The short version
This sheet is for the last pass before you book or sit AB-100. It lists the facts, the term pairs that get mixed up, simple decision rules, and a last-day checklist. It does not cover everything on the exam, and it is not a substitute for the Microsoft study guide. Rules of thumb here are our own teaching aids, not Microsoft statements.
Exam facts at a glance
| Item | What Microsoft publishes (checked October 22, 2026) |
|---|---|
| Exam | AB-100: Agentic AI Business Solutions Architect |
| Passing score | 700 on a scale of 1 to 1,000 (scaled, not 70%) |
| Languages | 13, including English, Spanish, French, Portuguese (Brazil) |
| Number of questions | Not published for AB-100; Microsoft says most exams have 40 to 60 |
| Time | Not published for AB-100; general table says 100 or 120 minutes, depending on labs |
| Also required | At least one qualifying associate certification |
| Renewal | Free online assessment, once a year |
| Update | English exam changes on November 16, 2026; localized about eight weeks later |
Weightings: two versions
Microsoft’s pages now show the outline that applies from November 16, 2026. The earlier outline (dated October 14, 2026) is the one we published before, and it applies until the switch.
| Domain | Until November 16, 2026 | From November 16, 2026 |
|---|---|---|
| Plan | 25-30% | 35-40% |
| Design | 25-30% | 30-35% |
| Deploy | 40-45% | 20-25% |
Check which version your exam date falls under. If you booked a date after the switch, weight your review toward Plan.
Domain 1: Plan
Key concepts
- Ask where agents help: task automation, data analytics, decision support.
- Grounding data is judged on accuracy, relevance, timeliness, cleanliness, and availability.
- Follow the AI adoption process from the Cloud Adoption Framework for Azure.
- Know the build, buy, or extend decision, and the cost side: total cost of ownership and an ROI case.
- A model router sends each request to the most suitable model.
- An AI Center of Excellence is the shared team and set of practices for AI work.
- Prompt libraries and prompt guidelines make results repeatable.
Easily confused pairs
- Build is for a gap nothing covers. Buy or use a prebuilt agent is for a common need. Extend is for a product you already own.
- A custom model is for a need that prompting and grounding cannot meet. A small language model is for a narrow, cheaper job.
- Grounding data is what the agent reads. A prompt is what you ask it.
Decision rules
- If the scenario says “common process, prebuilt option exists”, think prebuilt first.
- If it says “poor answers”, check data quality before changing the model.
- If it names costs beyond licenses, count them in total cost of ownership.
Domain 2: Design
Key concepts
- Agent kinds: task, autonomous, prompt-and-response.
- Copilot Studio pieces: topics (with fallback), agent flows, prompt actions, generative orchestration.
- Extensibility: Model Context Protocol (MCP), computer use for apps and websites, custom models in Microsoft Foundry, agents in Microsoft 365 Copilot.
- Prebuilt AI in Dynamics 365 for finance and supply chain, and for customer experience and service.
- The Power Platform Well-Architected Framework applies to intelligent apps.
Easily confused pairs
- A knowledge source is for what the agent reads. A tool or connector is for what the agent does.
- Standard natural language processing is for fixed, predictable phrases. Generative orchestration is for open requests and tool choice.
- MCP is for plugging in tools and data through a common protocol. Agent2Agent (A2A) is for agents talking to each other.
- Computer use is for systems with no API. An API or connector is the cleaner choice when one exists.
- A declarative agent is configured on a Microsoft base. A custom engine agent brings its own orchestration.
Decision rules
- If an API exists, prefer it over screen-driving.
- If the scenario says “must be generally available”, rule out preview-only answers.
- If the customer already owns a Dynamics 365 app, check its built-in AI before building new.
Domain 3: Deploy
Key concepts
- Monitoring: track agent performance, usage feedback, telemetry, and issues, then tune.
- Testing: define metrics and test sets, validate custom models, plan end-to-end scenarios across apps.
- ALM: separate approaches for data, Copilot Studio agents with connectors and actions, Microsoft Foundry Agent Service, custom models, and Dynamics 365 AI.
- Responsible AI, security, governance, and compliance: agent security, model security, prompt manipulation, data residency, access control on grounding data, audit trails.
Easily confused pairs
- Monitoring is for how a live agent behaves. Testing is for whether it is ready to go live.
- Security is for who and what can reach the agent. Governance is for the rules and approvals around it.
- Data residency is about where data sits. Access control is about who may read it.
- Development, test, and production are separate environments. Solutions and pipelines carry changes between them.
Decision rules
- If an answer edits production directly, be suspicious.
- If the scenario mentions regulation or region, check residency before anything else.
- If someone must prove who changed a model, think audit trail.
Common traps
- Picking the most impressive technology instead of the one that fits the stated constraint.
- Skipping the sentence in the stem that holds the deciding detail: region, license, preview status, or existing app.
- Treating security as a single control.
- Fixing the wrong layer, such as rewriting instructions when the knowledge source is stale.
- Ignoring that multi-select questions say how many to choose.
- Using older notes with outdated names. Microsoft now writes “Foundry Tools” and “Foundry Models”.
Last-day checklist
- Confirm your exam date against the November 16 update, and know which weightings apply.
- Re-read the current Microsoft study guide headings, not the old notes.
- Review the look-alike pairs above out loud.
- Check your booking for time, language, and delivery.
- Prepare your ID and your room if you test online.
- Skim your own notes on Deploy topics or Plan topics, whichever carries more weight on your date.
- Sleep. Scenario reading needs a clear head.
More guides for AB-100
- AB-100 Study Plan: 4-Week and 8-Week
- Is the AB-100 Exam Hard?
- AB-100 Prerequisites and Skills Check
- AB-100 Renewal and Retakes Explained
- Is AB-100 Worth It? An Honest Look
- AB-100 vs AB-620: Which Exam First?
Next step
Test the sheet against scenarios with the sample questions, or go deeper with domains explained and the exam format page. Our AB-100 practice tests are original sets for timed practice. The exam overview links everything together.