Table of Contents
🌏 中文版
This is a preparation path built from official material, not an exam-day account — I have not sat this exam. Every "what it tests" points back to the official study guide, and every "how to prepare" points to official Microsoft training. No leaked questions. Verified 2026-08-18.
Microsoft now has three agent-related certifications, and AB-620 is the only one that does not expect you to write Python. Its ground is Copilot Studio: agent flows, adaptive cards, computer use, MCP tools, the A2A protocol, and Fabric data agents.
For prices, validity, and gates across vendors, see What AI certifications engineers can take in 2026 — not repeated here.
Who This Is For
The certification page describes the candidate as a "professional developer or advanced builder" and lists what you should already know: Power Fx, Microsoft Dataverse, Power Platform environments and components, Microsoft 365 Copilot, Microsoft Foundry, and adaptive cards; intermediate generative AI concepts including models, orchestration, RAG, MCP, and the A2A protocol; plus prompt engineering and REST API integration experience.
The same page states plainly what you do in the role: integrate agents with Foundry, with MCP servers, with custom connectors, with APIs, with Microsoft Fabric, and automate tasks using computer use.
A fit for developers, consultants, and ISV partners building enterprise agents in Copilot Studio. One practical advantage: it is offered in Traditional Chinese, the only localized exam of Microsoft's three agent credentials — AI-500 and AB-100 are English-only.
Not a fit if you want to prove code-first ability. That track is AI-103 and AI-500 above it, which test Python, Agent Framework, LangGraph, and CI/CD — barely overlapping with this exam.
Official Specs at a Glance
| Item | Detail |
|---|---|
| Exam code | AB-620 |
| Certification | AI Agent Builder Associate |
| Status | Generally available (no beta label on the page) |
| Fee | $165 USD (priced by country or region) |
| Length | 120 minutes |
| Question count | Not published |
| Question types | Not disclosed in advance; the page says "You may have interactive components to complete as part of this exam" |
| Passing score | 700 (this figure appears in the study guide, not on the certification page) |
| Validity | 1 year (free online renewal assessment) |
| Languages | 13, including Traditional Chinese |
| Prerequisites | None |
| Practice assessment | Not yet available — "usually available within 8 weeks of the exam being out of beta and generally available" |
Plan around the missing practice assessment: there is no official mock exam for this one yet, so the learning-path labs and the exam sandbox are all you have for familiarizing yourself with the interface.
The Three Weighted Skill Areas
| Skill area | Weight |
|---|---|
| Plan and configure agent solutions | 30–35% |
| Integrate and extend agents in Copilot Studio | 40–45% |
| Test and manage agents | 20–25% |
Integration alone is nearly half the exam, which sets the strategy: rather than reading about planning, connect each integration type once by hand.
Area by Area
Plan and Configure Agent Solutions (30–35%)
What it tests: planning enterprise system integration, identity strategy, channels and deployment, and responsible AI strategy; evaluating security and governance considerations; planning reusable agent components; designing agents for internal versus external audiences.
The agent-flow half is concrete: create an agent flow, create a human-in-the-loop agent flow, configure actions and connectors, monitor agent flows, add input and output parameters, and implement error handling in agent flows.
The topics half: add agent flows to a topic, configure response formatting, add tools to a topic, configure advanced responses with custom prompts, custom knowledge sources, and APIs and Send HTTP requests, configure the generative answers node, configure adaptive cards, and manage variables.
How to prepare: this is the hands-on surface of Copilot Studio, and building it once through the official learning path covers most of it. Focus on human-in-the-loop and error handling — the line between a demo and something that survives production, and both are called out explicitly.
Integrate and Extend Agents in Copilot Studio (40–45%, the heaviest)
What it tests, in four directions:
Knowledge sources — Copilot connectors, Power Platform connectors, Azure AI Search.
Tools — configuring and monitoring computer use for an agent, configuring MCP tools, adding a tool via an existing custom connector, adding REST APIs to an agent.
Multi-agent collaboration — designing multi-agent solutions in Copilot Studio, integrating a Foundry agent, integrating an existing agent, integrating a Fabric data agent, and building a multi-agent solution using the A2A protocol.
