AB-100 Design AI solutions explained
The Design AI-powered business solutions domain is 25–30% of AB-100 and asks you to choose the right kind of agent, the right Copilot Studio construct, and the right extension or prebuilt feature for a stated requirement. It has three parts: designing AI and agents, designing extensibility, and orchestrating the configuration of prebuilt agents and apps.
Design AI and agents
Choosing an agent type
The outline lists task agents, autonomous agents and prompt-and-response agents. Treat them as a ladder of independence:
| Type | Triggered by | Typical use |
|---|---|---|
| Prompt and response | A user’s question | Answer from knowledge, summarise, draft |
| Task agent | A user or process request | Carry out a defined sequence of steps |
| Autonomous agent | An event or schedule, with little human input | Monitor, decide and act within set limits |
The more autonomy, the more you must design limits, escalation and audit. If a scenario says a person must approve every action, an autonomous design is probably wrong.
Copilot Studio choices
- Topics, including fallback. Topics handle defined conversations; the fallback topic decides what happens when nothing matches. A good design hands off or asks the user to rephrase, rather than guessing.
- Orchestration. The outline asks when to use standard natural language processing, conversational language understanding, or generative AI orchestration. Fixed intents and tight control favour the first two. Open-ended requests across many tools favour generative orchestration.
- Agent flows and prompt actions. Flows run deterministic steps; prompt actions use a model for a bounded task, such as extracting fields from text.
- Knowledge and grounding. Design how data is processed and grounded, so the agent answers from approved sources.
Dynamics 365 and Power Platform
You are expected to design business terms and customisations for Copilot in Dynamics 365 apps for customer experience and service, connectors for Copilot in Dynamics 365 Sales, and agent integration with Dynamics 365 Contact Center channels. For Power Platform, expect AI in a canvas app, code-first generative pages with an agent feed, and the Power Platform Well-Architected Framework applied to intelligent workloads.
Design extensibility
| Need | Direction |
|---|---|
| Behaviour no catalogue model delivers | Custom models in Microsoft Foundry |
| Agent inside a user’s daily tools | Agents in Microsoft 365 Copilot, including Teams and SharePoint |
| Reach tools through a standard interface | Model Context Protocol in Copilot Studio |
| Operate an application that has no API | Computer use in Copilot Studio |
| Richer conversation | Reasoning and voice mode in Copilot Studio |
The pattern across the table is to pick the lightest option that meets the requirement. Computer use is useful when no API exists, but a proper connector or MCP tool is more robust when one does.
Orchestrate prebuilt agents and apps
This sub-area is about configuring what Microsoft already provides:
- AI features in Dynamics 365 finance and supply chain, and in customer experience and service.
- Microsoft 365 agents proposed for business scenarios, and configuring Microsoft 365 Copilot for Sales and for Service.
- Power Platform AI features, including AI hub.
- Finance and operations agent chats using extra knowledge sources, and knowledge sources for in-app help and guidance in Dynamics 365 Finance or Supply Chain Management.
Questions here are mostly matching: a described business need and the prebuilt capability that covers it with the least custom work.
Example questions
Question 1. A company wants an agent that watches incoming supplier invoices, flags mismatches against purchase orders and starts an approval, with no one starting each run. Which agent type fits?
- A. A prompt-and-response agent
- B. An autonomous agent
- C. A task agent started by a user
- D. A fallback topic
Show answer
Answer: B
An autonomous agent is triggered by events and acts without a user starting it, which matches monitoring incoming invoices. A prompt-and-response agent only answers when asked. A task agent runs when a user or process requests it, so it does not fit the no-one-starts-it requirement. A fallback topic is a Copilot Studio conversation element, not an agent type.
Want more questions like this? Full AB-100 practice tests →
Question 2. A support agent in Copilot Studio must handle five well-defined request types with exact, auditable responses, and must say so politely when a request matches none of them. What should you design?
- A. Only generative orchestration with no topics
- B. A computer-use agent for each request type
- C. One topic per request type and a fallback topic for unmatched requests
- D. Topics only, and no fallback behaviour
Show answer
Answer: C
Defined topics give predictable, auditable behaviour for the five request types, and a fallback topic handles unmatched requests. Pure generative orchestration gives less control over exact responses. Removing fallback leaves unmatched requests unhandled. A computer-use agent is for applications with no API, not for conversation control.
Want more questions like this? Full AB-100 practice tests →
Question 3. An agent must update records in a legacy desktop application that has no API and no connector. Which Copilot Studio capability should you propose?
- A. Computer use
- B. An MCP tool connecting to the legacy system
- C. A prompt action
- D. A code-first generative page
Show answer
Answer: A
Computer use lets an agent operate application interfaces directly, which suits a system with no API. An MCP tool needs a server exposing the system’s functions, which does not exist here. A prompt action uses a model for a bounded task but cannot operate the application. Generative pages build app screens, not automation of a legacy tool.
Want more questions like this? Full AB-100 practice tests →
How to revise this domain
For each agent type and each Copilot Studio construct, write one scenario where it is the right answer and one where it is not. Then map the prebuilt features to their apps. Read Plan explained if the platform choice is still unclear, and continue to Deploy explained.