ExamSprintLab

AB-620 prerequisites: what you need first

Updated September 20, 2026

There are no formal prerequisites for AB-620. You can book it today with no certifications and nothing will stop you. But Microsoft’s audience profile for this exam is unusually long and unusually specific, and ignoring it is the main reason people fail.

Most exams assume a role. This one assumes a toolkit.

What Microsoft actually assumes

You should be familiar with:

  • Power Fx
  • Microsoft Dataverse
  • Microsoft Power Platform environments and components
  • Microsoft 365 Copilot
  • Microsoft Foundry
  • Adaptive cards

You need intermediate knowledge of generative AI concepts, including:

  • Models and orchestration
  • Retrieval-augmented generation (RAG)
  • Model Context Protocol (MCP)
  • Agent2Agent (A2A) protocol
  • Prompt engineering
  • REST APIs and integration patterns

And you need hands-on experience configuring agents in Copilot Studio with basic knowledge sources, instructions, tools and topics.

That last line is the real gate. AB-620 begins at the point where a beginner’s Copilot Studio course ends. If you have never built a working agent with a knowledge source and a tool, you are not ready to start studying — you are ready to start building.

What each item actually means in practice

Power Fx. You need to read and write expressions, not master the language. Variables, conditions, string handling, referencing values from earlier steps.

Dataverse. Understand tables, columns, relationships and that it is the Power Platform’s data store. You are not being examined as a data modeller.

Power Platform environments. Know what an environment is, why development, test and production are separate, and that solutions move between them. This underpins the whole ALM part of the exam.

Microsoft 365 Copilot. Know how agents surface there and what that implies for audience and permissions.

Microsoft Foundry. Enough to understand what a Foundry agent and the model catalog are, because a whole section covers integrating them.

Adaptive cards. Know what they are for — structured, interactive responses — and roughly how they are defined.

MCP and A2A. These are newer than most study material. MCP is a standard way to expose tools to models. A2A is a standard for agents to talk to other agents. Both appear in the outline as things you configure, so read the current documentation rather than a course recorded last year.

What you do not need

  • No prior certification. Not PL-200, not PL-400, not AI-103.
  • No deep coding. This is the low-code agent exam. Power Fx expressions and API concepts, not C# or Python.
  • No Azure administration. You integrate with Azure services; you do not administer them.
  • No data science. Nothing about training or evaluating models statistically.

If you are missing something

Rank the gaps by how much of the exam they block:

  1. No hands-on Copilot Studio agent experience. Fix this first — nothing else matters until you have built one. A day of building beats a week of reading.
  2. No Power Platform environment or solution experience. This blocks the whole test-and-manage domain, 20–25% of the exam, and it is quick to close.
  3. No MCP or A2A familiarity. Blocks part of the largest domain. Read the current docs and configure each once.
  4. Weak on Power Fx or adaptive cards. Narrower, and learnable alongside the rest.

What to do before you book

  1. Build an agent in Copilot Studio with a knowledge source, a tool and a topic.
  2. Put it in a solution and move it to a second environment.
  3. Configure one MCP tool and one connector-based tool.
  4. Try the free sample questions.
  5. If the integration questions felt like guesswork, work through the study plan before booking.

You will need a Copilot Studio environment you can build in. A developer or trial Power Platform environment is enough for everything on this exam — you do not need production access to anything.