How to use this post
Eight original AB-100: Agentic AI Business Solutions Architect scenario questions. Answer each one before reading its explanation. On this exam the wrong answers are usually the interesting part, so every explanation covers them too.
Most of what you'll find online about AB-100 is either a reworded copy of Microsoft's skills outline or someone claiming to sell real exam questions. The first tells you what is on the exam but not how it thinks. The second puts your certification at risk under Microsoft's exam policies.
So here are eight original scenario questions, written against the current skills outline, with explanations of why each wrong answer is wrong. They are not real exam questions.
The exam facts
| Exam | AB-100: Agentic AI Business Solutions Architect |
|---|---|
| Certification | Microsoft Certified: Agentic AI Business Solutions Architect Expert |
| Prerequisite | At least one Associate-level certification from Microsoft's list (for example PL-200, PL-400, MB-230, MB-310, MB-800, AI-103 or AB-620) |
| Passing score | 700 on Microsoft's scaled score. It isn't a percentage. |
| Question types | Multiple choice, multi-select and case studies |
| Price | Set by country; $165 in the United States |
| Renewal | Yearly, through a free online assessment on Microsoft Learn |
The weights change on 16 November 2026
Microsoft has published new domain weights that apply from 16 November 2026. Check your exam date and study to the split that applies to it:
| Domain | Until 15 November | From 16 November |
|---|---|---|
| Plan AI-powered business solutions | 25–30% | 35–40% |
| Design AI-powered business solutions | 25–30% | 30–35% |
| Deploy AI-powered business solutions | 40–45% | 20–25% |
The topics themselves barely move: the change log marks the content changes as minor. What changes is where the marks are. From 16 November, planning and design carry about two-thirds of the exam. (Source: the official AB-100 study guide.)
The eight questions below cover all three domains: two Plan, two Design and four Deploy.
The 8 questions
Question 1: Plan
A retail company wants to automate supplier-invoice exception handling. Today three analysts review flagged invoices against contract terms held in a SharePoint library, a Dynamics 365 Finance instance, and a set of PDF contracts scanned between 2019 and 2023 with inconsistent OCR quality. The CIO wants an autonomous agent to resolve exceptions end to end.
As the solution architect, what should you address first?
- A. Design a multi-agent orchestration in Copilot Studio with a supervisor agent routing to per-source retrieval agents.
- B. Assess the grounding data for accuracy, cleanliness, timeliness and availability, and scope automation to the exceptions where data quality supports it.
- C. Build a custom small language model fine-tuned on the historical invoice corpus.
- D. Implement a model router so simple exceptions go to a cheaper model and complex ones escalate.
Correct answer: B
The Plan domain explicitly calls for reviewing data for grounding (accuracy, relevance, timeliness, cleanliness, availability) before committing to an automation scope. Four years of inconsistently OCR'd PDFs is a loud data-quality signal. An autonomous agent grounded on unreliable contract text will produce confident, wrong resolutions on financial documents. Expect this pattern often: when a scenario contains a data-readiness red flag, deal with it before designing.
Why A is wrong: A plausible design, and maybe part of the final architecture, but it's a Design answer to a Plan question. Orchestrating retrieval across three sources doesn't help if one of them is unreliable.
Why C is wrong: Custom model creation is an expensive step the exam treats as something to justify, not a default. The problem here is the input data, and fine-tuning on a dirty corpus teaches the model the dirt.
Why D is wrong: Model routing is a real objective, but cost optimisation before feasibility is backwards.
Question 2: Plan
A manufacturing firm asks you to justify a proposed agentic solution to a sceptical CFO. It would automate about 60% of order-status enquiries handled by a 12-person service desk. Licensing and implementation are estimated at $340,000 in year one.
Which set of inputs best supports the ROI analysis?
- A. Projected licence savings alone, since headcount decisions are the CFO's to make.
- B. Total cost of ownership (licensing, implementation, ongoing monitoring and tuning, and model consumption) set against measurable deflection and handling-time outcomes.
- C. Benchmark ROI figures from published Microsoft customer case studies in the same industry.
- D. A pilot with 20 users, deferring the financial analysis until real usage data exists.
Correct answer: B
The outline asks you to "select ROI criteria for AI-powered business solutions, including the total cost of ownership". Agentic solutions have running costs that traditional automation doesn't: consumption that scales with usage, continuous monitoring, and periodic tuning. An ROI case that counts build cost but not run cost is the failure this objective tests.
Why A is wrong: It undercounts both sides: no running costs, and no benefits beyond licences.
Why C is wrong: Benchmarks are useful context, not evidence. The analysis needs this organisation's volumes and cost base.
Why D is wrong: Piloting is sound, but the question asks what supports the ROI analysis. In practice you'd do both: build the TCO model and validate it with a pilot.