Azure integration — configuring generative answers with Azure AI Search plus Foundry, pointing custom prompts at the Foundry model catalog, and monitoring agents with Application Insights.
How to prepare: no shortcut — wire up each of the four once. MCP tools and A2A are the newest objectives here; the harness layer of agent security on this site covers the protocol layer in practical terms. For computer use, run it once and then open the monitoring view — Microsoft writes "configure and monitor computer use" as a single skill.
Test and Manage Agents (20–25%)
What it tests: creating a test set, choosing an evaluation method, reviewing test results; and ALM — creating a solution, adding existing agents to a solution, creating and using environment variables, and implementing and extending Power Platform Pipelines.
How to prepare: this is classic Power Platform ALM and has little to do with agents as such. With a Power Platform background it needs almost no preparation; without one, budget real time, because solutions and pipelines are concepts from outside Copilot Studio.
A Five-Week Schedule and Its Derivation
Derivation: the three official learning paths total 8 hours 29 minutes of stated content, but that is content time, excluding hands-on work and absorption. Allocating by weight and doubling the heaviest area gives five weeks.
At 5–7 hours a week over five weeks:
| Week | Content | Reasoning |
|---|---|---|
| 1 | Read the study guide, run the exam sandbox, and take Design agent conversations and responses using topics (2h 17m, 3 modules) | With no practice assessment, the sandbox is the only way to meet the interface early |
| 2 | The rest of Plan and configure: agent flows and human-in-the-loop | 30–35% of the exam, and the base the integration work sits on |
| 3–4 | Integrate and extend (40–45%): multi-agent solutions (2h 54m) and enterprise system integration (3h 18m) | The heaviest area, split into "multi-agent" and "external integration" |
| 5 | Test and manage (20–25%) + full review | ALM concepts close it out |
The instructor-led course is not out yet: AB-620T00-A (three days) carries the notice "This course will be available on 9/18/2026" — but the three learning paths above are self-serve today, so there is nothing to wait for.
Failure cost: Microsoft's retake policy is 24 hours after a first failure, then 14 days between subsequent attempts, with at most five attempts per 12 months, paying each time.
Where It Sits Among Microsoft's Three Agent Credentials
| Certification | Position | Heaviest area | Languages |
|---|---|---|---|
| AB-620 (this article) | Low-code builder, Copilot Studio | Integration 40–45% | 13, incl. Traditional Chinese |
| AI-500 (beta) | Code-first engineer: Python, Agent Framework, LangGraph | Development 30–35% | English only |
| AB-100 | Architect: ROI and cross-product governance | Deploy and govern 40–45% | English only |
Exactly one official ladder exists: AI-103 → AI-500, where the latter's Certification prerequisites section requires the former. AB-620 has no prerequisite and can be taken directly.
AB-620 does appear in AB-100's list of "current possible associate certs that can be used for this expert certification," which suggests an intended Copilot-Studio-side path of AB-620 → AB-100 — but official text stops short of making it a requirement, so don't treat it as one.
One-Year Validity and Renewal
The same as AI-103: one-year validity, free renewal, taken online, unproctored and open book, available only in the six months before expiry, with a lapse forcing a full retake. The full rules are in the last section of the AI-103 article.
One small oddity worth knowing: AB-620's dedicated renewal page does not exist yet (/ai-agent-builder-associate/renew/ returns 404) — only the general renewal policy page. That normally means the first cohort has not reached its renewal window, not that there is no renewal path.
Things That Will Go Stale (Check These Next Time)
| Item | Status as of 2026-08-18 | When to re-check |
|---|---|---|
| The three weights | 30-35 / 40-45 / 20-25 | Quarterly |
| Study guide version | No "Skills measured as of" date, no change log | Quarterly |
| Practice assessment | Not yet available | Monthly |
| Dedicated renewal page | Not live yet (404) | Quarterly |
| Languages | 13, including Traditional Chinese | Every six months |
References
- AI Agent Builder Associate certification page
- AB-620 official study guide
- Course AB-620T00-A
- Microsoft certification renewal
- Microsoft exam retake policy
- Exam duration and experience
Related on this site
Loading...