In plain words
Designing a solution means deciding what to build before you build it. You read the requirements, split them into small needs, and match each need to a Power Platform part: a table, an app, a flow, a prompt, an agent or a site.
It is like an architect planning a house. Before anyone lays a brick, the architect decides where the kitchen, stairs and windows go. A mistake on paper costs minutes. A mistake in concrete costs weeks.
AB-410 adds a twist: you design with AI-enabled tools. Copilot can draft a plan, suggest tables and even propose which agents to use. Your job is to judge that draft and fix it.
Why it matters
Most failed projects fail at design, not at build. Someone builds a canvas app when staff needed a model-driven app. Someone writes a plug-in when a business rule was enough. Someone builds everything in the default environment and then cannot move it.
Good design saves licences, time and support effort. It also keeps AI features where they help, such as summarising a long ticket, and away from places where a simple rule is safer and cheaper.
How it works
1. Analyze requirements. Turn each sentence of the brief into a need, then into a component. Ask who uses it, on what device, how much data, and whether it must run without a person.
| Requirement clue | Likely component |
|---|---|
| Structured data, relationships, security by team | Dataverse tables |
| Staff work with many related records, views and forms | Model-driven app |
| Custom, mobile-friendly screens for a specific task | Canvas app |
| Something must happen when a row changes, on a schedule or on email | Cloud flow |
| Summarise, classify, extract or draft text | AI Hub prompt, prompt column or row summary |
| Answer questions in chat or act on its own | Copilot Studio agent |
| External users, such as the public, need a website | Power Pages |
2. Use AI-enabled design tools. Plan designer lets you describe a business problem in plain language, optionally with documents or images. Copilot drafts a plan with roles, user stories, a process map, a Dataverse data model and suggested apps, flows and agents. You review and edit the plan, then generate components from it. Copilot can also create tables from a description. Treat every AI draft as a first version that needs your review.
3. Evaluate built-in agents and AI features. Before building something new, check what already exists.
- Copilot chat in model-driven apps: users ask questions about their data in a side pane.
- Form fill assistance and smart paste: help users complete forms faster.
- Row summaries: AI-generated summaries of a record on forms and views.
- Autonomous agents in the agent feed: agents built in Copilot Studio can log work for review, and the model-driven app shows it in the agent feed. You add an agent to the feed from the app designer’s Agents area.
- Copilot Studio agents: for chat experiences in Teams, a website or an app.
Most of these must be enabled by an admin for the tenant and environment, and some need the modern look for model-driven apps.
4. Recommend extensibility. When low-code cannot meet a requirement, you recommend, not build, a code option. The AB-400 developer track builds them.
| Need | Extensibility option |
|---|---|
| Call an API that has no connector | Custom connector |
| A custom control on a form, such as a star rating | PCF code component |
| Server-side logic that must run in the same transaction as a save | Plug-in |
| A fully custom web front end, such as a React dashboard | Code app |
| Long-running or heavy processing outside Dataverse | Azure Functions or other Azure services |
5. Recommend environment types.
| Type | Best for | Key trait |
|---|---|---|
| Developer | One person learning or building | Created with the Developer Plan; personal |
| Sandbox | Team development and testing | Non-production; supports copy and reset |
| Production | Live business use | Permanent, backed up, fully governed |
| Trial | Short evaluations | Cleaned up after 30 days |
| Default | Personal productivity apps | One per tenant, shared by all licensed users, cannot be deleted |
| Dataverse for Teams | Simple apps inside a Teams team | Created from Teams; limited admin control |
Example: Campus Help Desk
- The brief says: students report issues on their phones; staff triage and resolve them; managers want weekly numbers; long descriptions slow staff down; some tickets arrive by email.
- You open Plan designer and paste the brief. It proposes Ticket, Category and a student table, a canvas app, a model-driven app and a flow.
- You review the plan. You replace the new student table with the standard Contact table. You keep Ticket and Category.
- You map the rest: phone screens go to the canvas app; triage goes to the Help Desk Staff model-driven app; email tickets go to a cloud flow with the Outlook "When a new email arrives" trigger; long descriptions go to a prompt column and row summaries; weekly numbers go to a model-driven dashboard.
- You evaluate built-in agents. Copilot chat in the model-driven app can answer "how many High tickets are open?" without building anything. An autonomous triage agent could log suggested categories to the agent feed for staff to confirm.
- One need remains: look up student details from the university’s Campus Directory API. No connector exists, so you recommend a custom connector.
- You recommend a developer environment to build, a sandbox to test and a production environment to go live.
Common mistakes
- Accepting the Plan designer output without review. It is a draft; check tables against existing standard tables.
- Jumping to code. A plug-in or PCF control is not the answer if a business rule or a built-in control works.
- Building a new AI feature when a built-in one, such as row summaries or Copilot chat, already meets the need.
- Choosing a trial environment for anything that must last beyond a month.
- Forgetting that AI features need admin enablement in the environment.
How the exam asks about it
- "Describe the process in natural language and get a draft data model and components" points to Plan designer.
- "Reset", "copy" or "test without affecting production" points to a sandbox. "Single maker learning" points to a developer environment.
- "No connector exists for the API" points to a custom connector. "Custom visual control on a form" points to a PCF component.
- "Review what an agent did" or "human in the loop inside the app" points to the agent feed.