Question 3: Design
A logistics company runs a customer-service agent in Copilot Studio. Most enquiries are handled well, but the agent must now also trigger shipment re-bookings in a third-party carrier system that has a documented API and a published MCP server.
What's the most appropriate design?
- A. Rebuild the agent in Microsoft Foundry, since Copilot Studio can't reach external systems.
- B. Extend the Copilot Studio agent with Model Context Protocol to connect the carrier's MCP server, keeping the existing conversational agent.
- C. Add a topic with a fallback telling the user to complete the re-booking manually in the carrier portal.
- D. Use Computer Use in Copilot Studio to drive the carrier's web portal.
Correct answer: B
"Design agent extensibility with Model Context Protocol in Copilot Studio" is a named objective, and an external system that already exposes an MCP server is its textbook trigger. MCP gives a structured, governable integration and keeps the working agent.
Why A is wrong: The premise is false: Copilot Studio agents connect to external systems routinely. Spotting a false capability claim is faster than reasoning about the rest of the option.
Why C is wrong: It avoids the requirement instead of meeting it.
Why D is wrong: Computer Use is on the outline and genuinely useful, for systems with no API. With an API and an MCP server available, driving the UI is the fragile choice. API or MCP available: integrate. No programmatic surface: Computer Use.
Question 4: Design
An insurance company wants an agent that monitors a claims queue overnight, gathers supporting documents, applies eligibility rules, and either settles low-value claims or escalates them. No employee is present during the run.
Which agent type should you design, and what must the design include?
- A. A prompt-and-response agent with retrieval over the claims knowledge base.
- B. A task agent, invoked by the claims handler at the start of each shift.
- C. An autonomous agent with explicit trigger conditions, defined action boundaries, human escalation paths and audit trails for its decisions.
- D. A Microsoft 365 Copilot declarative agent surfaced in Teams for the claims team.
Correct answer: C
Unattended, event-driven, multi-step action is the definition of an autonomous agent, and "design autonomous agents" is its own objective. But the answer isn't only the agent type, it's the guardrails. On AB-100, autonomy and controls come as a package; an option offering autonomy without controls is a trap.
Why A is wrong: Prompt-and-response agents answer on request. They don't monitor queues or act on triggers.
Why B is wrong: Invoking it at the start of a shift breaks the core requirement: the work happens overnight, unattended.
Why D is wrong: This is a human-in-the-loop assistant in Teams, the opposite of what was asked. Nobody is in Teams at 3 a.m.
Question 5: Deploy
Six weeks after go-live, a Copilot Studio agent that answers HR policy questions shows a rising rate of escalations to human agents, though user satisfaction scores are unchanged. Leadership wants to know what's wrong.
What should you do first?
- A. Expand the knowledge sources, since escalations imply missing content.
- B. Interpret the telemetry (escalation triggers, topic-level resolution rates and the enquiries preceding escalation) to find where the agent is failing.
- C. Switch to a more capable model to improve reasoning quality.
- D. Raise the confidence threshold so the agent answers more questions before escalating.
Correct answer: B
"Interpret telemetry data for performance and model tuning" is an explicit objective, and the domain is built on diagnosing before acting. Rising escalations have many possible causes: a content gap, a routing error, a policy the agent doesn't know, or a change in what users ask. Telemetry tells you which. Stable satisfaction scores are a clue too: they point away from answer quality and towards coverage or routing.
Why A is wrong: Maybe the most likely cause, but still a guess. Adding unconfirmed content is how knowledge bases bloat and retrieval quality drops.
Why C is wrong: Nothing indicates a reasoning failure, and a model change risks regressions in every topic that works today.
Why D is wrong: The dangerous option. The escalation metric improves while the real failure rate gets worse: you've silenced the alarm instead of fixing the fault.
Question 6: Deploy
A financial services client is deploying an agent that reads customer correspondence and drafts replies. Security review flags a concern: a customer could embed instructions in an email to make the agent disclose other customers' information.
Which mitigation set best addresses this?
- A. Restrict the agent to a read-only connection to the customer database.
- B. Treat all correspondence as untrusted input, apply prompt-injection defences, and enforce access controls on grounding data so the agent can only retrieve records for the customer in context.
- C. Add a disclaimer to all agent-drafted replies stating that they were AI-generated.
- D. Log all agent interactions for later audit.
Correct answer: B
The outline names "analyze solution and AI vulnerabilities and mitigations, including prompt manipulation" and "design access controls on grounding data". The expected answer is layered: input-side defences and a data boundary scoped to the customer in context. Even a successful injection can't leak what the agent was never allowed to retrieve.
Why A is wrong: Read-only stops modification. The threat is disclosure, and read access is exactly what the attacker wants.
