Chapter 18 — AB-731 Exam Readiness¶
Part IV — Exam readiness
In 30 seconds¶
- The core idea: consolidate everything for AB-731 (AI Transformation Leader) into a checklist, a set of high-yield facts, and a full mock exam.
- Why it matters: this is your final rehearsal before exam day.
- The exam angle: pass mark is 700; skills measured as of July 22, 2026.
Exam overview¶
- Domain 1 — Identify the business value of generative AI solutions (35–40%).
- Domain 2 — Identify benefits, capabilities, and opportunities for Microsoft's AI apps and services (35–40%).
- Domain 3 — Identify an implementation and adoption strategy (20–25%).
Objective checklist¶
- [ ] Foundational concepts of generative AI (Ch 1)
- [ ] Benefits and capabilities of generative AI solutions: prompting, grounding, RAG, secure AI (Ch 2, 3, 4)
- [ ] Benefits and capabilities of Microsoft 365 Copilot and Microsoft Copilot (Ch 11)
- [ ] Extending Copilot: Copilot Studio, Microsoft Graph, build/buy/extend (Ch 12)
- [ ] Benefits and capabilities of Foundry Tools (Ch 13)
- [ ] Business case, cost drivers, and licensing (Ch 14)
- [ ] Align an AI strategy with responsible AI policies (Ch 15)
- [ ] Plan for AI adoption across the organization (Ch 16)
High-yield facts¶
Business value of generative AI (Domain 1)
- Generative AI creates content; other AI classifies/predicts. Pretrained = general, ready; fine-tuned = specialized via extra training. Prefer pretrained + prompting/grounding first.
- Cost driver = tokens (input + output). ROI = (value − total cost) / total cost. Value comes from scale and automation.
- Challenges: fabrications, reliability, bias. Prompt engineering = impact + techniques (Goal · Context · Source · Expectations).
- Grounding / RAG supplies data at query time; Azure AI Search is the RAG tool for custom apps. Data quality and representative datasets matter. Secure AI: application, data, authentication.
- Know when ML adds value (repeatable pattern + representative data) and the ML lifecycle (define → data → train → evaluate → deploy → monitor, iterative).
Microsoft AI apps & services (Domain 2)
- Microsoft 365 Copilot = work-grounded productivity (per-user). Microsoft Copilot (Chat) = web/ mobile, work data with a license. Researcher = deep multi-source research; Analyst = quantitative data analysis.
- Copilot Studio = low-code build/extend agents; agent builder = no-code. Microsoft Graph = data fabric; connectors bring external data in.
- Build / buy / extend: prefer buy → extend → build. Extensibility = agents, connectors, plugins.
- Microsoft Foundry = build custom AI; Foundry Tools = model catalog, Azure AI Search, Azure AI Vision. Match a model to the need (fit, not size). Benefits = scalability + security.
- Integrated Microsoft AI solution = risk mitigation + safety from a shared security/compliance foundation.
Implementation & adoption (Domain 3)
- Responsible AI at scale = governance + an AI council (cross-functional strategy, oversight, alignment) + ensuring the six standards (fairness, reliability & safety, privacy & security, inclusiveness, transparency, accountability).
- Adoption = program: executive sponsorship + adoption team + AI champions program + training + measurement. Address barriers (skills, trust, use cases, data/governance).
- Plan for impacts to data, security, privacy, cost.
- Licensing: Copilot = per-user subscription (monthly / included); Foundry = pay-as-you-go or commitment tiers.
Mock exam — AB-731¶
40 original questions, weighted toward the exam's domains. Answers with explanations are under each item. Target ≥ 70% before sitting the real exam.
Domain 1 — Business value of generative AI¶
1. Which task is best suited to generative AI rather than predictive ML?
- A. Forecasting next quarter's demand
- B. Drafting a customer proposal
- C. Classifying support tickets
- D. Detecting anomalies
Answer
B. Drafting = content creation (generative). The others are predictive.
2. A pretrained model differs from a fine-tuned model in that:
- A. Pretrained is trained on a broad dataset and works out of the box; fine-tuned adds narrow domain training
