Case studies
The real AI-103 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 Field Services is building an internal support agent on Microsoft Foundry that answers technician questions grounded in thousands of internal manuals stored in Azure Blob Storage and SharePoint. Different technicians have different access rights, and the company must ensure the agent only returns content each user is authorized to see, honoring existing ACLs and Microsoft Purview sensitivity labels.
Security policy forbids storing API keys in application configuration and requires that no inbound traffic to the Foundry hub traverse the public internet. The team must keep per-token inference cost reasonable for steady, predictable daytime traffic, monitor the agent for drift and harmful outputs in production, and ship agent definitions through their existing CI/CD pipeline with version control and code review. Leadership also wants adversarial testing before go-live.
Contoso Travel is building a customer-support agent on Microsoft Foundry Agent Service. The agent must answer questions grounded in Contoso's internal policy PDFs, look up live flight availability through an existing partner REST API described by an OpenAPI 3.1 specification, and provide current public-web information with citations. The team wants a fully managed runtime with no containers to patch and the ability to swap models without changing code.
Sensitive actions such as issuing refunds must require human approval before execution. Before go-live, the team must evaluate the agent for fabrications and correct tool usage, and in production they need distributed tracing, token and latency dashboards, and alerts when outputs are unsafe or low-quality. Security requires that each agent have its own scoped identity and that downstream calls run with the signed-in user's permissions.
Lumora Media runs a stock-content marketplace. They are building an Azure pipeline that (1) generates new promotional images and short clips from text prompts, (2) catalogs an existing library of photos with searchable descriptions and accessibility alt-text, and (3) ingests user-submitted videos to extract scene-level metadata. All AI runs through a Microsoft Foundry project.
Compliance requires that generated and ingested content be screened for unsafe material, and that the team be able to trace extracted metadata back to its source for audit. Editors also need to replace specific regions of generated images (for example, swapping a background) without regenerating the whole image. The marketplace search must let customers find images using plain-language text queries even when images have no tags.
Helios Travel is building a voice-enabled support agent in Microsoft Foundry for international travelers. Callers phone in speaking many different languages, sometimes switching languages mid-call, and the agent must respond by voice in the caller's language. Helios also wants live, searchable transcripts of every call for quality review.
Compliance requires that personal identifiers (names, phone numbers, passport numbers) be removed from stored chat transcripts before analytics, and that recurring complaint themes be tracked by the specific aspect customers mention (for example, 'baggage handling' vs. 'check-in'). The team also maintains a short list of Helios-specific product and destination names that the base recognizer often mishears, but they want to avoid training a full custom model at first.
Northwind Legal is building an internal assistant that answers attorneys' questions over a large archive of contracts, scanned correspondence, and invoices. Many older documents are scanned image PDFs containing both printed and handwritten text, and some include tables and signature checkboxes. The assistant is implemented as a Foundry agent that must ground its answers in the firm's documents and cite the exact source location for compliance.
The engineering team uses Azure AI Search for retrieval and Azure document services for extraction. Source files land in Azure Blob Storage. The team wants automatic ingestion, OCR for scanned content, chunking, and embedding generation inside the indexer pipeline. Queries mix exact clause/case numbers with natural-language questions, and reviewers must be able to trace any extracted field back to where it appears in the original document.
Harborline Logistics is a freight forwarder headquartered in Rotterdam. Its legal team has ruled that customer shipping instructions may be processed only within the European Union, and a signed data protection assessment prohibits inference traffic leaving the EU Data Boundary. Harborline is building a Microsoft Foundry agent that answers booking questions for its European desk. Between 07:00 and 19:00 CET the agent must hold a predictable p95 response time under a steady, high call volume; outside those hours traffic is negligible.
The agent is grounded on an Azure AI Search index built from 12 million documents in Azure Blob Storage — bills of lading, tariff sheets, and customs paperwork. A large share of the older customs paperwork is scanned paper stored as image-only PDFs. A nightly indexer refreshes the index while the daytime query load is still tailing off, and the platform team has noticed that the indexer reports Success even on nights when hundreds of documents produce messages. Separately, a one-off project needs to reclassify all 12 million archived documents against a new tariff taxonomy; the results are not needed for two days and the CFO has asked for the cheapest option that still respects the EU processing rule.
Two access requirements are outstanding. The Rotterdam operations desk — about 200 staff plus a service principal used by the booking portal — must be able to call the agent's endpoint but must never be able to change the agent, deploy models, or see other agents in the project. And the agent exposes an MCP tool that can amend a live booking; Harborline's risk committee requires that a named human reviews and approves every amendment before it is submitted, with an auditable record.
