AB-100 open-book experiment
Before sitting AB-100, I ran timed Microsoft Learn searches that imitated exam-day lookups across four question types, to find out which kinds of question are worth searching. The short answer: product facts sometimes, tool matching usually, and architectural judgment or ordering never.
Experiment Summary
Date: 2026-03-08 Method: Automated Playwright browser searches on learn.microsoft.com simulating exam-day lookup behavior Scope: 2 practice scenarios per category across 4 question types
Category-Level Findings
Product Features: MIXED (1 EASY, 1 HARD)
- When searchable: Features with dedicated doc pages (NLU+, knowledge sources, voice config) are findable via in-site doc search within Copilot Studio docs
- When NOT searchable: Niche feature names ("business terms") that don't have prominent doc pages. Global MS Learn search drowns these in noise
- Key insight: The in-site search within a product's doc tree is dramatically better than global search. Go to the product doc landing page FIRST
Tool/Service Matching: MEDIUM (both MEDIUM)
- Individual tools are well-documented and findable with targeted searches
- BUT matching questions require 2-3 separate searches to confirm each mapping
- The knowledge sources summary page is a gold mine — bookmark it
- Time cost: ~2-3 minutes per matching question (too slow for frequent use)
Architectural Judgment: HARD-IMPOSSIBLE (1 HARD, 1 IMPOSSIBLE)
- Docs describe WHAT each feature does, not WHEN or WHY to combine them
- ROI metric selection, architecture trade-offs, and design decisions are NOT codified in docs
- Security/governance pages list available controls but don't prescribe combinations for specific scenarios
- Verdict: Don't search for these — trust your preparation
Ordering/Process: IMPOSSIBLE (both IMPOSSIBLE)
- Docs DO NOT present deployment or governance steps as numbered sequences
- Lifecycle content is descriptive, not prescriptive ("here are the phases" not "do step 1, then step 2")
- Verdict: Never search for ordering questions — total time waste
Grading Scale Used
| Grade | Criteria | Exam-Day Action |
|---|---|---|
| EASY | Found exact answer in first search, < 90 seconds | Search if unsure |
| MEDIUM | Found after refining search or 2+ results, < 3 min | Search only if completely stuck |
| HARD | Found partial info (helps eliminate, not definitive) | Don't search — too slow |
| IMPOSSIBLE | Learn didn't help, or would take > 5 min | Never search |
Exam-Day Decision Tree
Is the question about a SPECIFIC PRODUCT FEATURE or NAMED CAPABILITY?
├── YES → Do you know the product area (Copilot Studio, Purview, etc.)?
│ ├── YES → Go to that product's doc page → use IN-SITE search → ~60 sec
│ └── NO → Try global search with exact feature name → ~90 sec max
│ └── If no hit in top 3 results → STOP, guess and move on
└── NO → Is it a TOOL MATCHING question?
├── YES → Can you eliminate 2+ wrong answers from memory?
│ ├── YES → Don't search, go with your best answer
│ └── NO → Search for the ONE tool you're least sure about → 1 search max
└── NO → Is it JUDGMENT, ORDERING, or PROCESS?
└── YES → DO NOT SEARCH. Trust your preparation. Move on.Search Technique Ranking (Most → Least Effective)
- In-site doc search (search box within product docs) — Found NLU+ instantly
- Direct URL navigation (go to known doc page) — Knowledge sources page was perfect
- Global search with Documentation tab — Helps for well-known tools (Purview, Content Safety)
- Global search unfiltered — Drowns you in 400K+ results, rarely useful
- AI-generated answers (when they appear at top of search) — Free cheat code, use if shown
- Cmd+F within a page — Once on the right page, scan for keywords fast
Pre-Bookmarkable URLs
Open these tabs BEFORE the exam starts:
| Tab | URL | Use For |
|---|---|---|
| Copilot Studio Docs Home | learn.microsoft.com/en-ca/microsoft-copilot-studio/ | In-site search for ANY Copilot Studio feature |
| Knowledge Sources Summary | learn.microsoft.com/en-ca/microsoft-copilot-studio/knowledge-copilot-studio | Data source matching questions |
| Security & Governance | learn.microsoft.com/en-ca/microsoft-copilot-studio/security-and-governance | DLP, data residency, compliance questions |
| NLU/NLU+ Configuration | learn.microsoft.com/en-ca/microsoft-copilot-studio/nlu-plus-configure | Language model questions |
| Managed Environments | learn.microsoft.com/en-ca/power-platform/admin/managed-environment-overview | Power Platform governance features |
| AI Content Safety | learn.microsoft.com/en-ca/azure/ai-services/content-safety/overview | Harmful content detection |
| Purview Info Protection | learn.microsoft.com/en-ca/purview/information-protection | Data classification & labeling |
| CAF AI Adoption | learn.microsoft.com/en-ca/azure/cloud-adoption-framework/scenarios/ai/ | CAF strategy/governance (rarely useful) |
| MS Learn Search | learn.microsoft.com/en-ca/search/ | Global fallback search |
Time Budget Rules
Total exam time for MS Learn searches: 10-15 minutes MAX across entire exam
| Rule | Rationale |
|---|---|
| Max 3-4 searches total across the entire exam | Each search costs 1-3 min including reading; more than 4 eats into answering time |
| 90-second hard cutoff per search | If you haven't found it in 90 sec, you won't find it |
| Never search for ordering/process questions | 0% success rate in experiment |
| Never search for "which combination" architecture questions | Docs describe features, not prescriptions |
| DO search for: specific named features you can't remember | Highest ROI searches |
| DO search for: which tool does X when stuck on one specific tool | One targeted search can confirm/deny |
| Use in-site search, not global search | 10x more effective based on experiment |
Hypothesis Validation
| Category | Predicted Grade | Actual Grade | Prediction Accurate? |
|---|---|---|---|
| Product Features (Q-A, Q-B) | EASY–MEDIUM | HARD and EASY | Partially — depends on feature prominence |
| Tool Matching (Q-C, Q-D) | MEDIUM–HARD | MEDIUM and MEDIUM | Yes |
| Architectural Judgment (Q-E, Q-F) | HARD–IMPOSSIBLE | IMPOSSIBLE and HARD | Yes |
| Ordering/Process (Q-G, Q-H) | IMPOSSIBLE | IMPOSSIBLE and IMPOSSIBLE | Yes |
Overall hypothesis confirmed: MS Learn is useful for ~25% of AB-100 questions (product features with dedicated doc pages), marginally useful for ~25% (tool matching if you have time), and NOT useful for ~50% (judgment + ordering). Budget 3-4 searches maximum across the entire exam.
Bottom Line
MS Learn is a verification tool, not a discovery tool for AB-100. If you already know roughly what to search for, it can confirm. If you're lost on a question, searching will waste time without helping. The exam tests architectural judgment — docs describe features, not judgment.