AI on Azure — AI Foundry, Azure OpenAI & Azure AI Services¶
What this is: a basic-to-solid tour of Microsoft Azure's AI portfolio — the GenAI platform (AI Foundry + Azure OpenAI), the custom-ML platform (Azure Machine Learning), and the pre-built AI Services shelf. See the cross-cloud equivalents to map these to AWS/GCP.
1. The Two Pillars¶
| Pillar | Service | Use it when… |
|---|---|---|
| Generative AI (managed models) | Azure AI Foundry (+ Azure OpenAI Service) | You want to build GenAI apps/agents on foundation models with enterprise governance |
| Custom ML (build your own) | Azure Machine Learning | You want to build, train, and deploy your own ML/DL models end to end |
Azure's distinctive edge: exclusive access to OpenAI's GPT-4o / o-series models, deep Entra ID (identity) integration, and tight ties to Microsoft 365 / GitHub.
2. Azure AI Foundry — The GenAI Platform¶
Azure AI Foundry (the evolution of Azure AI Studio) is the unified platform to build, evaluate, and deploy generative AI apps and agents. It wraps several pieces:
| Foundry piece | Role |
|---|---|
| Azure OpenAI Service | Hosted OpenAI models — GPT-4o, GPT-4.1, o3, o4-mini, embeddings, DALL·E — with enterprise SLAs, in-region hosting, and no-training terms |
| Model Catalog | 1,800+ models beyond OpenAI — Llama, Mistral, Cohere, and open models — deployable as managed or serverless endpoints |
| Agent Service | Build and orchestrate tool-using agents |
| Prompt Flow | Visual authoring, chaining, and evaluation of prompt pipelines |
| Evaluation | Built-in evaluators — groundedness, relevance, coherence, plus risk & safety (violence, self-harm, jailbreak) — pre-prod and in production (details) |
| Content Safety | Filters for hate/violence/self-harm, prompt-shield (injection defense), protected-material detection |
Enterprise properties QA cares about¶
Entra ID (Azure AD) identity, private networking (Private Link / VNet), in-region data residency, customer-managed keys, and content-safety governance evidence for regulated sign-off.
3. Azure Machine Learning — The Custom-ML Platform¶
The end-to-end classic-ML platform (equivalent to AWS SageMaker / Vertex training):
| AML piece | Role |
|---|---|
| Studio / Notebooks | The workbench IDE |
| Designer | Low-code drag-drop ML pipelines |
| Automated ML (AutoML) | Auto model selection + hyperparameter tuning |
| Compute & Training | Managed clusters for distributed training |
| Endpoints | Real-time and batch model deployment |
| MLOps | Pipelines, model registry, versioning, CI/CD |
| Responsible AI dashboard | Fairness, explainability, error analysis |
Foundry vs Azure ML: Foundry = generative AI apps on managed foundation models; Azure ML = build/operate your own models. They interoperate (Foundry can consume AML-deployed models).
4. Azure AI Services (Pre-Built — formerly "Cognitive Services")¶
Call an API, no training:
| Service | Does |
|---|---|
| Azure AI Document Intelligence | OCR / structured extraction from forms & docs |
| Azure AI Vision | Image analysis, OCR, spatial analysis |
| Azure AI Speech | Speech-to-text, text-to-speech, translation |
| Azure AI Language | NLP — entities, sentiment, PII, summarization, Q&A |
| Azure AI Translator | Machine translation |
| Azure AI Search | Enterprise search + vector search (the RAG retrieval layer) |
| Azure AI Content Safety | Standalone moderation for text & images |
Assistants: Microsoft 365 Copilot (productivity) and Copilot Studio (build custom copilots/agents low-code).
5. A Typical Azure GenAI Architecture¶
flowchart LR
U["App / user"] --> ENTRA["Entra ID auth"]
ENTRA --> AOAI["Azure OpenAI<br/>(GPT-4o / o-series)"]
DOCS["Blob docs"] --> SRCH["Azure AI Search<br/>(chunk + vector index)"]
AOAI <--> SRCH
AOAI --> CS["AI Content Safety<br/>(prompt shield + filters)"]
AOAI --> EVAL["Foundry Evaluation (CI gate)"]
This is the classic "Azure OpenAI + Azure AI Search" RAG pattern — the most common enterprise design on Azure.
6. QA / Testing Focus on Azure¶
- Azure OpenAI calls — quality (Foundry evaluators), latency, token cost, content-safety enforcement.
- AI Search RAG — retrieval quality, hybrid (keyword+vector) tuning, RAG metrics.
- Agents / Copilot Studio — tool-call correctness, task adherence (AI Test Strategy).
- Entra ID & residency — RBAC on retrieval, region-boundary tests (AI Rollout Part 1).
- Safety — jailbreak/prompt-shield evaluators; prompt injection guide.
Where to Go Next¶
- AI on AWS · AI on GCP — the equivalents
- Cloud AI Platforms Overview — the cross-cloud mapping table
- Enterprise LLM Platforms — deeper platform comparison
- Cloud LLM Evaluation Tools — Azure AI Foundry evaluation in depth