The AB-100 exam is Microsoft's first Expert-level agentic AI credential, certifying the ability to design, prototype, and govern AI-driven business solutions that transform enterprise processes across the Microsoft stack. Launched in beta in late 2025 (content current as of July 22, 2026), it requires at least one qualifying associate certification as a prerequisite and targets solution architects who already know Dynamics 365, Power Platform, Copilot Studio, and Microsoft Foundry individually and must now combine them into agentic-first architectures. It tests three areas: Plan AI-powered business solutions (25-30%), Design AI-powered business solutions (25-30%), and Deploy AI-powered business solutions (40-45%), with the heaviest weight on deployment, monitoring, testing, ALM, and responsible AI/security/governance rather than initial design. Many questions reward the architect's judgment call - when to extend Microsoft 365 Copilot versus build a custom agent, when a task agent suffices versus an autonomous one, when a small language model beats a frontier model - so expect scenario-driven tradeoff questions more than feature recall.
What This Cheat Sheet Covers
This topic spans 18 focused tables and 197 indexed concepts, 6 practice tests with 238 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: Analyze Requirements for AI-Powered Business Solutions
AB-100's Plan domain task Analyze requirements for AI-powered business solutions starts with judging where an agent genuinely fits (task automation, data analytics, decision-making) before writing a single instruction. It also covers whether an agent's grounding data is trustworthy enough to build on, and how to organize business data so other AI systems in the tenant can reuse it too.
| Concept | Example | Description | |
|---|---|---|---|
Employee asks an agent to summarize open tickets across three systems: fits a Productivity agent, not Action or Automation. | Three agent types by autonomy: Productivity (retrieve/synthesize for faster decisions), Action (a specific task inside a defined workflow), Automation (complex multi-step work with minimal oversight, needing real governance). | ||
A support chat that recalls what the user asked two turns ago relies on the Memory component, not Retrieval. | Five components make up an agent: model, instructions, retrieval (grounding data and context), actions (functions/APIs/tools), and memory (conversation history and state). Not to be confused with each other. | ||
Copying an approved form's fields into a database every time it's submitted fits a Power Automate cloud flow, not an agent flow. | Agent flows suit AI-driven work that needs reasoning through ambiguity; cloud flows suit fixed-rule, repeatable automation. Microsoft treats them as complementary, not one replacing the other. | ||
A grounding source with correct facts from 14 months ago still fails the freshness check even though nothing in it is wrong. | Grounding data is reviewed against defined criteria (accuracy, relevance, timeliness/freshness, cleanliness, availability) on an ongoing basis, not as a one-time gate before launch. | ||
An agent quotes a retired refund policy verbatim: technically grounded, but grounded on a stale source. | Ungroundedness is the model deviating from what its source material says; an answer built faithfully on outdated or dirty source data is still "grounded", the problem sits in the data, not the model. | ||
A CRM record only grounds an agent's answer once it's connected and indexed as a knowledge source through a Copilot connector. | Data isn't usable for grounding just because it exists somewhere in the tenant; it must be connected/indexed as a knowledge source, and access still respects each source's own per-user permissions. | ||
Sales, finance, and support publish governed data products to one OneLake instead of three separate, disconnected data stores. | OneLake is the central data lake where data domains publish governed data products that AI agents consume via Fabric IQ or Foundry IQ; Microsoft 365 content still flows separately through Microsoft Graph. | ||
Default to Fabric IQ or Foundry IQ for retrieval; build a custom pipeline only when a documented gap remains. | Microsoft's default is built-in retrieval (Fabric IQ, Foundry IQ, SharePoint connection); reserve custom integration for cases where built-in controls fail a documented regulatory or operational requirement. | ||
How many units of SKU123 are in stock right now? calls for an MCP tool call, not a static knowledge search. | MCP suits actions and real-time data (live inventory, ticket creation) with identity, policy, and audit controls; RAG-style knowledge search suits static content like policy documents. |