Skip to content

AB-620 open-book strategy ​

Microsoft role-based exams like AB-620 give you a restricted Microsoft Learn pane during the exam. This page is the plan I used: which pages to open, what to Ctrl+F, and which questions aren't worth searching. Every page and term was checked against live Learn in September 2026. Pages change, so re-verify before you rely on it.

Learn is a verification tool, not a discovery tool. Use it to confirm a documented limit, prerequisite or known limitation you half-remember. Never use it for judgment, "best approach", or ordering questions. That rule comes from my AB-100 attempt, where searches for judgment and ordering questions found nothing useful.


Why AB-620 is more Learn-friendly than AB-100 ​

Many of the facts that are easy to forget are stated word for word on a single Copilot Studio page: 25 websites, 2 × 15 tables, one index, 15 credits, 512 MB, Official source, the Fabric limits. On AB-100 those facts were buried; here they're in "Known limitations" and limit tables. A targeted lookup costs about 60 seconds and turns a coin-flip into a sure point.


Exam-room mechanics ​

  • Learn opens in a restricted pane inside the exam. You can have several Learn tabs; community Q&A is blocked.
  • Ctrl+F / Cmd+F searches the Learn page, not the question. Always navigate to the page, then Ctrl+F the term below.
  • Case studies can't be reopened once you leave them. Do any case-study lookups before you leave that section.
  • Use in-site Copilot Studio search (the search box on the Copilot Studio docs), not global Learn search. In practice, in-site search takes about 60 s; global search is noise.

The core loop: answer first, look up last ​

  1. First pass: answer every question from memory at about 45 s each. If a question hinges on a lookup-able fact (table below), pick your best guess, mark it for review, and write the lookup code on the whiteboard/notepad (e.g. K1).
  2. Review pass: with the remaining time, run only the marked lookups, cheapest first.
  3. Hard stops: 90 s per lookup; about 12–15 min total. If the first page + Ctrl+F doesn't hit, keep your guess and move on.

Lookups can never cost you an unanswered question, because you already answered everything.


Decision tree ​

Does the answer hinge on a NUMBER, LIMIT, PREREQUISITE, or "can/can't" fact?
├── YES → Is it in the lookup table below?
│   ├── YES → mark for review; look it up in the review pass (≤ 90 s)
│   └── NO  → one in-site Copilot Studio search in the review pass; no hit in top 3 → keep guess
└── NO → Is it ORDERING, "which should you do / best approach", mapping of roles, or a case-study judgment?
    └── DO NOT SEARCH. Use the decision rules. Move on.

Tabs to open at the start (in this order) ​

#TabURL (under learn.microsoft.com/en-us/microsoft-copilot-studio/ unless full)
1Knowledge sources (highest yield)knowledge-copilot-studio
2Add other agents (child/connected/Fabric limitations)authoring-add-other-agents
3Azure AI Search knowledgeknowledge-azure-ai-search
4Computer usecomputer-use
5REST API toolsagent-extend-action-rest-api
6Evaluation methodsanalytics-agent-evaluation-overview
7Copilot Studio docs home (for in-site search)/microsoft-copilot-studio/

Open the others only on demand: MCP, A2A, Foundry agent, App Insights, pipelines, authentication, flows, HTTP node.


Lookup table (page → Ctrl+F → what you're confirming) ​

Say the answer column to yourself before the exam. If you're sure, don't look it up.

Tab 1: Knowledge sources (knowledge-copilot-studio) ​

CodeCue in the questionCtrl+FExpected answer
K1How many websites / SharePoint sources?2525 generative; 4 classic
K2Classic Dataverse knowledge limits15 tables2 sources × 15 tables
K3"Official source" option greyed outOfficial sourceClassic orchestration only
K4Answers blocked / "can't find" with ungrounded offungroundedUncited answers withheld → add citation instructions; Fallback
K5Bing Custom Search / custom data in a generative agentBing Custom SearchGenerative answers node → Classic data
K6Large SharePoint files512200 MB; 512 MB for PDF/PPTX/DOCX (tenant graph grounding)
K7Tenant graph grounding requirementstenant graphAuthenticate with Microsoft; extra cost
K8More than 25 sources25Description-based filtering; uploaded files don't count

Tab 2: Add other agents (authoring-add-other-agents) ​

CodeCueCtrl+FExpected answer
M1Anything "Fabric Data agent can/can't"Known limitationsNo Redirect, no / reference, no M365 Copilot main agent
M2Connected agent inside another connected agentconnected agent (Known limitations)A main agent with connected agents can't itself be connected
M3Citations missing after a hopcitationMight not be maintained
M4Sub-agent and general knowledgegeneral knowledgeInherits the main agent's setting
M5Too many tools/topics30~30–40 choices → split into connected agents

Evaluation + Fabric: analytics-agent-evaluation-intro → Ctrl+F Fabric → not supported.

CodeCueCtrl+FExpected answer
S1How many indexes per connectiononly oneOne vector index
S2No citations from AI Searchmetadata_storage_pathNeeds a full-URL field
S3Private endpoint / network-isolated searchvirtual networkPower Platform VNet support
S4Keyless / governed authservice principalService principal or Entra ID integrated

Tab 4: Computer use (computer-use) ​

CodeCueCtrl+FExpected answer
C1Credits for N stepsCopilot Credits5 per step standard, 15 premium
C2Restrict what it can act onallowAccess control allow list
C3HTTPS-onlyEnforce HTTPSEnforce HTTPS setting
C4Stored sign-in secretKey VaultStored credentials (Key Vault)

Tab 5: REST API tools (agent-extend-action-rest-api) ​

CodeCueCtrl+FExpected answer
R1Solution field left blanksolutionSolution auto-created (action name + default publisher)
R2Blank parameter descriptionsdescriptionMust be filled in; names/types fixed by spec
R3Per-user delegated accessOAuth 2.0OAuth 2.0

Tab 6: Evaluation (analytics-agent-evaluation-overview) ​

CodeCueCtrl+FExpected answer
E1Which grader for XGeneral quality / Compare meaning / CustomGrader by what "correct" means; refusals → Custom
E2Result marked InvalidInvalidMissing expected answer

On-demand pages ​

CodePageCtrl+FExpected answer
P1agent-extend-action-mcpresourcesMCP = tools + resources; needs generative orchestration
P2add-agent-agent-to-agentchat history / custom connectorA2A sends full history + contextId; works in VNet
P3add-agent-foundry-agentActivityNeeds the Activity endpoint; project endpoint + Agent ID
P4advanced-bot-framework-composer-capture-telemetrydesignMode / randomExclude test chats; anonymous = random user ID
P5https://learn.microsoft.com/en-us/power-platform/alm/pipelinesManaged Environment / delegatedTargets must be Managed; delegated = SP + approval
P6configuration-end-user-authenticationAuthenticate with MicrosoftTeams/M365 only; publish to apply
P7flows-overviewWhen an agent calls the flow / capacityRequired trigger; Copilot Studio capacity
P8authoring-http-node30 / Continue on error30 s default timeout; handle errors yourself

Never search these (answer from rules) ​

Question shapeWhy notUse instead
Ordering / sequenceLearn describes steps, not the exam's numbered order (searching these didn't pay off on AB-100)App Insights: resource → connection string → enable → publish. Agents SDK: publish → channel → config → SDK → validate → deploy. AI Search: Add knowledge → connection → auth → index → add. REST: spec → auth → operations → connection. RAI: DLP → moderation → prompt mods
Connected-agent role mappingLoosely worded keyPublish → callable; Add → collaboration; Connections → invocation
"Best approach" / case-study designDocs don't prescribe combinationsIntegration shape: indexed → Copilot connector; live → connector knowledge; write → tool; no API → computer use; foreign agent → A2A
Disclaimers / tone / compliance placementJudgmentInstructions / prompt modifications + DLP; never per-topic edits
Model selection for promptsJudgmentPrompt model settings (Foundry catalog), not instructions

Final 60-second checklist ​

  • [ ] Open tabs 1–7 as soon as the exam starts.
  • [ ] Pace ≈ 45 s/question; strike unlimited / never / only / nothing options first.
  • [ ] Mark lookup questions and write their code (K1, M1, …); answer them anyway.
  • [ ] Do case-study lookups before leaving the case study.
  • [ ] Review pass: marked lookups only, ≤ 90 s each, ≤ 15 min total.

Personal study notes written from official documentation. Not affiliated with Microsoft or GitHub. Contains no exam questions.