Why C is wrong: Disclosing AI generation is a real responsible-AI practice, but it doesn't stop manipulation.
Why D is wrong: Logging is detective, not preventive. It belongs alongside B, not instead of it.
Question 7: Deploy
An organisation has Copilot Studio agents, agent flows, connectors and a custom model in Microsoft Foundry, all built directly in production by a fast-moving team. It now needs a repeatable release process across development, test and production.
What should the ALM strategy include?
- A. Export and re-import agents manually at each release, with a change log kept by the team lead.
- B. A solution-based ALM process for the Copilot Studio agents, connectors and actions, plus a defined ALM process for the custom model, including versioning of the model and the data used to train and ground it.
- C. Rebuild each agent in the target environment so the configuration matches.
- D. Keep production as the only environment and use scheduled maintenance windows for changes.
Correct answer: B
The outline asks for ALM designs for Copilot Studio agents, connectors and actions, for Microsoft Foundry Agents Service, for custom AI models, and, separately, "design the ALM process for data used in AI models and agents". That last item is what makes AI ALM different. Without versioned training and grounding data you can't reproduce a model's behaviour or explain a regression.
Why A is wrong: Manual export and a hand-kept log isn't a repeatable process, and it gives no reliable rollback.
Why C is wrong: Manual rebuilding maximises configuration drift, the opposite of the goal.
Why D is wrong: It refuses the requirement and sends every change to customers untested.
Question 8: Deploy
A healthcare provider in the EU plans to deploy an agent grounded on patient-care documentation, using a model hosted in a US region because of capacity constraints.
What must the architect do?
- A. Proceed, applying encryption in transit and at rest.
- B. Validate data residency and cross-border movement compliance before deployment, and keep grounding data and inference within the permitted region, escalating the capacity constraint rather than accepting a compliance exception.
- C. Anonymise the patient data before it reaches the model, which permits the US-hosted deployment.
- D. Obtain written sign-off from the provider's CIO accepting the residency risk.
Correct answer: B
"Validate data residency and movement compliance" is a named objective. Regulatory constraints are boundaries on the design, not something to trade against capacity. The architect makes the constraint visible, designs within it, and pushes the capacity problem back to where it can be solved.
Why A is wrong: Encryption protects against unauthorised access. Residency rules govern where data may be stored and processed. Encrypted data in the wrong jurisdiction is still in the wrong jurisdiction.
Why C is wrong: The most tempting distractor. Anonymisation isn't a switch: rich clinical text carries re-identification risk, and the result may still count as personal data.
Why D is wrong: Sign-off transfers accountability. It doesn't make the processing lawful.
What these questions tell you about the exam
- Diagnose before you act. Telemetry before tuning, data assessment before design.
- Extend before you rebuild. Working solutions get extended through MCP, connectors and Copilot Studio extensibility.
- Autonomy comes with controls. An option that gives an agent independence without boundaries, escalation and audit trails is incomplete.
- Constraints are boundaries, not trade-offs. Residency, responsible AI and access control aren't negotiable against cost.
- Beware the option that improves a metric by making the outcome worse.
- AI ALM includes the data.
A study plan for the new weights
If you sit the exam on or after 16 November, spread about four weeks roughly like this:
Weeks 1–2: Plan (35–40%). Grounding-data assessment, the AI adoption process in the Cloud Adoption Framework, build vs. buy vs. extend, when to create custom models or small language models, model routers, prompt libraries and prompt-engineering guidance, the Microsoft AI Center of Excellence, multi-agent design across Microsoft 365 Copilot, Copilot Studio and Microsoft Foundry, and ROI with total cost of ownership.
Week 3: Design (30–35%). Task, autonomous and prompt-and-response agents and when each applies; Copilot Studio topics, fallback, agent flows, prompt actions, reasoning and voice mode; extensibility through MCP and Computer Use, and the line between them; Copilot in Dynamics 365; Microsoft Foundry Tools; and the Power Platform Well-Architected Framework.
Week 4: Deploy (20–25%) and case studies. Monitoring and telemetry, testing, ALM (including the data), security, governance, prompt-manipulation defences, data residency and audit trails. Spend the second half of the week on case studies: read the requirements before the questions, and note every constraint, because the answer usually sits in one you skimmed.
Sitting the exam before 16 November? Swap the order: Deploy is 40–45% until then. Either way, try Microsoft's exam sandbox once so the interface doesn't cost you minutes.
Try 20 more, free
→ Take the free AB-100 practice test: 20 questions, no signup, and your score as soon as you submit. Give your email at the end to see every answer explained.
Full AB-100 practice tests
599 questions across 6 full-length practice tests, each with an explanation for every answer. See the AB-100 course →
This article is not affiliated with or endorsed by Microsoft. All questions above are original, written from Microsoft's published AB-100 skills outline.