AB-620 · Associate

AB-620 Cheat Sheet: Task-to-Feature Quick Reference

Organized by what you're trying to do, not by domain — find the Copilot Studio feature fast, then check the pattern table for how the exam phrases it.

Exam Snapshot

Certification: AI Agent Builder Associate
Passing score: 700
Price: $165 USD
Delivery: Pearson VUE
Domain 1: Plan & configure (30–35%)
Domain 2: Integrate & extend (40–45%)
Domain 3: Test & manage (20–25%)
Objectives checked: 2026-09-07

Domain weights, the passing score, and every referenced feature trace back to Microsoft's official AB-620 study guide and Microsoft Copilot Studio documentation. Microsoft does not publish an exact question count or duration for AB-620 as of this writing.

How to Use This Page

Instead of grouping by exam domain, the table below groups by the thing you're trying to accomplish — because that's how the exam actually poses most questions: a scenario describing a goal, then four options for how to achieve it. Scan the left column for your scenario, then check the right column for the trap.

Task Ledger

You want to…Configure thisWatch out for
Bring in ServiceNow, SAP, or SharePoint content as broad searchable knowledgeCopilot connector (Enterprise knowledge sources)Indexes content into Microsoft Graph — best for broad grounding, not live transactional data
Read or write live, up-to-the-second data from a line-of-business systemMicrosoft Power Platform connectorCalls the API in real time — no data replication, ideal for current facts and transactions
Ground answers in your own custom vector/semantic search indexAzure AI Search knowledge sourceRequires a Foundry-connected Azure AI Search resource, not a Copilot connector
Let the agent call tools from an existing MCP serverTools → Add a tool → Model Context ProtocolCopilot Studio negotiates the handshake and lists the server’s tools; you can disable individual tools per agent
Automate a legacy app or website with no APIComputer use tool (Computer-Using Agent model)Perceives the screen, reasons, then drives a virtual mouse/keyboard — needs generative orchestration turned on
Call your own internal REST endpoint directlyAdd REST API as a toolDifferent path from a custom connector — use when you don’t need the connector wrapper
Delegate a task to an agent built on another framework or hosted elsewhereA2A (Agent2Agent) protocol connectionSupports multiturn interactions and rich metadata — a plain HTTP connector does not
Connect to a Microsoft Foundry agentFoundry agent integrationFoundry agents expose only Responses + A2A by default — the Activity protocol endpoint must be enabled via REST API/Python SDK first, or the connection 400s
Pull structured analytics from a Fabric workspaceFabric data agent integrationA distinct integration path from Foundry or A2A agents
Let a person approve or supply info mid-automationHuman-in-the-loop agent flowDifferent from a standard agent flow — pauses for input instead of running straight through
Show buttons, images, or structured input in a responseAdaptive card node inside a topicConfigured at the topic level, not the agent flow level
Answer from a narrow, specific, bounded sourceGenerative answers node with a scoped knowledge sourceDifferent from a broad custom prompt — the exam tests this distinction directly
Measure whether the agent gives accurate, grounded answersEvaluation method + test setMatch the method to what you’re measuring — accuracy, tool-call correctness, or task completion aren’t the same check
Watch for latency, failures, or usage trends over timeApplication Insights monitoringOperational telemetry — not the same tool as answer-quality evaluation
Move an agent from dev to test to production without hardcoding valuesSolution + environment variables + Power Platform PipelinesStandard Power Platform ALM applied to agents

Knowledge-Source Decision Table

Domain 2's single most-tested distinction — three ways to ground an agent, one job each:

SourceMental modelBest for
Copilot connectorSearch & index — copies content into Microsoft GraphBroad, searchable knowledge with citations across Microsoft 365
Power Platform connectorLive API bridge — no data replicationUp-to-date facts, transactions, read/write actions
Azure AI Search + FoundryYour own managed semantic indexCustom-built retrieval you control end to end

If the Question Says X, Think Y

Exam cueThink
The question says “ServiceNow,” “SAP,” or “index a large document set”Think Copilot connector, not Power Platform connector
The question says “real-time,” “current inventory,” or “write data back”Think Power Platform connector, not Copilot connector
The question says “no API exists” or “legacy desktop app”Think computer use, not MCP or REST API
The question says “existing MCP server” or “tools exposed by a server”Think MCP tools, not a custom connector
The question says “agent built on another framework” or “hosted outside Copilot Studio”Think A2A protocol, not a plain HTTP connector
The question describes a Foundry connection that fails with a 400 error despite correct credentialsThink missing Activity protocol endpoint on the Foundry agent
The question says “approval before continuing” or “needs more info from a person”Think human-in-the-loop agent flow, not a standard topic
The question says “buttons,” “structured input,” or “rich response”Think adaptive card node
The question says “move to production” or “different environment values”Think solution + environment variables, not manual reconfiguration
The question says “track latency or failures over time”Think Application Insights, not agent evaluation

Quick Terms

TermMeaning
MCPModel Context Protocol — a standard for exposing tools/resources to an agent from a connected server
A2AAgent2Agent protocol — an open standard for one agent to delegate tasks to another across frameworks
RAGRetrieval-augmented generation — grounding a model’s answer in retrieved external content
CUAComputer-Using Agent — the model powering the computer use tool’s screen perception and control
ALMApplication lifecycle management — solutions, environment variables, and pipelines for agents
ISVIndependent software vendor — one of AB-620’s named candidate audience roles
FoundryMicrosoft Foundry — Azure’s AI platform for models, agents, and the model catalog
FabricMicrosoft Fabric — the data platform whose data agents can be integrated into a Copilot Studio agent

Quick Answers

How is this cheat sheet different from the AB-620 study guide?

The study guide explains each domain's reasoning and common traps in narrative form. This page is organized by task — what you're trying to do — for a fast lookup during final review.

What's the single most-confused topic?

Copilot connectors versus Power Platform connectors as knowledge sources — see the decision table above.

Does this cheat sheet cover everything the exam tests?

No single page replaces hands-on Copilot Studio experience, which Microsoft's audience profile explicitly expects. Use this as a review aid alongside actually building an agent and connecting an MCP server yourself.

Can I print this AB-620 cheat sheet?

Yes — every table here is plain HTML, so a browser print or "print to PDF" renders cleanly.

MSCertQuiz sells practice-exam access for AB-620 and other Microsoft certifications; this cheat sheet is written by the same team that builds those questions.

Continue Your Prep

Glossary down. Now spot the pattern live.

Start with 40 free AB-620 questions covering every task above.

Start Free AB-620 Practice