Case studies
The real AB-100 exam includes case studies: a single scenario described up front, followed by several questions that test whether you can apply the concepts in context. Read the scenario, then answer the linked questions; the scenario stays on screen the whole time.
Contoso Retail runs a customer-facing e-commerce platform and operates internal teams in sales, support, finance, and operations. Leadership wants to adopt AI to raise productivity and automate business processes, and has asked the enterprise architecture team to plan the program using the Cloud Adoption Framework. Contoso already licenses Microsoft 365 Copilot for knowledge workers and uses Dynamics 365 Sales and Customer Service.
Priorities include: an internal HR and policy assistant for employees inside Teams; a high-volume customer-support chat workload where cost must be controlled without sacrificing answer quality; a sales lead-management capability across Dynamics 365 and Salesforce; and a loan-style credit decisioning agent with strict business rules and external system integrations. The team must also ensure grounding data is trustworthy, business data is reusable across AI systems, and responsible AI accountability is clearly owned. Several knowledge sources have not been updated in over a year.
Northwind Traders is modernizing customer operations. Sellers want AI assistance inside Dynamics 365 Sales to summarize opportunities, catch up on recent record changes, prepare for meetings, and draft emails, while strictly respecting each seller's record-level access. The contact center wants AI agents across voice and chat that resolve common issues and hand off to human representatives when needed.
Northwind also runs a legacy on-premises billing desktop application with no API. The finance team must automate nightly invoice data entry and extraction from this app. The platform team is standardizing on Copilot Studio for low-code agents and Microsoft Foundry for any custom-model, pro-code work, and it wants to connect a partner's existing Model Context Protocol (MCP) server that already exposes shipment-tracking tools. All solutions must pass a Power Platform Well-Architected review, and governance requires that connector data loss prevention policies extend to any new integrations.
Contoso Financial, a regulated lender operating in the EU and US, is deploying AI agents built in Microsoft Copilot Studio that ground answers on SharePoint policy libraries and Dataverse, plus a custom-tuned model hosted in Microsoft Foundry. The agents integrate with Dynamics 365 Customer Service and Finance. Makers currently build directly in the tenant's default environment, and the chief architect has been asked to establish ALM, governance, monitoring, and responsible AI controls before go-live.
Compliance requires that EU customer data not move outside the EU for generative features, that Microsoft engineer access to customer data require explicit approval, that all maker activity be auditable, and that a documented responsible AI process govern the custom model. Leadership also wants reliable production deployments and post-launch monitoring of adoption and satisfaction, with the ability to roll back quickly if an incident occurs.
Halvorsen Marine Supply is a 4,200-employee distributor of marine parts operating in nine countries. It runs Dynamics 365 Sales and Dynamics 365 Customer Service on a single Dataverse environment, and Dynamics 365 Supply Chain Management for procurement and inventory. Historical sales, margin, and shipment data lands in a Microsoft Fabric lakehouse, and the finance team publishes a certified Power BI semantic model over it. Roughly 400 head-office knowledge workers have Microsoft 365 Copilot licenses; the remaining staff do not.
The company's 61,000 service and product knowledge articles live in ServiceNow. Leadership wants those articles usable not only by a new Copilot Studio support agent but also by Microsoft 365 Copilot Chat and by any agent the company builds later, with citations back to the original article. A separate data-residency policy forbids replicating live purchase order records out of Supply Chain Management into Microsoft 365.
Three outcomes are in scope for the first release: cut quote turnaround time, stop purchasers from chasing vendors by hand, and let regional sales directors ask margin and shipment questions in plain language instead of filing report requests. An AI Center of Excellence already exists and has asked the architect to standardize how employees prompt Copilot so that quality doesn't depend on who is typing.
Brightmoor Logistics is a mid-sized freight forwarder running Dynamics 365 Supply Chain Management and Dynamics 365 Finance version 10.0.47, with a Dataverse environment in Canada that hosts several model-driven apps used by dispatchers and account managers. Vendor change requests arrive both as free-text emails to a shared mailbox and through the vendor collaboration interface, and purchasers currently open every purchase order by hand to work out whether a promised-date change will break a customer delivery.
The company has three parallel initiatives. First, the app team wants a new AI-generated dashboard page inside the dispatcher model-driven app; the developers are comfortable with TypeScript and CLI tooling but the Generative page option does not appear in their app designer. Second, an autonomous Copilot Studio agent should triage vendor emails, propose purchase-order updates, and hand exceptions to a named purchaser inside the app rather than by email — and it must never write a record that nobody has looked at. Third, a partner customs-brokerage firm operates its own agent, built on a non-Microsoft framework and hosted in the partner's cloud, whose internal tools and pricing logic Brightmoor is contractually forbidden from seeing; Brightmoor's agent must be able to hand it a whole classification task and get a structured answer back over multiple turns.
Compliance has two additional requirements. Support questions asked in the ERP sidecar must be answerable from Brightmoor's own operating procedures, which exist as Word and PDF documents, and the security team insists that no agent be able to reach ERP forms or entities beyond what its assigned role permits.
