AB-100 cheat sheet
Last-minute reference for AB-100. Deploy is the largest domain at 40–45%; Plan and Design are 25–30% each.
Exam facts
| Passing score | 700 / 1000 |
| Duration and number of questions | Not published on the exam page |
| Price | Set per country or region; shown at booking |
| Languages | 13 |
| Renewal | Free, annually |
| Also required for the credential | One qualifying associate certification |
| Outline | Skills measured as of October 14, 2026 |
Skill areas
| Area | Weight |
|---|---|
| Deploy AI-powered business solutions | 40–45% |
| Plan AI-powered business solutions | 25–30% |
| Design AI-powered business solutions | 25–30% |
Requirement to answer
| Requirement | Answer |
|---|---|
| Fastest, cheapest start | Prebuilt agent, then extend |
| Users live in Microsoft 365 | Extend Microsoft 365 Copilot |
| Agent acts on events, no user trigger | Autonomous agent, with limits and approvals |
| Agent answers a user’s question | Prompt-and-response agent |
| Defined steps on request | Task agent |
| Handle unmatched conversation | Fallback topic |
| Standard interface to tools | Model Context Protocol |
| Agents talk to agents across platforms | Agent2Agent |
| Application with no API | Computer use |
| Cheaper model for simple requests | Model router |
| Narrow, high-volume task, low cost | Small language model |
| Wrong answers, normal latency | Fix grounding data |
| Same promotion for many objects | Versioned, coordinated ALM |
| Prevent users seeing restricted content | Permissions on grounding data |
| Who changed the model or data | Audit trail |
| Region and data movement | Residency and movement validation |
Agent types
| Type | Trigger | Control to design |
|---|---|---|
| Prompt and response | User question | Grounding and answer quality |
| Task agent | User or process request | Defined steps, error handling |
| Autonomous agent | Event or schedule | Limits, approvals, escalation, audit |
Platforms at a glance
| Platform | Think of it as |
|---|---|
| Microsoft 365 Copilot | Where employees already work; extend it with agents |
| Copilot Studio | Low-code agents, topics, flows, connectors, MCP |
| Microsoft Foundry | Models, custom models and the Agents Service |
| Dynamics 365 apps | Built-in copilots and agents for finance, supply chain, sales and service |
| Power Platform | Apps and flows with AI features, governed by environments |
Grounding data properties
Accuracy, relevance, timeliness, cleanliness, availability.
ALM objects
Data for models and agents; Copilot Studio agents with connectors and actions; Foundry Agents Service; custom models; AI in Dynamics 365 finance and supply chain; AI in Dynamics 365 customer experience and service.
Deploy checklist
- Monitor: usage, quality, latency, errors, feedback.
- Test: fixed set, pass criteria, end-to-end across apps.
- Promote: development, test, production; versioned packages.
- Secure: agent, model and data access; prompt manipulation mitigations.
- Govern: responsible AI review, ownership, audit.
Common traps
- A custom model where grounding would do.
- Autonomy where approval is required.
- A bigger model before checking data.
- Editing production directly.
- Mixing up A2A (agent to agent) and MCP (model or agent to tools).
Practise with the practice test and re-read the exam tips.