9.2. High-Frequency Traps and Exam Patterns
These are the distinctions the exam tests most reliably: the questions that catch candidates who learned features in isolation but didn't internalize the differences.
The most important distinctions for AB-730:
| Topic | The Trap | The Correct Understanding |
|---|---|---|
| Chat vs agent vs Cowork | "Ask Researcher to send the email and book the meeting" | Chat drafts, agents specialize (Researcher reports, Analyst analyzes), and Cowork acts across apps with your approval |
| Cowork access | "A Copilot license is enough for Cowork" | Cowork needs the license and usage-based billing granted by a spending policy, and students aren't eligible |
| Cowork approvals | "Always allow means forever" | Approvals and Always allow last for the current session and can be revoked under Permissions |
| Scheduled prompt limits | "Chat and Cowork share one scheduler" | Copilot Chat: up to 10, in Settings. Cowork: up to 25, on the Automations page. Event triggers exist only in Cowork |
| Save vs Schedule | "I saved the prompt, so it will run automatically" | Saving ≠ automating. Scheduling runs it, and needs a license |
| Prebuilt agents | "Customize Researcher with our rules" | Prebuilt agents can't be customized. Use Agent Builder (small team, Q&A) or Copilot Studio (org-wide, external, workflows) |
| Custom skill vs custom agent | "Build a Cowork skill in Copilot Studio" | A Cowork custom skill is a SKILL.md in OneDrive, discovered at session start |
| Work IQ | "Turning off Work IQ turns off personalization" | Work IQ off stops grounding on work data. Memory and profile preferences can still apply. It's on by default for licensed users |
| Memory | "Deleting the chat deletes what Copilot remembers" | Saved memories persist until deleted in Manage saved memories. Turning memory off doesn't delete it |
| Facilitator | "Add Facilitator to a Meet now call" | Scheduled meetings only: not channel meetings, instant meetings or calls |
| Intelligent recap | "Recap needs Teams Premium" | Teams Premium or a Copilot license, plus a recorded or transcribed meeting of 5+ minutes |
| Connectors | "Users install connectors themselves" | Admins enable connectors. They're read-only by default, permission-trimmed, and need a license to view |
| Notebook sharing | "Share the notebook read-only with the group" | Notebook invitees always get edit access, no groups, no external sharing |
| Fabrication frequency | "Hallucinations are rare edge cases" | Fabrications are structural and can occur on any topic without grounding |
| Permission inheritance | "Copilot or an agent can surface files I can't open" | Copilot, agents and Cowork inherit your exact permissions |
| DLP and output | "DLP only restricts what Copilot can access" | DLP can also exclude protected content from what Copilot outputs |
| Prompt injection | "Prompt injection only affects developers" | Any user who asks Copilot to process external content is exposed |
Recognizing scenario question patterns:
AB-730 scenarios follow predictable patterns. Once you recognize the pattern, the correct answer becomes clearer.
Pattern 1 — "Which feature or agent should they use?" → The question describes a workflow goal. First decide the level of delegation (answer, specialist or doer), then match the specific feature using the frameworks from Phases 4–8.
Pattern 2 — "What went wrong?" → Describes a suboptimal output or a feature that "didn't work." Diagnose the root cause: missing GCSE elements, Work IQ off, a scoped chat pane, an unsupported meeting type, a missing license or admin setting, or an AI risk not mitigated.
Pattern 3 — "What should they do next?" → After Copilot or Cowork produces output. The answer almost always involves human review before action, especially for external, high-stakes or factual content, and reviewing Cowork's approvals rather than approving everything.
Pattern 4 — "What is the risk?" → Identify the risk category: fabrication, injection, over-reliance, sensitive data exposure, a permission boundary issue, or an untrusted custom skill.
Pattern 5 — "What is the best prompt?" → The correct answer has a clear goal plus the context, source and expectations the task needs. Wrong answers are long but unstructured, or missing the source or audience.