- B. They are identical
- C. Pretrained needs your data to function
- D. Fine-tuned is always cheaper
Answer
A. Fine-tuning adds specialized training on top of a general pretrained model.
3. For a general drafting need, the most cost-effective approach is usually:
- A. Fine-tune a custom model first
- B. Use a pretrained model with good prompting and grounding
- C. Build a model from scratch
- D. A rule-based engine
Answer
B. Prefer pretrained + prompting/grounding before fine-tuning.
4. The primary cost driver of consumption-based generative AI is:
- A. Employee headcount
- B. Tokens (input + output)
- C. Number of slides
- D. Office size
Answer
B. Tokens drive consumption cost.
5. ROI for an AI initiative is best expressed as:
- A. Cost only
- B. (Value created − total cost) / total cost
- C. Tokens per second
- D. Number of licenses
Answer
B. ROI weighs value against total cost.
6. Generative AI creates business value primarily through:
- A. Scale and automation of knowledge work
- B. Replacing all employees
- C. Eliminating governance
- D. Reducing internet usage
Answer
A. Value comes from scaling and automating repetitive knowledge work.
7. Which is a known challenge of generative AI solutions?
- A. Perfect accuracy
- B. Fabrications, reliability issues, and bias
- C. Zero cost
- D. No need for data
Answer
B. Fabrications, reliability, and bias are named challenges.
8. Retrieval-augmented generation (RAG) improves answers by:
- A. Retraining the model each time
- B. Retrieving relevant data and adding it to the prompt before generation
- C. Removing all context
- D. Encrypting the model
Answer
B. RAG grounds the prompt with retrieved data.
9. Which most affects the quality of an AI solution's output?
- A. Data quality and representative datasets
- B. Monitor size
- C. Office location
- D. The model's name
Answer
A. Representative, high-quality data drives output quality.
10. When does machine learning add value?
- A. When a fixed rule already solves it
- B. When there's a repeatable pattern and enough representative data to learn from
- C. For displaying today's date
- D. Never
Answer
B. ML fits learnable patterns with representative data.
11. The machine-learning lifecycle is best described as:
- A. A one-time training event
- B. Iterative: define → data → train → evaluate → deploy → monitor
- C. Only deployment
- D. Only data collection
Answer
B. It's an iterative loop including monitoring.
12. "Secure AI" security considerations include:
- A. Application security, data security, and authentication
- B. Font choice
- C. Slide transitions
- D. Keyboard layout
Answer
A. Those three layers are named in the objectives.
13. Prompt engineering, for a business leader, means:
- A. Writing code to train models
- B. Crafting clear instructions and choosing good sources to improve output
- C. Configuring servers
- D. Building a data center
Answer
B. It's about better instructions and sources — no code.
Domain 2 — Microsoft AI apps & services¶
14. A leader needs a cited market briefing synthesizing internal docs and the web. Best fit?
- A. Analyst
- B. Researcher
- C. Excel
- D. A saved prompt
Answer
B. Researcher does deep, multi-source research and synthesis.
15. A team needs quantitative analysis of a large sales dataset. Best fit?
- A. Researcher
- B. Analyst
- C. Copilot Pages
- D. Outlook
Answer
B. Analyst performs data analysis.
16. The main difference between free Microsoft Copilot chat and Microsoft 365 Copilot is:
- A. Color
- B. Work-data grounding via Microsoft Graph (licensed) vs primarily web
- C. Only free chat is secure
- D. None
Answer
B. Work-data grounding requires the license.
17. A benefit of an integrated Microsoft AI solution is:
- A. Separate security models per tool
- B. Shared security/compliance foundation → risk mitigation and safety
- C. No governance needed
- D. Offline-only operation
Answer
B. Integration means a consistent, secure foundation.
18. To let Copilot answer from a non-Microsoft CRM, the best approach is:
- A. Build a new AI app from scratch
- B. Use a Microsoft 365 Copilot connector to bring CRM data into Microsoft Graph
