Agentic AI Business Solutions Architect Practice Test & Intelligent Application Sandbox

The requirement is to secure agents in a Microsoft Power Platform environment so that they:

use only approved connectors and services

are prevented from accessing sensitive data

The correct recommendation is B. Deploy data loss prevention (DLP) policies in Power Platform.

Why B is correct:

DLP policies in Power Platform are specifically designed to control which connectors can be used together and which services are allowed in an environment. They help administrators classify connectors as business or non-business and restrict unsafe data flows. This directly supports both requirements:

limiting agents to approved connectors/services

preventing data from being exposed through unapproved or risky connector usage

Why the other options are not correct:

A. Enable customer-managed keys in Microsoft Dataverse This helps with encryption control, not with restricting connector usage or preventing data movement through agents.

C. Configure Azure Monitor to capture connector activity logs Monitoring logs is useful for visibility, but it does not enforce connector restrictions or prevent sensitive data access.

D. Enable a Microsoft Dataverse audit Auditing records activity after the fact. It does not proactively block unapproved connectors or sensitive data exposure.

To improve performance → Move to a multi-agent architecture

To improve accuracy → Add a grounding data source

The current design uses a single agent and a single prompt to complete a series of tasks . That is often a bottleneck. When one agent is responsible for everything, it has to manage multiple steps, multiple reasoning modes, and multiple task transitions in one flow. This commonly leads to:

slower response times

task overload

incomplete outputs

reduced efficiency as complexity grows

Moving to a multi-agent architecture helps performance because tasks can be separated by function. For example:

one agent can handle task planning

another can retrieve domain knowledge

another can perform structured reasoning

another can prepare the final response

From an agentic AI systems perspective, decomposition improves throughput and execution quality. Instead of one overloaded agent trying to do everything, specialized agents handle narrower responsibilities. That often reduces latency in practical enterprise designs and improves the reliability of task completion.

This also addresses the symptom of incomplete results , because a multi-agent architecture can break a large workflow into smaller, controlled substeps.

Why “Add a grounding data source” improves accuracy

The agent struggles with domain-specific reasoning . That strongly suggests it lacks sufficient domain context during inference.

The best way to improve accuracy in this case is to add a grounding data source .

Grounding means giving the model access to trusted, relevant business knowledge at runtime, such as:

internal documentation

product specifications

policy manuals

knowledge bases

industry-specific reference data

This improves domain-specific reasoning because the model no longer relies only on general pretrained knowledge. Instead, it can anchor its responses in authoritative content.

From an AI business solutions standpoint, grounding is one of the most important mechanisms for improving:

factual relevance

domain accuracy

consistency

trustworthiness

explainability in enterprise contexts

When a model is inaccurate because it lacks business context, grounding is usually a better first fix than simply scaling model size.

Why the other actions are not the best fit

Add a prebuilt connector

A prebuilt connector helps with integration to systems and services, but it does not directly solve slow reasoning, incomplete output, or weak domain-specific reasoning unless the issue is specifically missing access to an external system. That is not the main problem described here.

Upgrade to a larger generative AI model

A larger model may sometimes improve reasoning quality, but it usually comes with higher cost and often slower response times , which works against the stated performance issue. It is not the best recommendation when the current agent is already slow.

Also, when domain-specific reasoning is the problem, grounding is usually more efficient and more controllable than simply choosing a larger model.

Expert reasoning shortcut

Use this exam logic:

Slow and overloaded single agent handling many tasks → move to multi-agent architecture

Weak domain-specific reasoning → add grounding data source

Need system integration → prebuilt connector

Need raw generative capability increase, but can accept more cost/latency → larger model

The goal is to use generative AI to answer questions based on a subset of corporate data , and to ensure that this data is available as a grounding data source .

The solution includes:

Dataverse tables

a Power BI workspace

a canvas app

a Power BI semantic model using DirectQuery to Dataverse

The best action is D. Endorse the semantic model.

Why D is correct:

Endorsing a semantic model makes it a trusted, discoverable enterprise data asset

It is the appropriate step when you want approved corporate data to be used reliably by downstream AI and analytics experiences

It fits the requirement of exposing a curated subset of data as a grounding source rather than duplicating or manually exporting it

Why the other options are not correct:

A. Export the semantic model Exporting does not make it a governed grounding source.

B. Share WS1 Sharing the workspace grants access, but it does not establish the semantic model itself as the trusted data source for grounding.

C. Populate a Dataverse table The data already exists and is modeled through the semantic layer; creating another table is unnecessary for this requirement.