cause of the abandonment. What should you review first?

• AA. Capacity consumption by environment in the Power Platform admin center
• BB. Customer satisfaction scores for the sessions that the agent escalated
• CC. Trigger frequency for the Conversation Start system topic
• DD. Generative answer quality by knowledge source in Copilot Studio analytics

Fabrikam trains a custom classification model to flag fraudulent card transactions. Only 0.5 percent of the transactions in the validation set are fraudulent. Missing a fraudulent transaction costs far more than reviewing a legitimate one. Which metric should you define as the primary validation criterion for the model?

• AA. Mean squared error
• BB. Recall
• CC. Accuracy
• DD. Precision

Contoso uses Microsoft Dynamics 365 Customer Service. Support representatives use Copilot to ask questions about products. Copilot currently answers only from knowledge articles stored in Dataverse. Contoso also keeps 3,000 product manuals in a Microsoft SharePoint site, and representatives want Copilot to answer from those manuals. The solution must minimize development effort. What should you do?

• AA. Convert each manual into a Dataverse knowledge article and retire the SharePoint site
• BB. Fine-tune a model in Microsoft Foundry on the manuals and register it as the Copilot model
• CC. Add the SharePoint site as a knowledge source for Copilot in the Customer Service admin center
• DD. Build a Copilot Studio agent that calls the SharePoint REST API and publish it to the service workspace

Contoso has a Microsoft Copilot Studio agent that answers questions from a SharePoint knowledge source. The knowledge source is updated every week. You need to recommend a regression test process that runs after each update. The process will use a test set of 200 representative questions that each have an expected answer. The agent's answers are free text, so the wording differs between runs. Which test method should you recommend?

• AA. Exact match evaluation of each response against its expected answer in the test set
• BB. Manual testing of a sample of questions in the test pane after each update
• CC. Compare meaning evaluation of each response against its expected answer in the test set
• DD. An alert on the abandoned session rate in production analytics after each update

Fabrikam uses Microsoft Dynamics 365 Customer Service, Dynamics 365 Field Service, and Dynamics 365 Finance. Copilot summarizes cases in Customer Service, and an agent recommends technician schedules in Field Service. You need to design an end-to-end test scenario for the case-to-invoice process that validates the AI features. The scenario must use only masked data and must not affect production. Which four actions should you perform in sequence? To answer, move the appropriate actions from the list of actions to the answer area and arrange them in the correct order. Actions: - Complete the work order and verify that the invoice is posted in Finance with the expected amount - Copy production data into a separate sandbox environment for each app - Load masked test data into a single sandbox environment that contains all three apps - Convert the case to a work order and verify that the recommended technician and time slot meet the scheduling rules - Validate the case summary prompt in the Copilot Studio test pane - Create a case from a test customer email and verify that the Copilot case summary matches the email

Put the items in order, then reveal.

Fabrikam is starting its first AI initiatives and plans to follow the AI adoption process in the Cloud Adoption Framework for Azure. The first use cases will use Microsoft 365 Copilot. Later use cases will use custom agents built in Microsoft Foundry. For each of the following statements, select Yes if the statement is true. Otherwise, select No. 1. Fabrikam should identify and prioritize AI use cases and define success metrics as part of its AI strategy. 2. Fabrikam should prepare an AI landing zone for the Foundry workloads during the Ready phase. 3. Fabrikam can postpone AI governance policies until it starts to build custom agents, because the first use cases use only Microsoft 365 Copilot. 4. Fabrikam should build custom models in Foundry by default and consider software as a service (SaaS) agents only when a custom model cannot meet a requirement.

Answer each statement, then reveal.

1. Fabrikam should identify and prioritize AI use cases and define success metrics as part of its AI strategy.YesNo
2. Fabrikam should prepare an AI landing zone for the Foundry workloads during the Ready phase.YesNo
3. Fabrikam can postpone AI governance policies until it starts to build custom agents, because the first use cases use only Microsoft 365 Copilot.YesNo
4. Fabrikam should build custom models in Foundry by default and consider software as a service (SaaS) agents only when a custom model cannot meet a requirement.YesNo

Fabrikam is building a prompt library for sales managers who use Microsoft 365 Copilot to review customer meetings. Before a prompt is published, testers must confirm that it states one clear goal, gives specific context and a source, and sets expectations for the output. Which prompt should the testers approve?

• AA. Using the transcript of Tuesday's Contoso renewal meeting, summarize it, draft the follow-up email, update the forecast, and score the deal.
• BB. Using the transcript of Tuesday's Contoso renewal meeting, tell me what happened and anything a sales director might want to know.
• CC. Write a five-bullet summary for a sales director that lists the open action items and owners from my recent customer meetings.
• DD. Using the transcript of Tuesday's Contoso renewal meeting, write a five-bullet summary for a sales director with action items and owners.

Fabrikam has a product support agent in Microsoft Foundry. Tracing is enabled, and about 8,000 conversations are logged each day. Users report that some answers include details that do not appear in the product documentation that the agent retrieves. You need to identify the affected responses across all logged conversations. The solution must minimize manual review effort. What should you do?

• AA. Have subject matter experts score a weekly random sample of conversations against the documentation
• BB. Review the content filter annotations on each response for protected material detections
• CC. Run the groundedness evaluator in Microsoft Foundry against the traced conversations
• DD. Create an Azure Monitor alert on agent request latency and failed request counts for the project

Contoso runs a customer service agent in Microsoft Foundry that calls a model deployment named chat-prod. Users report slow responses during the morning peak. Monitoring data for the peak hour shows the following: Deployment type: Standard Tokens-per-minute (TPM) quota: 300,000 Peak TPM consumed: 296,000 Average prompt tokens per request: 720 Average completion tokens per request: 180 HTTP 429 responses: 14 percent of requests Content filter blocks: 0.1 percent of requests You need to recommend a change that reduces the slow responses. What should you recommend?

• AA. Add retry logic with exponential backoff to the agent
• BB. Increase the TPM quota for chat-prod
• CC. Fine-tune the model to return shorter completions
• DD. Increase the max_tokens value for each request