7.1 Service Comparison and Selection Guide
Key Takeaways
- Service selection is the single most-tested skill: the "Plan and manage" domain (20-25%) opens with "Select the appropriate Microsoft Foundry Services" for vision, NLP, speech, information extraction, and knowledge mining.
- Azure AI Language handles pre-built deterministic NLP (NER, sentiment, key phrases, PII); Azure OpenAI handles generative text (open summarization, drafting, chat) — purpose-built is cheaper and faster than a generative prompt.
- Azure AI Vision Read does general OCR on photos/signs; Azure AI Document Intelligence extracts structured key-value fields from forms; Content Understanding now extracts from documents, images, video, and audio.
- CLU does structured intent + entity recognition; Custom Question Answering returns curated FAQ answers; Azure OpenAI grounded with Azure AI Search (RAG) handles open-ended enterprise Q&A.
- Deprecated names are exam traps: LUIS to CLU, QnA Maker to Custom Question Answering, Form Recognizer to Document Intelligence, Cognitive Search to AI Search.
Quick Answer: Roughly one in five AI-102 questions is "which service?". Memorize the boundaries: Language for deterministic NLP, Azure OpenAI for generative text, Vision Read for general OCR, Document Intelligence for structured fields, Custom Vision for custom image classes, Face for identity, CLU for intents, Custom Question Answering for FAQs, and AI Search for retrieval.
Why "Select the Appropriate Service" Dominates
The December 2025 skills outline lists six service-selection bullets in the Plan and manage an Azure AI solution domain alone (20-25% of the exam). Microsoft writes these items as business scenarios, not API trivia. The trap is that Azure OpenAI can technically do almost anything — so it appears as a tempting distractor. The right answer is usually the purpose-built service because it is cheaper, deterministic, lower-latency, and does not need prompt engineering or grounding.
Text Processing Decision Matrix
| Scenario | Correct Service | Why NOT the alternative |
|---|---|---|
| Sentiment of customer reviews | Azure AI Language (sentiment) | OpenAI works but Language is cheaper, faster, deterministic |
| Open-ended summary of a long brief | Azure OpenAI (chat completions) | Language abstractive summarization is extractive/limited |
| Detect people, places, orgs | Azure AI Language (NER) | OpenAI output is non-structured and varies run to run |
| Extract policy or case IDs | Azure AI Language custom NER | Trained model is consistent; a prompt drifts |
| Route support tickets to categories | Azure AI Language custom text classification | Trained on your labels, unlike a generic prompt |
| Generate marketing copy | Azure OpenAI | Language cannot author creative text |
| Translate text/documents | Azure AI Translator | Language has no translation feature |
| Detect and redact PII | Azure AI Language (PII) | Deterministic, comprehensive, returns offsets |
| Understand "Book a flight to Paris" | CLU | OpenAI is generative, not structured intent + entity |
| Answer questions from an FAQ KB | Custom Question Answering | Returns curated answers; CLU has no Q&A pairs |
Image and Document Decision Matrix
| Scenario | Correct Service | Why NOT the alternative |
|---|---|---|
| Read text off a street-sign photo | Azure AI Vision Read (OCR) | Document Intelligence is for structured forms |
| Pull vendor, total, line items from invoices | Document Intelligence prebuilt-invoice | Raw OCR has no field understanding |
| Classify product photos into custom classes | Azure AI Custom Vision | Vision tags are generic, not your categories |
| Verify a person's identity (1:1) | Azure AI Face | Vision detects people but cannot recognize |
| Count/track shoppers on camera | Azure AI Vision Spatial Analysis | Face does recognition, not movement tracking |
| Insights from a video file | Azure AI Video Indexer | Vision processes still images |
| Extract across docs, images, audio, video | Content Understanding (Foundry Tools) | New multimodal extraction service |
| Generate an image from text | Azure OpenAI (DALL-E) | No other Azure service generates images |
Knowledge and Generative Decision Matrix
- Full-text + vector search over enterprise content -> Azure AI Search (with skillsets for OCR/NER/embeddings during indexing).
- Direct curated answers from FAQs -> Custom Question Answering.
- Open-ended chat grounded in your data -> Azure OpenAI + Azure AI Search (RAG) — Search retrieves, OpenAI generates the grounded answer with citations.
- High-volume forms at scale -> Document Intelligence (prebuilt or composed model).
Deprecated Services and Their Replacements
| Deprecated | Current name | Migration |
|---|---|---|
| LUIS | CLU (Conversational Language Understanding) | Export LUIS app, import into CLU |
| QnA Maker | Custom Question Answering | Recreate KB as a Q&A project |
| Form Recognizer | Document Intelligence | Rename only; same models |
| Cognitive Search | Azure AI Search | Rename only; same API |
| Cognitive Services | Azure AI Services / Microsoft Foundry | Multi-service resource |
| Face emotion / age / gender | Retired | Removed for Responsible AI |
On the Exam: If LUIS, QnA Maker, or Form Recognizer appear as options, they are almost always wrong unless the stem asks specifically about migration. Watch the new Foundry branding: "Microsoft Foundry Services", "Azure Vision in Foundry Tools", and "Content Understanding" are the current 2025-2026 names you must recognize.
Worked Example: Reading the Scenario for Disqualifiers
Microsoft writes selection questions so that the disqualifier is a single phrase. Train yourself to find it. Consider: "A retailer receives thousands of supplier invoices in different layouts and must extract totals and line items with the least development effort." The phrase "least development effort" rules out a custom-trained model and rules out an OpenAI prompt you would have to engineer and validate. The phrase "line items" rules out plain OCR, which returns unstructured text. That leaves Document Intelligence prebuilt-invoice — a model that ships ready to recognize invoice fields across layouts.
Now change one word: "...extract a non-standard field that appears only on this retailer's contracts." The phrase "non-standard field" now eliminates the prebuilt model and points to a custom extraction model you train on your own labeled samples. Single-word swaps like this flip the answer, so read every qualifier before scanning the options.
Cost and Latency: Why Purpose-Built Wins
When two services can both produce a result, the exam reward goes to the cheaper, faster, deterministic option. Pre-built Language operations are billed per 1,000 text records and return structured JSON with confidence scores in a single synchronous call. A generative model billed per token must read a system prompt plus your instructions plus the input on every request, can vary its output between runs, and may need a second groundedness or validation pass.
For high-volume, repeatable extraction or classification, Language or Document Intelligence is the defensible answer; reserve Azure OpenAI for tasks that genuinely require open-ended language generation, reasoning over free text, or multimodal understanding that no pre-built model exposes.
Speech, Translation, and Multimodal Edge Cases
A handful of scenarios trip candidates because the obvious service is wrong. Translating a Word or PDF document while preserving formatting is Translator's document translation feature, not Language and not OpenAI. Real-time captioning of a live meeting is Speech (speech-to-text), not Video Indexer, which works on recorded files. Identifying who is speaking across a recording is Speech speaker diarization. Extracting fields from a scanned form, a photo, and an audio note in one ingestion is Content Understanding, the new multimodal extractor in Foundry Tools.
Memorize these so the distractors that pair a plausible-but-wrong service with the right verb do not catch you.
A company needs to extract vendor names, invoice numbers, and total amounts from thousands of scanned invoices. Which approach should they use?
A travel app must understand the utterance "Book a flight to Paris" and extract the destination as a structured slot. Which service fits best?
Which deprecated service must be migrated to Custom Question Answering?