AB-731 certifies Microsoft's AI Transformation Leader: a business decision-maker who can spot where generative AI creates real value, choose the right Microsoft AI apps and services for the job, and lead an organization's adoption of them without writing a line of code. The exam (skills measured as of July 22, 2026; a score of 700 or greater is required to pass) splits almost evenly across three areas: the business case for generative AI, the capabilities of Microsoft 365 Copilot and Foundry Tools, and the strategy for rolling both out responsibly. The recurring trap is treating this as a product-features quiz when it is really a judgment exam - most questions hinge on picking the right AI approach or governance move for a business scenario, not memorizing a UI. Expect heavy emphasis on grounding, responsible AI principles, and the practical mechanics of AI adoption (champions programs, licensing models, build-vs-buy calls) that separate a leader who can talk about AI from one who can actually run a rollout.
What This Cheat Sheet Covers
This topic spans 10 focused tables and 96 indexed concepts, 91 flashcards, 3 practice tests with 121 questions. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.
A jump-to index of every table row in this cheat sheet.
An interactive map of every table and concept in this topic.
Table 1: Foundational Concepts of Generative AI
AB-731 Business Value of Generative AI, Task: Identify the foundational concepts of generative AI. Covers how generative AI differs from other AI types, when to choose a pretrained versus fine-tuned model or a given Microsoft adoption path, and the cost, ROI, and risk factors (fabrication, bias, overreliance) that shape its business value.
| Concept | Example | Description | |
|---|---|---|---|
Traditional AI scores a customer's churn risk from past data. Generative AI drafts a plain-English summary of that risk report. | • Creates new content (text, images, audio, code) instead of only classifying or predicting from existing data. &bull • Not to be confused with predictive AI, which returns a score or label, never new content | ||
An LLM handles broad, open-ended writing tasks. An SLM runs one narrow support scenario on-device. | • LLMs encode broad linguistic patterns from internet-scale training • SLMs are compact and optimized for one scenario. &bull • Multi-modal models add text, image, and speech handling on top of either | ||
A base model trained on internet-scale text, ready to answer prompts with no company-specific training. | • A foundation model trained on large, general data before any task-specific customization. &bull • Most business use starts here via prompting, not custom training | ||
Training a small model on a few hundred labeled support tickets to match a brand's tone. | • Further trains a pretrained model on a smaller, task-specific dataset to specialize its behavior. &bull • Not always cheaper: it adds training and hosting cost, and needs repeating when data or the base model changes | ||
I heard a dog bark splits into 5 word-based tokens. | • The billed unit of input and output text a generative AI service processes. &bull • Cost and rate limits scale with token count, not with the number of users | ||
Compare hours saved per employee before and after a pilot against its adoption and scaling cost. | • The measurable value gained from AI weighed against the cost to adopt, manage, and scale it. &bull • Needs a defined baseline and success metric set before launch, not just a look at tool price | ||
An assistant states a policy clause exists in a document when it was never written. | • Confidently wrong, ungrounded output not supported by the source material. &bull • Not to be confused with bias (unfair treatment) or overreliance (a user trusting the output). | ||
A hiring tool scores equally qualified resumes differently based on demographic signals in the names. | • AI output that treats similar people or cases differently instead of consistently. &bull • One of Microsoft's six Responsible AI principles, distinct from transparency (explainability) and reliability (consistent function). | ||
A team ships a report built on an unverified AI-generated statistic that turns out wrong. | • Users accept incorrect or incomplete AI output because the system makes errors hard to spot. &bull • Mitigated by verification steps and setting realistic expectations, not by fine-tuning the model | ||
Need Word and Outlook writing help fast? Copilot. Need a custom RAG app? Foundry (PaaS). | • Four Microsoft adoption models trading customization for simplicity: ready-to-use Copilots, low-code SaaS, managed PaaS, and Azure infrastructure. &bull • Moving toward infrastructure gains control but costs setup speed | ||
Auto-routing support tickets by category is business automation. One employee drafting an email with Copilot is individual work. | • Use cases that change how the organization operates, often integrating multiple systems and AI types. &bull • Separate from individual work, which improves how one person or team works inside existing tools | ||
One Copilot-drafted reply template personalizes thousands of customer responses without adding staff. | • Cloud-hosted generative AI scales to meet growing demand without proportional headcount or infrastructure increases. &bull • A core business-value driver alongside automation, distinct from accuracy techniques like grounding |