Meridian Rail Freight moves bulk cargo across three continents. Customers currently phone a service desk to ask where a consignment is and whether a delivery window has slipped. Meridian is replacing that call flow with a Copilot Studio agent that uses generative orchestration. The agent grounds itself on a SharePoint library of service commitments, calls a shipment-status action backed by Dynamics 365 Supply Chain Management, and hands off to a connected agent that raises a case in Dynamics 365 Customer Service when the customer wants a credit note. A separate Microsoft Foundry prompt-based agent, published as an agent application, is used internally by planners to summarise disruption reports.
Meridian's development team works in one development environment; a QA environment and a production environment complete the chain, and solutions are promoted with Power Platform pipelines from a host environment. The SharePoint library differs in every environment, and the shipment-status action authenticates with a different service account per environment. Legal requires that, for every release, Meridian can produce evidence two years later showing which acceptance tests passed. The risk committee is most worried about two things: that a future release quietly stops calling the shipment-status action and starts answering from stale documents, and that the internal planner agent behaves differently once published than it did in the playground.
The programme has eight weeks before go-live. There is no automated testing today. The team has licences for the Copilot Studio Kit, Azure DevOps, and Application Insights, and the Foundry project already has an Application Insights resource connected.
Larkspur Energy is a regulated utility serving customers in Germany, France, and Spain. It already operates Azure landing zones with a platform team owning connectivity and identity, and separate application landing zone subscriptions per workload. Meter telemetry sits in a third-party cloud object store, asset and maintenance records live in an on-premises ERP, and curated Power BI semantic models over both are widely trusted by the analytics community. Definitions of "Service Point" differ between the billing and field-operations teams, which has already produced two contradictory pilot chatbots.
The programme has three initiatives. First, a public outage-status agent on the customer website. Second, an internal field-operations agent used in Microsoft Teams by 900 engineers, all of whom will be assigned Microsoft 365 Copilot add-on licenses; it must answer from asset records and maintenance history. Third, a storm-response capability where the remediation path is not known in advance, agents hold tools that dispatch crews and update the outage management system, and the regulator requires a reviewable plan and full audit trail for every automated action.
Constraints: inference for any workload touching customer data must be processed inside the EU Data Boundary. Finance wants a defensible cost model before approval and has asked the architect to justify every consumption line. The AI Center of Excellence has 12 weeks to reach a decision gate and has been told not to overstate delivery timelines.
Calderon Freight is a regional haulier running 41 depots and a contact centre built on Dynamics 365 Contact Center. Roughly two-thirds of contacts arrive by phone; the rest come through web chat on the customer portal. Callers are usually drivers or dispatchers standing in noisy loading yards, often on hands-free devices. The board has funded a two-year agentic programme and appointed an architect to design it.
The functional requirements are mixed. Callers must be able to enter a six-digit consignment number on the phone keypad because speech recognition is unreliable in the yards. Before any change to a booking, the agent must verify the caller's identity and capture spoken consent to recording in a fixed sequence that auditors can replay. At the same time, general "how do I…" questions arrive in wildly varied phrasing and the business does not want to author a topic for each one. When the agent cannot resolve a call, it must reach the best-suited representative using the same assignment engine the contact centre already uses for chat, and the representative must arrive with full context.
In the back office, proof-of-delivery paperwork is scanned by dozens of customers in dozens of different layouts, and clerks retype the key fields by hand. Separately, the booking team wants a Book a collection topic that needs a postcode and a collection date; the agent already runs on generative orchestration and the team is unclear how those two topic inputs will be filled.
Meridian Grid is a regulated utility operating in two geographies. Its first production agent, built in Microsoft Copilot Studio, helps field engineers look up asset histories from SharePoint and Dataverse, raise work orders through Power Platform connectors, and escalate to a human dispatcher. A second, autonomous agent built on Microsoft Foundry Agent Service summarises outage telemetry overnight; it was provisioned with standard agent setup because the regulator requires that all conversation state remain in Meridian's own Azure subscription.
The platform team runs three Power Platform environments — Dev, QA, and Prod — with deployments driven by pipelines in Power Platform. During a recent incident, a support engineer with elevated rights fixed a broken adaptive card directly in Prod. Two releases later, the same card was corrected properly in Dev, the pipeline reported a successful deployment to Prod, and the defect was still visible to field engineers. Separately, after the agent's first month live, Prod telemetry in Application Insights showed adoption numbers roughly 30% higher than the dispatch team's own conversation counts, and the agent stopped invoking its work-order agent flows mid-month while continuing to answer questions normally.
Legal and the CISO have now joined the programme board. Legal must be able to preserve and export a named engineer's agent interactions for an ongoing dispute. The CISO wants to know exactly which connectors the published agent can reach and to require compliant devices before those connector calls succeed, and wants an answer that survives an audit rather than a screenshot of the Power Platform admin center.
How to approach case studies
Read the scenario twice and note the hard constraints. Most case-study questions hinge on matching one of those constraints to the right feature.