Meridian Freight is building a shipment exceptions desk on Microsoft Foundry. A triage prompt agent receives customer messages and delegates to three specialist prompt agents: customs, claims, and tracking. All agents live in a single Foundry project and are called from a Python service that constructs an AIProjectClient with DefaultAzureCredential and obtains an OpenAI client from the project client.
Business requirements are firm. A customer's conversation must stay in context across a multi-day exception case, not just within a single web session, without the service re-uploading transcripts. Every resolution the desk produces must be handed to the ERP system as a machine-readable object with a fixed set of fields, because the ERP import job fails whenever a field is missing. Meridian's compliance team requires that inferencing traffic be processed only within the EU data zone, and finance has capped spend: roughly 70 percent of incoming questions are trivial status lookups while the rest need genuine reasoning, and the team does not want to pay reasoning-model rates for the simple ones.
Operationally, the support console must show the agent's answer appearing progressively rather than after a long pause, because agents-in-training complain about the perceived latency. Traffic is bursty — quiet overnight, heavy on Monday mornings — and the team wants to avoid buying reserved capacity in the first release.
Halden Renewables operates onshore wind farms and inspects turbine blades with drones. Each inspection produces several hundred high-resolution JPEG stills and one continuous flight recording, typically a 45-minute MP4 of around 900 MB, which the field crews upload into Azure Blob Storage at the end of a shift. All AI workloads run in a single Microsoft Foundry project, and the platform team writes Python.
The engineering portal must publish an inspection report for every blade. Accessibility is a contractual requirement: each still needs a one-sentence alt-text string, and the reviewers have asked that captions not assert the gender of any technician who appears in frame. Alongside the caption, the inspection record must carry a structured defect classification drawn from Halden's fixed taxonomy of Crack, Erosion, Lightning strike and None, plus a short free-text note describing what the reviewer should look at.
Two other workstreams touch the same platform. Marketing generates promotional imagery from text prompts and must never publish an asset containing a competitor's logo or an unapproved brand mark. Separately, field technicians paste screenshots of vendor emails and maintenance PDFs into an assistant that summarises them and files work orders; security has flagged that those screenshots are untrusted third-party content.
Kestrel Utilities runs a 400-seat contact centre serving customers in English, French, and Spanish. Every call is recorded as a stereo WAV file with the agent on one channel and the customer on the other, and web chats are captured as multi-turn transcripts. Today a quality team listens to a 2% sample by hand. Leadership wants automated coverage of 100% of interactions within one quarter, built on Microsoft Foundry with Python.
The analytics platform team has scoped four workstreams. First, the archive of roughly 90,000 historical recordings must be transcribed in bulk overnight; separately, supervisors want to paste a single call recording into an internal tool and get the transcript back immediately, faster than real time. Second, each chat transcript must be reduced to a segmented set of topic headings plus a short summary per segment so supervisors can skim a 40-turn chat in seconds. Third, free-text comments arriving from the post-call survey have no reliable language field, and several of them contain words spelled identically in English and French.
Fourth, the regulator requires that a structured record be emitted for every escalated interaction, with fixed fields (account reference, issue category, commitment made, follow-up date) that must be present and correctly typed on every single record — a downstream ingestion job hard-fails on a missing property. The team also plans a field-technician mobile app for remote sites where cellular coverage is unreliable, and has begun the Microsoft limited-access review for on-device speech.
Kestrel Marine Logistics moves refrigerated cargo for about 200 corporate customers. Every voyage generates a bundle of PDFs — bills of lading, customs declarations, inspection reports, and rate sheets — that operations staff currently hunt through by hand. The company is building a Python agent on Microsoft Foundry that answers questions such as "what temperature deviation was recorded on voyage KM-4471?" and must cite the exact document it used.
Source PDFs land in an Azure Blob Storage container, partitioned by vessel and voyage. Each blob carries metadata for vessel name, voyage number, and the owning customer's Microsoft Entra group IDs. An Azure AI Search indexer with a skillset chunks the extracted text, generates embeddings, and projects the chunks into a single index. Answers must always show the vessel and voyage alongside the retrieved passage, and a user must never see a chunk belonging to another customer.
Two operational problems have surfaced since the pilot. Documents that operations deletes from the container still appear in answers weeks later, and queries about pricing frequently surface a superseded rate sheet instead of the current one. The team also needs a reliable way to tell whether a nightly run genuinely processed new paperwork.
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.