AB-620 is Microsoft's new Associate-level exam for Microsoft Certified: AI Agent Builder Associate. It tests whether you can plan, build, integrate, and manage production-grade agents in Microsoft Copilot Studio — MCP tool connections, Agent2Agent (A2A) multi-agent collaboration, Foundry and Fabric integration, and computer-using agents — not whether you can click through basic topic authoring. This guide walks each of the three domains in the depth Microsoft's own skills-measured list describes.
TL;DR
- • AB-620 has 3 domains: planning agent solutions (30–35%), integrating/extending agents in Copilot Studio (40–45%, the largest), and testing/managing agents (20–25%)
- • Passing score is 700, delivered through Pearson VUE like other Microsoft role-based exams
- • The exam entered beta in late June 2026 — confirm current beta/GA status before booking
- • The hardest material is the newest: MCP tools, A2A multi-agent collaboration, Foundry/Fabric agent integration, and computer use
- • No enforced certification prerequisite, but Microsoft expects real hands-on Copilot Studio experience, not just reading
Exam Snapshot
| Detail | Information |
|---|---|
| Exam Code | AB-620 |
| Exam Title | Designing and Building Integrated AI Agent Solutions in Copilot Studio |
| Certification | Microsoft Certified: AI Agent Builder Associate |
| Level | Associate (Intermediate) |
| Passing Score | 700 (Microsoft’s scaled 1000-point system) |
| Price | $165 USD (per Microsoft’s AB-620 sweepstakes voucher terms) |
| Delivery | Pearson VUE — online proctored or test center |
| Prerequisite Certification | None enforced |
| Renewal | Annual, free online assessment via Microsoft Learn |
Sourced from Microsoft's AI Agent Builder Associate certification page and official AB-620 study guide, checked 2026-09-07. Microsoft does not publish an exact question count or exam duration for AB-620 as of this writing — that is normal for a recently launched exam and worth re-checking before you register.
Plan and Configure Agent Solutions (30–35%)
This domain front-loads decisions you make before you open the Copilot Studio canvas, then moves into building the agent flows and topics that carry those decisions out.
Planning an agent solution
Microsoft's objectives here cover planning integration with enterprise systems, identity strategy, channel and deployment planning, responsible AI strategy, security and governance evaluation, reusable agent components, and designing for internal versus external audiences. The exam treats these as judgment calls, not vocabulary recall — expect a scenario describing a constraint (a regulated industry, an external customer-facing channel, a need to reuse a component across agents) and asking which planning decision follows. Common trap:conflating "responsible AI strategy" questions with generic content-moderation settings; Microsoft's framing is broader, including transparency, fairness, and governance review before an agent ships, not just a profanity filter.
Creating and monitoring agent flows
Agent flows are the automation layer — multistep, tool-calling sequences with their own monitoring, input/output parameters, and error handling, distinct from a conversational topic. Objectives include creating an agent flow, creating a human-in-the-loop flow (where a person approves or supplies information mid-flow), configuring actions and connectors, monitoring flows, adding input/output parameters, and implementing error handling. Common trap:candidates who have only built topics assume a topic and an agent flow are interchangeable; the exam expects you to know when a task needs a flow's structured, monitorable execution instead of a topic's conversational branching.
Configuring topics
This is the densest sub-area in the domain: adding agent flows to a topic, configuring response formatting, adding tools to a topic, configuring advanced responses with custom prompts, custom knowledge sources, and API/Send HTTP requests, configuring the generative answers node, configuring adaptive cards, and managing variables. Common trap:the generative answers node and a custom-knowledge-source-backed prompt look similar in a scenario description — the exam differentiates them by whether the answer needs to be grounded in a specific, bounded source (generative answers) versus a broader, instruction-driven response (custom prompt).
Integrate and Extend Agents in Copilot Studio (40–45%)
The largest domain, and the one that separates AB-620 from a basic Copilot Studio credential. It is built almost entirely from capabilities that did not exist, or were niche, a year ago.
Connecting to enterprise knowledge sources
Three knowledge-source options are explicitly named: Copilot connectors, Microsoft Power Platform connectors, and Azure AI Search. Microsoft's own documentation is explicit about the distinction: Copilot connectors (formerly Microsoft Graph connectors) index external content into Microsoft Graph for broad, searchable, semantically grounded knowledge; Power Platform connectors call APIs in real time for up-to-date facts, transactions, and read/write actions without replicating data. Common trap:a scenario needing live, transactional data (e.g. current inventory counts) gets answered with a Copilot connector distractor, or a scenario needing broad indexed knowledge across many documents gets answered with a real-time connector distractor — know which direction each one solves.
Adding tools to agents
Four tool types: computer use, MCP tools, a custom-connector-based tool, and REST APIs added directly. Computer use is powered by a Computer-Using Agent (CUA) model that perceives the screen, reasons about it, and drives a virtual mouse/keyboard — Microsoft's own framing is that it exists specifically for systems "where APIs aren't available." MCP tools work differently: you connect to an existing MCP server from the agent's Tools page, and Copilot Studio negotiates the protocol handshake and lists out the tools and resources that server exposes, which you can then selectively enable per agent. Common trap:a scenario describing a legacy system with no API is a computer-use answer, not an MCP or REST API answer — MCP and REST APIs both assume a programmatic interface already exists.
Configuring multi-agent collaboration
Five sub-objectives: designing multi-agent solutions, integrating a Foundry agent, integrating an existing Copilot Studio agent, integrating a Fabric data agent, and creating a multi-agent solution using the A2A protocol. Microsoft's own comparison is direct: A2A is designed for agent workflows, supports multiturn interactions, carries rich contextual metadata, and works across frameworks — a plain HTTP connector supports none of those by default. Common trap, and a real one:a newly created Foundry agent exposes only its Responses and A2A protocol endpoints by default; connecting it to Copilot Studio requires the Activity protocol endpoint to be enabled first (currently only possible via REST API or the Python SDK, not the Foundry portal UI), or the connection fails at runtime with a 400 error. A question describing a Foundry-agent connection that "isn't working" despite correct credentials is very plausibly testing this exact gap.
