The decision tables — memorise these five

This card gathers the five decisions that recur across every domain. If you are short of time before the exam, this is the one to read last.

1 · Which service for the task

What the requirement describes	The service
Deterministic, high-volume entity, personal-information or sentiment extraction	Azure AI Language — not a model
Standard text-from-image or generic object detection	Azure AI Vision
Long-video insights — speaker or celebrity identification, sentiment	Video Indexer
Custom schema-based extraction from any modality, or retrieval-ready Markdown	Content Understanding
A known structured document type with a prebuilt — invoice, receipt, identity	Document Intelligence
Grounding, indexing and the retrieval backbone	Azure AI Search — Foundry IQ sits on top of it

2 · Prompt, retrieval or fine-tune

• A little extra knowledge → prompt engineering
• A lot of knowledge, or knowledge that changes → retrieval
• Different behaviour, tone or format → fine-tuning
• Never fine-tune to add facts

3 · Deployment SKU

• Highest quota, general use → Global Standard
• Predictable throughput → Provisioned
• Cheap overnight batch → Batch, around half price, 24-hour turnaround
• EU or US residency → Data Zone
• UK-only residency → Regional, because the UK is not a data zone
• Testing a fine-tuned model → Developer

4 · Tool choice for agents

• Local code → a custom function, which you dispatch
• Offloaded or asynchronous → Azure Functions
• An existing REST API → the OpenAPI tool
• Low-code → Logic Apps
• Standardised discovery or reuse → MCP

5 · Multi-agent orchestration

• A pipeline → sequential
• Parallel independent work → concurrent
• Runtime routing to a specialist → handoff
• Collaborative, shared history → group chat
• Dynamic planning with a ledger → Magentic

And the query mode, once more

Exact terms → keyword. Meaning → vector. Best recall → hybrid. Precision re-rank → semantic ranking, which is not vector.

The trap, and the best single piece of exam technique here

Two options are often both technically valid. What separates them is a qualifier in the question: minimise cost, least administrative effort, most secure, real-time, deterministic.

Read the last sentence of the question first. It usually carries the qualifier, and the qualifier tells you which constraint is actually being scored. Two answers that both work are not a badly written question — they are a question about priorities, and the priority is stated.