azure-ai

---
name: azure-ai
description: "Use when standing up or operating Azure AI services: choosing between Azure OpenAI, AI Foundry, AI Search, and Cognitive Services, then deploying, securing, and observing a chosen path."
source: https://github.com/microsoft/azure-skills/tree/main/skills/azure-ai
generated: 2026-05-25T02:43:51.821Z
category: platform
audience: devops
---

## When to use

- Picking between Azure OpenAI, AI Foundry, and AI Search for a new feature instead of reaching for the wrong one
- