Free AB-620 Cheat Sheet 2026: Exam Quick Reference

Plan and configure agent solutions

30-35%of exam

Agent Building Blocks Orchestration Modes Topic Triggers Topic Canvas Nodes Variables and Entities Agent Flows Authentication Options Channels and Publishing

Integrate and extend agents in Copilot Studio

40-45%of exam

Knowledge Sources Knowledge Limits Tool Types MCP and Computer Use Multi-Agent Collaboration Azure Integration

Test and manage agents

20-25%of exam

Evaluation Test Methods Monitor Metrics ALM and Solutions Data Policy Connectors Copilot Credits

Quick Facts

Exam: AB-620
Credential: AI Agent Builder Associate
Time: 120 min
Pass: 700/1000 scaled
Level: Intermediate associate
Delivery: Pearson VUE proctored
Price: Set by country/region
Renewal: Free annual assessment
Retake: 24 hours, then 14 days
Blueprint: April 9, 2026
Heaviest area: Integrate and extend 40-45%

Domain Weight Order

PIT: Plan 30, Integrate 40, Test 20

Plan 30-35%Integrate 40-45%Test 20-25%Integrate is heaviest

Generative vs Classic Orchestration

Generative

Model picks topics and tools
Descriptions drive selection
Generates questions and replies
25 knowledge sources searched

Classic

Trigger phrases pick one topic
Tools called only from topics
You author every message
Four websites, four SharePoint URLs

Descriptions select; phrases match

Pick an Authentication Option

Anonymous public website agent→No authentication(No user credential tools)
Teams and Microsoft 365 only→Authenticate with Microsoft(Zero configuration Entra)
Custom website needs sign-in→Authenticate manually(Entra ID or OAuth)
Topic needs an access token→Authenticate manually(User.AccessToken variable)
Control who can chat→Require users to sign in(Entra ID sharing)
Certificate based identity→Entra ID V2 certificates(Key Vault URL)

Agent Building Blocks

Instructions: Agent-level natural language guidance
Knowledge: Grounding data for answers
Tools: Actions the agent calls
Topics: Authored conversation paths
Triggers: What starts a topic
Child agent: Lightweight subagent inside agent
Connected agent: Separate reusable published agent
Agent flow: Deterministic rule-based automation
Solution: Dataverse container for ALM

What Generative Orchestration Picks

KATT: Knowledge, Agents, Tools, Topics

Knowledge sourcesChild and connected agentsToolsTopicsDescriptions decide

Topic vs Tool

Topic

Authored conversation path
Nodes on a canvas
Always has a trigger

Tool

Calls an external system
Inputs filled by AI
Reusable across topics

Topic talks; tool acts

Pick a Deployment Channel

Internal employees in Teams→Teams and Microsoft 365(Microsoft authentication)
Stakeholder feedback round→Demo website(Not for production)
Public customer website→Custom website(Direct Line channel)
Native mobile application→Mobile app(Direct Line API)
Contact centre handoff→Omnichannel for Customer Service(Dynamics 365)
Consumer messaging→WhatsApp(Separate channel connector)
Slack, Telegram, email→Azure Bot Service channels(Configured in Azure)

Orchestration Modes

Generative: Model picks topics, tools, knowledgeDefault
Classic: Trigger phrases pick one topic
Descriptions: Drive generative selection accuracy
Multi-intent: Generative chains several steps
Plan: Ordered steps for one turn
Activity map: Test view of steps
Admin off switch: Environment can force classic
Conversational boosting: Skipped under generative orchestration
Prebuilt agent: Its config sets the mode

Three Authentication Choices

NMM: None, Microsoft, Manual

No authenticationAuthenticate with MicrosoftAuthenticate manually

Microsoft vs Manual Authentication

Authenticate with Microsoft

Zero configuration Entra ID
Teams and Microsoft 365
Only User.Id, User.DisplayName

Authenticate manually

Any channel supported
Entra ID or generic OAuth
Adds User.AccessToken, User.IsLoggedIn

Need a token, go manual

Topic Triggers

The agent chooses: Generative default, description match
User says a phrase: Classic default, trigger phrases
A message is received: Any user message activity
A custom client event occurs: Event activity, filter by name
An activity occurs: Any activity type received
The conversation changes: Conversation update, user joins
It's invoked: Teams invoke activity received
It's redirected to: Called explicitly by another topic
User inactive for a while: Fires after configured inactivity
A plan completes: All planned steps finished
AI response generated: Fires before orchestration reply sent
Priority: Overrides default execution order

Agent Flow vs Cloud Flow

Agent flow

Built in Copilot Studio
Consumes Copilot Credits
13 credits per 100 actions

Power Automate cloud flow

Power Automate licensing
Not subject to enforcement
Conversion is one way

Credits versus Power Automate licensing

Topic Canvas Nodes

Message: Send text or card
Question: Ask and store answer
Condition: Branch on a value
Set variable value: Assign literal, variable, formula
Parse value: Convert JSON to record
Clear variable values: Can clear conversation history
Redirect: Call another topic or agent
End all topics: Cancels remaining planned steps
Generative answers: Advanced node searching knowledge
Send HTTP request: Direct REST call from topic
Adaptive card: Rich card inside message
Add a tool: Call tool explicitly from topic

Variables and Entities

Topic variable: Scoped to one topic
Global variable: Shared across the session
System variable: Platform supplied, Activity.Text
Environment variable: Per-environment setting for ALM
Key Vault secret: Environment variable pointing at vault
Prebuilt entity: Person name, City, Date
Custom entity: Closed list or regex
Receive values: Accept input from other topics
Return values: Pass output to caller
Unknown type: Variable set before value assigned
Custom entity gap: Not tool or topic inputs

Agent Flows

Deterministic: Same input, same output
When an agent calls the flow: Trigger exposing flow as tool
Human in the loop: Approval or information request
Built-in tools: Loops, branches, data operations
AI capabilities: Prompts, agent calls, replies
Connectors: Microsoft, third-party, custom
Input and output parameters: Passed to and from agent
Convert cloud flow: One-way switch, cannot reverse
Test runs: Do not consume capacity

Authentication Options

No authentication: Anyone with the link
Authenticate with Microsoft: Entra ID, Teams focused
Authenticate manually: Any channel, full token
Entra ID V2: Secret, certificate or federated
Generic OAuth 2: Any compliant identity provider
Require users to sign in: Manual authentication only setting
User.Id: Available in both Microsoft options
User.AccessToken: Manual authentication only
Publish to apply: Auth changes need a publish
Tool credentials: End user or maker-provided

Channels and Publishing

Teams and Microsoft 365 Copilot: Default with Microsoft authentication
SharePoint: Publish inside SharePoint sites
WhatsApp: Messaging channel, separate connector
Demo website: Stakeholder testing, not production
Custom website: Embed on your site
Mobile app: Custom application via Direct Line
Facebook: Messenger channel connector
Azure Bot Service channels: Slack, Telegram, Twilio, Email
Publish first: Channels appear after first publish
Session timeout: New content after 30 minutes

Child vs Connected Agent

Child agent

Lives inside the main agent
Shares settings and publishing
Delete removes it

Connected agent

Separate agent, own settings
Own channels and ALM
Disconnect leaves it intact

Reuse and publish means connected

Pick a Knowledge Source

Public web pages you own→Public website(Bing indexed)
Files stored in SharePoint→SharePoint(Entra user auth)
Ad hoc reference documents→Documents(Uploaded to Dataverse)
Records in Dataverse tables→Dataverse(Up to 15 tables)
ServiceNow, SAP, Confluence→Copilot connectors(Indexed in Graph)
Live data, no index→Power Platform connectors(Real-time knowledge)
Your own vector index→Azure AI Search(One index only)
Open internet answers→Web Search(Generative mode only)

Knowledge Sources

Public website: Bing results limited to sites
Documents: Files uploaded to Dataverse
SharePoint: Graph search over sites
Dataverse: Retrieval over environment tables
Copilot connectors: External data indexed in Graph
Power Platform connectors: Real-time connector knowledge
Azure AI Search: Your own vector index
Web Search: Grounding with Bing Search
Official source: Trusted, classic orchestration only
Allow ungrounded responses: Permits model general knowledge

Copilot vs Power Platform Connectors

Copilot connectors

Index data into Microsoft Graph
Permission trimmed grounding
Admin configures the connector

Power Platform connectors

Call APIs as tools
Also real-time knowledge
Maker adds a connection

Indexed knowledge versus live calls

Pick an Extension Method

Call one REST endpoint→REST API tool(Endpoints become tools)
Reuse an existing connector→Custom connector(Share with organization)
Server exposes many tools→Model Context Protocol(Tools refresh automatically)
Deterministic multi-step process→Agent flow(Same input, same output)
No API exists→Computer use(Screen automation)
Single model instruction→Prompt tool(One turn only)
External agent framework→A2A agent(Endpoint plus agent card)
Group skills inside agent→Child agent(No separate publishing)

Knowledge Limits

Websites generative: 25 websites
Websites classic: Four public URLs
SharePoint generative: 25 URLs
SharePoint classic: Four URLs per node
Dataverse classic: Two sources, 15 tables each
Connectors classic: Two per custom agent
Over 25 sources: Model filters which to search
Uploaded files: Exempt from the 25 limit
Teams citations: At most 20 per response
Semantic search files: 512 MB PDF, PPTX, DOCX

MCP vs A2A

MCP

Server exposes tools, resources
Your agent still reasons
Built on custom connectors

A2A

Delegates a task outward
Remote agent reasons
Multiturn with chat history

Call tools versus delegate tasks

Tool Types

Prompt: Single-turn model call
Agent flow: Deterministic multi-step automation
Computer use: Drives websites and desktop apps
Custom connector: Your own Power Platform connector
Model Context Protocol: Connect to an MCP server
REST API: Endpoints added as tools
Prebuilt connector: Thousands of ready APIs
Skills: Azure Bot Service container
Client tools: Event handled by client
Tool limit: 128 maximum per orchestrator
Recommended tools: 25 to 30 per agent
Ask before running: Confirmation prompt, off by default

MCP and Computer Use

MCP tools: Supported, callable functions
MCP resources: Supported, file-like context
MCP prompts: Not supported today
Dynamic refresh: Server changes appear automatically
MCP prerequisite: Generative orchestration must be on
CUA: Vision model driving the screen
Machine: Power Automate managed target machine
Access control: Allow list of sites, apps
Enforce HTTPS: Blocks plain HTTP sites
Stored credentials: Internal storage or Key Vault
Human supervision: Email reviewer for risky instructions
Credits per step: 5 standard, 15 premium

Multi-Agent Collaboration

Child agent: Lives inside the main agent
Connected agent: Separate agent, own settings
Foundry agent: Preview, Microsoft Foundry connection
Fabric data agent: Preview, no topic redirect
A2A agent: Endpoint URL plus authentication
Agents SDK agent: Preview, Activity protocol
Agent card: Well-known metadata for A2A
A2A authentication: None, API key, OAuth
Split threshold: Above 30 to 40 choices
No chaining: Main agent cannot be connected
Cost of splitting: Extra hops add latency

Azure Integration

Azure AI Search: One vector index per connection
Integrated vectorization: Same model embeds query
Semantic ranker: Optional Azure-side relevance boost
Search auth types: Key, certificate, principal, Entra
metadata_storage_path: Index field used as citation
Foundry model catalog: Model choice for custom prompts
Application Insights: Connection string in Advanced settings
Enable logging: Logs incoming and outgoing messages
Node execution events: Event per topic node run
designMode: Custom dimension marking test chat
Copilot Studio Dashboard: Built-in Application Insights workbook
Virtual network: Private endpoint indexes supported

Copilot Credit Ladder

1 classic, 2 generative, 5 action, 10 graph

Classic answer 1Generative answer 2Agent action 5Tenant graph 10Flow 13 per 100

Environment vs Solution

Environment

Dataverse container and boundary
Development, test, production
Holds capacity and policies

Solution

Package of components
Moves between environments
Managed downstream, unmanaged in dev

Environment hosts; solution travels

Pick an Evaluation Method

No single correct answer→General quality(Default method)
Same idea, different wording→Compare meaning(Set a pass score)
Right tool must fire→Tool use(Pass or fail)
Answer must contain terms→Keyword match(Any or all)
Wording must match closely→Text similarity(Cosine score)
Codes or fixed phrases→Exact match(Character for character)
Your own compliance rule→Custom(Instructions plus labels)

Evaluation Test Methods

General quality: Relevance, groundedness, completeness, abstentionDefault
Compare meaning: Intent similarity to expected answer
Tool use: Expected tools or topics used
Keyword match: Any or all keywords present
Text similarity: Cosine similarity of wording
Exact match: Character for character match
Custom: Your instructions and labels
Pass score: Threshold, default 50
Test set types: Single response or conversation
Invalid result: Missing expected answer or tools

Conversation Outcome Buckets

REAU: Resolved, Escalated, Abandoned, Unengaged

Resolved confirmed or impliedEscalated has three reasonsAbandoned after timeoutUnengaged never started

Test Panel vs Demo Website

Test panel

Authoring-time chat
Activity map available
Excluded from Monitor analytics

Demo website

Needs a publish first
Share with stakeholders only
Not for production traffic

Build in test; validate published

Monitor Metrics

Resolved confirmed: User confirmed successful end
Resolved implied: Timed out, no active plan
Escalated: Escalate topic or transfer node
Abandoned: Engaged session timed out
Unengaged: Session never became engaged
System intended: Business rule escalation, expected
System unintended: Threshold escalation, user stuck
User requested: User asked for human
CSAT: Survey score out of five
Answer rate: Answered versus unanswered questions
Data retention: 360 days, transcripts 28
Test panel: Excluded from Monitor analytics

ALM Promotion Order

SEP: Solution, Environment variable, Pipeline

Work inside a solutionEnvironment variables for settingsPipelines to promoteExport as managed

ALM and Solutions

Solution: Container that moves components
Publisher prefix: Always use a custom publisher
Unmanaged: Development environment only
Managed: Export for test, production
Environment variable: Value set per environment
Connection reference: Environment-specific connection pointer
Power Platform Pipelines: In-product deployment for makers
Azure DevOps: Enterprise CI/CD with build tools
Three environments: Development, test, production
Component collection: Reusable topics, tools, knowledge
Not solution-aware: App Insights, manual auth, channels
Sharing: Reapply after every deployment

Data Policy Connectors

Data groups: Business, Non-business, Blocked
Chat without Entra ID auth: Block to require sign-in
Direct Line channels: Demo, custom site, mobile
Knowledge source connectors: SharePoint, websites, documents
Microsoft Copilot Studio: Blocks event triggers, evaluations
Skills with Copilot Studio: Blocks Bot Service skills
HTTP: Blocks Send HTTP request
Application Insights connector: Blocks agent telemetry export
Endpoint filtering: Allow or deny specific URLs
Analytics Viewer: Monitor page access only
Bot Transcript Viewer: Needed to read transcripts

Copilot Credits

Classic answer: 1 credit
Generative answer: 2 credits
Agent action: 5 credits
Tenant graph grounding: 10 credits
Agent flow actions: 13 per 100 actions
Premium AI tool: 100 credits per 10 responses
Computer use step: 5 standard, 15 premium
M365 Copilot users: Employee scenarios at no charge
Overage enforcement: Agents disabled at 125%
Flow enforcement: Blocks new runs only
Pay-as-you-go: Azure meter avoids enforcement

Common Traps

700 scaled, not 70 percent

700 of 1,000 scaled ≠ Not 70% of items

Generative 25, classic 4

Generative mode: 25 websites ≠ Classic mode: four URLs

Knowledge is not a tool

Knowledge grounds the answer ≠ Tools take an action

Child agent is not connected

Child cannot publish separately ≠ Connected has own ALM

MCP is not A2A

MCP exposes tools, resources ≠ A2A delegates whole tasks

Agent flow is not cloud flow

Agent flow bills Copilot Credits ≠ Cloud flow uses Power Automate

Save is not publish

Auth and channels need publish ≠ Users get last published version

Not everything travels in solutions

App Insights, manual auth excluded ≠ Channels and sharing reapplied

Authenticate with Microsoft limits channels

Teams and Microsoft 365 focused ≠ Other channels need manual auth

Enforcement thresholds differ

Agents disabled above 125% ≠ Agent flows blocked at 100%

Last Minute

1.Plan 30-35, Integrate 40-45, Test 20-25
2.Pass = 700 of 1,000 scaled
3.120 minutes, Pearson VUE, proctored
4.Generative = descriptions; classic = trigger phrases
5.Knowledge grounds; tools act
6.MCP = tools; A2A = agents
7.Child agent inside; connected agent separate
8.Computer use needs generative orchestration
9.Auth changes apply only after publish
10.Credits: 1, 2, 5, 10, 13
11.General quality is default evaluation
12.Outcomes: Resolved, Escalated, Abandoned, Unengaged
13.Environment variables carry per-environment settings
14.Export managed to test, production
15.First retake after 24 hours
16.Certification renews free every year
17.Blueprint dated April 9, 2026

Microsoft AI Agent Builder Associate (AB-620, Copilot Studio)

Practice Questions100 questions Study Guide Cheat Sheet Flashcards50 cards

Microsoft AB-620 Designing and Building Integrated AI Solutions In Copilot Studio - Practice Test Prep Guide

$28.99 · Buy on Amazon

Same family resources

Explore More Microsoft Certifications

Continue into nearby exams from the same family. Each card keeps practice questions, study guides, flashcards, videos, and articles in one place.

More From This Family

Videos and articles for deeper review.

VideoFREE DP-600 Exam Guide 2026: Pass Microsoft Fabric Analytics Engineer (Direct Lake, DAX, Semantic Models)VideoFREE MD-102 Exam Guide 2026: Pass Microsoft Endpoint Administrator (Intune, Autopilot, Windows 11)VideoFREE MS-102 Exam Guide 2026: Pass Microsoft 365 Administrator (Entra, Defender XDR, Purview)VideoIs PL-300 Worth It in 2026? Power BI Salary & Career Reality ArticlePL-200 Exam Guide 2026: Skills, Retirement, and Renewal12 min read ArticlePL-400 Exam Guide 2026: Skills, Format, and Renewal12 min read ArticlePL-900 Exam Guide 2026: Current Outline and Free Prep19 min read ArticleMS-900 Retired 2026: FREE Guide + AB-900 Pivot17 min read

AB-620 Agent Builder Cheat Sheet