AB-100 Exam Guide 2026: Microsoft Agentic AI Business Solutions Architect Certification

If you're staring at the AB-100 exam page on Microsoft Learn trying to figure out what you actually need to know, this guide is for you. It's the exam guide we'd hand a senior architect who just told us "I have eight weeks and a credit card — what do I do?"

We're going to cover the exam structure exactly as Microsoft tests it, the three skill domains in proportion to their weight, the specific question types you'll see, the July 22, 2026 content refresh, and a study plan that respects how busy you actually are.

For the "should I even take this?" question, we have a separate piece on AB-100's cost-benefit math. This post assumes you've already decided yes.

The exam at a glance

One detail that catches people off guard: AB-100 does not currently include labs, but Microsoft has signaled labs may be added in future updates. If you see a prep guide claiming a specific lab count, treat it as out of date.

The three skill domains (with realistic study weighting)

Microsoft publishes the skill weights as ranges, not fixed percentages. Treat the upper end of each range as the planning number — you don't want to be the candidate who allocated for the lower bound and saw the actual exam at the upper bound.

Plan AI-powered business solutions (25–30%)

This is the "before we build" domain. The exam tests whether you can take a vague business requirement and translate it into a concrete agentic architecture decision.

Critical sub-topics:

• Cloud Adoption Framework's AI adoption process — you need to know the phases by name (Strategy, Plan, Ready, Adopt, Govern, Manage, Secure) and what activities sit in each.
• When to build vs. buy vs. extend — a common case-study fork. Microsoft expects you to default to "extend" (use Copilot or Foundry building blocks) before "build" (custom agents from scratch).
• Multi-agent architecture design — orchestrating Microsoft 365 Copilot, Copilot Studio, and Foundry Agent Service. Know which platform "owns" which kind of agent.
• ROI analysis — total cost of ownership math, including infrastructure, model inference costs, and human-in-the-loop oversight time.
• Microsoft AI Center of Excellence (CoE) model — its purpose, structure, and the specific deliverables a CoE produces. This shows up in scenario questions where you're asked who should sign off on what.
• Prompt library guidelines and prompt engineering techniques — at the architectural level, not as a developer.

Common mistake: Candidates from a pure data-science background over-index on model selection questions and under-prepare for governance topics. The exam rewards governance fluency over model fluency.

Design AI-powered business solutions (25–30%)

This is the "how the pieces fit together" domain. Heavy emphasis on Copilot Studio extensibility, Foundry custom models, and integration patterns.

Critical sub-topics:

• Copilot Studio agent design — topics, fallback, voice mode, reasoning, agent flows. Know when to use natural language processing vs. conversational language understanding vs. generative AI orchestration.
• Model Context Protocol (MCP) integration in Copilot Studio — new in the July 22, 2026 skills update. Expect at least one question on when MCP is the right choice over a custom connector.
• Agent extensibility via Model Context Protocol — the architectural patterns, not the protocol details.
• Computer Use in Copilot Studio — when to use Computer Use to automate browser/app interactions vs. a connector-based approach.
• Custom AI models in Microsoft Foundry — including the new Agent Service, prompt flows, and model router patterns.
• Power Platform Well-Architected Framework — applied to intelligent application workloads. Five pillars: reliability, security, operational excellence, performance efficiency, experience optimization.
• Cross-app design for Dynamics 365 — Sales, Service, Field Service, Customer Service, Finance, and Supply Chain. You don't need encyclopedic depth, but you do need to know which Copilot lives where and what it talks to.

Deploy AI-powered business solutions (40–45%)

This is the largest domain. Microsoft is explicit: shipping, monitoring, and governing matters more than designing. Be ready.

Critical sub-topics:

• Application Lifecycle Management (ALM) — for Copilot Studio agents, Foundry Agents service, custom AI models, and data used in AI models. Each has its own ALM nuances; expect questions that test whether you know the right ALM pattern for each surface.
• Testing strategy — validation criteria for custom AI models, end-to-end test scenarios across multiple Dynamics 365 apps, Copilot prompt best practices, and using Copilot itself to generate test cases.
• Monitoring and telemetry — recommending tools and processes for agent monitoring, analyzing backlog and user feedback, AI-based tools for tuning, performance metrics interpretation.
• Responsible AI — designing for the Microsoft Responsible AI guidelines, including security, governance, audit trails, prompt manipulation defense, data residency, and access controls on grounding data.
• Vulnerability analysis — analyzing solutions for AI-specific vulnerabilities and proposing mitigations. Prompt injection is the headline topic but not the only one — also expect questions on data poisoning, model extraction, and model tuning compromise.

Question types you'll actually encounter

Microsoft uses nine recognized question types. AB-100 uses six of them. Knowing each format in advance saves you time on exam day — you don't want to be reading the navigation instructions instead of the question.

Each AB-100 exam contains at least one case study. The case study introduction screen tells you how many to expect on your specific form of the exam. Treat each case study as a contained workload — read the scenario, the requirements, and the technical constraints completely before opening the first question.

The July 22, 2026 skills refresh

Microsoft refreshed AB-100's skills measured on July 22, 2026. Most older prep content was written against the pre-July version. The biggest additions:

• Microsoft Foundry rename — what was previously called Azure AI Foundry. Same product family, but exam questions now use the new branding. Prep content referencing only "Azure AI Foundry" needs to be cross-referenced against Microsoft Foundry's updated docs.
• Agent2Agent (A2A) protocol — a new open standard for inter-agent communication. Expect at least one design-domain question on when A2A is the right choice vs. a custom orchestration layer.
• Model Context Protocol (MCP) in Copilot Studio — MCP is now a first-class extensibility option. The exam tests architectural decisions, not protocol details — you need to know when to recommend it.
• Computer Use in Copilot Studio — the capability to drive UIs autonomously. Tested at the design level (when to recommend it, what the security implications are).
• Microsoft Responsible AI Standard — re-emphasized across multiple sub-domains. Read the current version of the Standard, not the 2024 version.

If your prep materials don't mention A2A or MCP in Copilot Studio, they predate the July 2026 refresh. Treat them as supplementary, not authoritative.

A pragmatic 6-week study plan

This assumes you already hold one of the prerequisite Associate certs and have day-job exposure to Copilot Studio or Foundry. If you don't, double the timeline.

Weeks 1–2: Foundations + the official study guide

Read the official Microsoft Learn study guide end-to-end. For each sub-topic, flag it as one of three categories: can teach it, can recognize it, have never heard of it. The "never heard of it" list becomes your week-3-and-4 priority.

Spend at least 10 hours in hands-on Copilot Studio building a multi-topic agent with at least one fallback and one external connector. The exam will quietly assume you've done this.

Weeks 3–4: Plan + Design domains

Work through the Plan domain (25–30%) and the Design domain (25–30%) systematically. For each skill bullet:

• Read the official Microsoft documentation page
• Find one real-world product example
• Write a one-paragraph summary of when you'd use it

The third step is the secret weapon — being able to articulate "when" rather than "what" is the difference between recognizing exam answers and reasoning about them.

Week 5: Deploy domain (the heavy one)

The Deploy domain is 40–45% of the exam and is where most candidates lose points. Dedicate a full week to it:

• ALM patterns for each agentic surface (Copilot Studio, Foundry Agents, custom models, Dynamics 365)
• Microsoft Purview integration for AI data governance
• Responsible AI Standard — specifically the requirements, not just the principles
• Prompt injection mitigation patterns
• Telemetry interpretation and tuning loops

Week 6: Practice tests and exam-day prep

Use practice tests aggressively. Aim for 80%+ on practice exams before you sit the real one. If a practice question surfaces a topic you can't explain in two sentences, return to the study guide for that sub-topic before continuing.

Two days before the exam, schedule a final read-through of the official study guide. Don't try to learn new material at this point — you're consolidating, not expanding.

Exam-day pitfalls (and how to avoid them)

• Spending too long on the first case study. Allocate roughly 30 minutes per case study, not more. Time you spend in a case study you can't get back, and you can't return to questions after leaving.
• Skipping the exam introduction screen. It tells you how many questions, how many case studies, whether labs are present, and the time you have. Read it.
• Over-reading drag-and-drop questions. If the correct ordering is "obvious from the Microsoft framework you studied", it's the right answer. Microsoft rarely uses drag-and-drop for trick questions.
• Defaulting to "custom model" answers. Microsoft tests whether you choose the simplest solution that works. "Use a prebuilt Copilot in Dynamics 365" beats "build a custom Foundry agent" almost every time, unless the scenario explicitly rules out the prebuilt path.
• Ignoring data residency clues. If the scenario mentions an EU-based customer, regulated industry, or specific compliance framework, expect at least one question on data residency or sovereignty. Answer choices that violate residency requirements are wrong even if they're otherwise correct.