- C. Email the data around
- D. Fine-tune a model
Answer
B. Connectors bring external data into Graph.
19. Which tool suits makers/IT building connector-rich agents with actions?
- A. The no-code agent builder
- B. Microsoft Copilot Studio
- C. Excel
- D. Outlook
Answer
B. Copilot Studio is low-code with connectors and actions.
20. The recommended default order for meeting an AI need is:
- A. Build → extend → buy
- B. Buy/adopt → extend → build
- C. Always build custom
- D. Never extend
Answer
B. Adopt if it fits, extend to close gaps, build only for bespoke.
21. Which platform is for building custom, customer-facing AI solutions?
- A. Microsoft 365 Copilot
- B. Microsoft Foundry
- C. Outlook
- D. A notebook
Answer
B. Foundry is the build platform for custom AI.
22. Which Foundry tool provides retrieval/grounding (RAG) over your data?
- A. Azure AI Vision
- B. Azure AI Search
- C. PowerPoint
- D. Copilot Pages
Answer
B. Azure AI Search handles retrieval/grounding.
23. How should a leader match a model to a need?
- A. Always pick the largest/newest
- B. Pick the model meeting requirements (quality, latency, modality) at acceptable cost
- C. Always pick the cheapest
- D. Let it choose itself
Answer
B. Fit-for-purpose, not maximalism.
24. Headline benefits of Microsoft Foundry are:
- A. Scalability and security
- B. Free unlimited usage
- C. No responsible AI needed
- D. Replaces Copilot for everything
Answer
A. Enterprise scalability and security.
25. Microsoft Graph is best described as:
- A. A charting library
- B. The data fabric/API for Microsoft 365 that Copilot grounds on and connectors feed
- C. A spreadsheet
- D. A meeting app
Answer
B. It's the data gateway for Microsoft 365.
26. Extending Copilot (vs building) is preferred when:
- A. The need is bespoke and unique
- B. A connector or Copilot Studio can close the gap on the existing foundation
- C. You want maximum cost
- D. Never
Answer
B. Extend to close gaps before building custom.
27. Azure AI Vision in Foundry Tools is used to:
- A. Understand images (e.g., extract text from scans)
- B. Send email
- C. Schedule meetings
- D. Store passwords
Answer
A. Vision handles image understanding.
Domain 3 — Implementation & adoption¶
28. Which body guides AI strategy, oversight, and cross-functional alignment?
- A. The help desk
- B. An AI council
- C. A single power user
- D. The vendor
Answer
B. A cross-functional AI council.
29. Ensuring solutions meet responsible-AI standards means checking against:
- A. Fairness, reliability & safety, privacy & security, inclusiveness, transparency, accountability
- B. Speed, color, size, price
- C. CPU, memory, disk, network
- D. Tokens, weights, layers, epochs
Answer
A. The six responsible-AI principles.
30. A company adopting AI at scale should first:
- A. Ban all AI
- B. Establish governance principles and an AI council
- C. Let everyone decide alone
- D. Buy the priciest product
Answer
B. Governance + council enable safe, aligned adoption.
31. Licenses are bought but usage is low after months. Missing ingredient?
- A. More expensive licenses
- B. A structured adoption program (sponsorship, training, champions, use cases)
- C. Faster internet
- D. Disabling the tool
Answer
B. Low adoption reflects missing change management.
32. An AI champions program is:
- A. A replacement for IT
- B. Peer advocates who model good use, coach colleagues, and share feedback
- C. The team that builds models
- D. A budget committee
Answer
B. Grassroots peer advocates.
33. Which four impact areas must adoption planning address?
- A. Data, security, privacy, cost
- B. Fonts, colors, themes, icons
- C. CPU, memory, disk, network
- D. Tokens, weights, layers, epochs
Answer
A. Data, security, privacy, and cost.
34. To give 1,000 employees Copilot in their apps with predictable cost, use:
- A. Pay-as-you-go tokens
- B. Per-user (monthly) Microsoft 365 Copilot subscription
- C. Foundry commitment tier
- D. Free web chat only
Answer
B. Copilot is per-user subscription.
35. A custom Foundry app with variable usage is best priced initially as:
- A. Per-user subscription
- B. Pay-as-you-go (consumption)
- C. One-time license
- D. Unpriceable
Answer
B. Consumption suits variable workloads; commitment tiers later.
36. Why pair governance with enablement?
- A. To slow adoption
- B. Rules without approved tools/training push people to ungoverned "shadow AI"
- C. They're unrelated
- D. To avoid using AI
Answer
B. Guardrails + enablement = fast and safe.
37. A sound AI business case:
- A. Chooses the cheapest regardless of outcome
- B. Weighs value (time saved, quality, capability) against total cost, including adoption
- C. Ignores adoption cost
- D. Always builds custom
Answer
B. Balance value against total cost, including adoption.
38. Foundry commitment tiers are attractive when:
- A. Usage is unpredictable and tiny
- B. Usage is steady and predictable, to lower the rate
- C. There is no usage
- D. Only for Copilot
Answer
B. Commitment tiers discount steady, predictable usage.
39. A proposed lending model may disadvantage a group. Governance should:
- A. Launch it anyway
- B. Flag it for a fairness review and remediate before launch
- C. Ignore it
- D. Blame the model
Answer
B. Review against the standards (fairness) before launch.
40. The strongest predictor of adoption success is:
- A. No executive involvement
- B. Visible executive sponsorship plus measurement of usage and value
- C. Secrecy
- D. No training
Answer
B. Sponsorship + measurement drive and prove adoption.
\u2705 Ready check: if you can explain why each wrong option is wrong, you understand the material, not just the answer. Revisit any chapter where you missed two or more questions in its domain.