An insurer wants to introduce agents in Dynamics 365 and Power Platform to automate claim triage, produce trend analytics for supervisors, and…

This design uses a multi-agent approach, which is best practice for separating concerns and applying specific controls. A task agent with approval gates (B) directly meets the requirement for human-in-the-loop on high-risk actions like payouts. An analytics agent grounded on a Fabric semantic model (D) provides a secure and efficient way to generate trend analytics while respecting data residency. A decision-support agent with policy tool gating (E) allows for audited, controlled assistance to human adjusters, aligning with the need for both support and oversight. This composite architecture respects all constraints, including security, auditing, and environment isolation.

A. A single autonomous agent is a poor design for separating high-risk and low-risk tasks, complicating auditing, security, and adherence to varied business rules.

C. Shared memory without separation controls directly violates the requirements for data regulation, in-region residency, and dev/test/prod isolation, creating a major compliance risk.

F. A prompt-only copilot cannot perform the required actions, such as claim triage or initiating payouts, as these require tool use to interact with other systems.

1. Microsoft Power Platform. (2024). Responsible AI FAQs for Power Automate. In "Process and task mining," Section: "What are the capabilities of the AI-powered process and task mining feature?". This document discusses the importance of human oversight for AI-generated suggestions, aligning with options B and E.

2. Microsoft Copilot Studio. (2024). Generative answers with Search and summarize content. Section: "Boost conversational coverage over multiple sources". This documentation explains grounding on trusted data sources, which is the principle behind option D's use of a Fabric semantic model for reliable analytics.

3. Microsoft Cloud Solution Architects. (2023). Microsoft Power Platform Well-Architected Framework - Reliability. Section: "Design for evolution". The framework advocates for decomposing workloads into smaller, independent components (like separate agents for different tasks) to improve reliability and maintainability.