Integrating agents with Azure
Generative answers grounded through Azure AI Search combined with Foundry, custom prompts against the Foundry model catalog, and monitoring agents with Application Insights. Common trap:Application Insights is for operational telemetry (latency, failures, usage over time) — don't confuse it with the in-product evaluation tooling covered in the next domain, which measures answer quality rather than system health.
Test and Manage Agents (20–25%)
The smallest domain by weight, but it's where hands-on agent builders who have never operated agents at team scale tend to lose points.
Evaluating agent performance
Creating a test set, choosing an evaluation method, and reviewing test results. Common trap:the exam expects you to match an evaluation method to what's actually being measured (factual accuracy of a grounded answer versus whether a tool was invoked correctly versus overall task completion) rather than treating "evaluation" as one undifferentiated pass/fail step.
Application lifecycle management (ALM) for agents
Creating a solution, adding existing agents to a solution, creating and using environment variables, and implementing and extending Microsoft Power Platform Pipelines. This is standard Power Platform ALM applied specifically to agents. Common trap:candidates who have only built agents in a single default environment underestimate how much of this sub-area is really "how do I move this agent from dev to test to production without hardcoding values," which is exactly what environment variables and a solution-based pipeline solve.
Where to Put Your Study Hours
This is MSCertQuiz's own recommended allocation, not a Microsoft-published figure — built by weighting a roughly 30-hour study budget against each domain's exam share, tilted further toward domains where hands-on build time matters more than reading time.
| Domain | Exam Weight | Suggested Hours | Where to spend it |
|---|---|---|---|
| Plan and configure agent solutions | 30–35% | 9–11 hrs | Planning artifacts (identity, channels, responsible AI, governance) plus hands-on time building agent flows and topics |
| Integrate and extend agents in Copilot Studio | 40–45% | 12–14 hrs | MCP tool setup, one real A2A connection, one Foundry agent connection, one Fabric data agent — build each once rather than only reading about it |
| Test and manage agents | 20–25% | 6–8 hrs | Building a test set, running an evaluation, and moving one agent through a solution and a Power Platform Pipeline |
Candidates already building and integrating agents in Copilot Studio day-to-day can compress this toward 3–4 weeks of part-time study; candidates newer to the platform should plan closer to 6–8 weeks, with extra time set aside to actually connect an MCP server and an A2A agent once, rather than only reading Microsoft's documentation about them.
Frequently Asked Questions
What is the AB-620 exam?
AB-620, officially Designing and Building Integrated AI Agent Solutions in Copilot Studio, is the exam for Microsoft Certified: AI Agent Builder Associate. It tests whether a developer or advanced maker can plan, build, integrate, and manage enterprise-grade agents in Microsoft Copilot Studio — not just author a basic topic-based chatbot.
What is the passing score for AB-620?
Microsoft's AB-620 study guide states a score of 700 or greater (on Microsoft's scaled 1000-point system) is required to pass. That is the same threshold used across nearly all Microsoft role-based certification exams.
How much does the AB-620 exam cost?
Microsoft's own AI Agent Skills Sweepstakes rules page lists a $165 USD retail value specifically for the AB-620 exam voucher, matching the standard Associate/Expert exam price Microsoft publishes generally. Microsoft's AB-620-specific exam listing does not itself publish an exact duration or question count as of this writing, so confirm both figures on the official exam page before you book.
Are there prerequisites for AB-620?
No certification prerequisite is enforced at registration. Microsoft's audience profile does, however, expect intermediate knowledge of generative AI concepts (models, orchestration, RAG, MCP, A2A protocol), Power Fx, Microsoft Dataverse, Power Platform environments, prompt engineering, REST APIs, and hands-on experience configuring knowledge sources, instructions, tools, and topics in Copilot Studio.
How is AB-620 different from just knowing basic Copilot Studio topics?
The largest domain (40–45%) is entirely about integration and extension — MCP tools, computer use, custom connectors, REST APIs, multi-agent collaboration with Foundry/Fabric/A2A, and Azure AI Search grounding. A candidate who has only built simple trigger-phrase topics without touching any of these surfaces will find well over a third of the exam unfamiliar.
Is AB-620 a beta exam right now?
AB-620 entered beta in late June 2026, per Microsoft's own AI Agent Skills Sweepstakes rules and community moderator guidance. Microsoft's standard beta-to-general-availability window runs roughly 10–14 weeks, so treat any exam-day detail Microsoft hasn't published on its official AB-620 pages as still settling, and check the current status before you register.
Does the AB-620 certification expire?
Yes. Like other Microsoft Associate and Expert certifications, AB-620 expires annually. Renewal is a free online assessment through your Microsoft Learn profile — no need to retake the full proctored exam.
MSCertQuiz sells practice-exam access for AB-620 and other Microsoft certifications; this guide is written by the same team that builds and verifies those questions against live Microsoft Learn documentation.
Related Resources
Task-to-feature reference tables for a final review pass.
12 scenario questions across Easy, Medium, and Hard tiers.
Full exam details and the 500-question practice bank.
Browse every Microsoft certification MSCertQuiz covers.
Ready to Practice AB-620?
500 scenario-based questions across all 3 domains — MCP tools, A2A protocol, Foundry and Fabric integration, and full ALM coverage. Start with 40 questions free.
Start Free AB-620 Practice →