Sample question patterns by domain

We can't reproduce real exam items here, but these are the patterns you'll see, paraphrased:

Plan-domain pattern

"Contoso wants to automate three customer support workflows. The IT team has two senior developers and limited Foundry experience. The Chief Risk Officer requires SOC 2 attestation for any system handling customer data. Which architectural decision should the AI Center of Excellence recommend first?"

The right answer almost always nudges toward the lowest-risk, highest-reuse option — prebuilt Copilot extensions before custom Foundry agents. The CoE governance signal is your hint that the question is testing decision authority, not technical capability.

Design-domain pattern

"You're designing an agent in Copilot Studio that needs to invoke an internal pricing engine. The pricing engine exposes a REST API with OAuth 2.0 authentication. Which extensibility option should you recommend?"

This is a "custom connector vs. MCP vs. plugin" question. The correct answer depends on subtle hints in the scenario — frequency of change, who maintains the engine, whether multiple agents will use it. Know the decision tree.

Deploy-domain pattern

"After deploying a multi-agent solution to production, end users report inconsistent responses for the same input. Telemetry shows similar prompts produce different model outputs. What should you investigate first?"

Microsoft is testing whether you can identify the simplest plausible root cause first — usually configuration drift between environments, not a model issue. The right answer is rarely "retrain the model."

Where to learn more

If you want broader context on the certification ecosystem:

• Is AB-100 worth it? The honest ROI breakdown — for the cost-benefit math
• Microsoft Agentic AI certification path — for the full ladder from AI-900 through AB-100
• AB-100 vs. AI-102: which Microsoft AI cert should you choose? — for the architect vs. engineer decision

FAQ

How many questions are on the AB-100 exam?

Microsoft doesn't publish a fixed count. Most candidates report 40–60 questions across multiple choice, multi-response, drag-and-drop, build list, hot area, and case study items. The exam introduction screen on test day will tell you exactly how many.

Are there labs on AB-100?

Not as of the current exam version. Microsoft has historically added labs to architect-level exams over time, so this could change. The exam introduction screen will tell you definitively on the day.

How is the case study scored?

Each case study contains 4–8 questions scored individually using the same scaled scoring as the rest of the exam. Case studies aren't timed separately — you must manage the overall 120-minute clock.

Can I go back to a case study after I've moved past it?

No. Once you leave a case study, you cannot return. This is the single biggest source of regret-after-the-fact among AB-100 candidates. Read every question in a case study and finalize your answers before leaving.

What's the best way to practice case studies before exam day?

The Microsoft Learn study guide doesn't include case studies, but practice tests do. Read the entire case study before answering anything — that's the discipline that translates directly to exam day.

How long are the breaks?

You can take multiple unscheduled breaks during the exam, but the clock doesn't stop. Most candidates take one short break around the 60-minute mark. You cannot return to questions seen before a break, so finish a case study before stepping away.

If I fail, how soon can I retake?

Microsoft's retake policy requires a 24-hour wait after your first failure and progressively longer waits after subsequent failures (14 days after a second failure, 14 days after a third, and so on, up to a maximum of five attempts per 12-month period). Expect to pay the exam fee for each retake.

Will the exam be updated again before I sit it?

The skills measured were updated on July 22, 2026. Microsoft typically refreshes architect-level exams every 12–18 months. Watch the official certification page for "skills measured update" announcements; updates are posted with at least 6 weeks of notice.

The bottom line

AB-100 rewards candidates who can read a scenario and identify the simplest Microsoft-aligned answer rather than the most technically interesting one. The Deploy domain dominates — if your study time isn't 40–45% on operations, ALM, and governance, you'll feel that imbalance in the second half of the exam.

Budget six weeks of focused prep, work the official study guide as your spine, build at least one real multi-agent solution end-to-end, and use practice tests aggressively in the final week. Most candidates who fail AB-100 fail because they over-prepared for Design and under-prepared for Deploy. Don't be one of them.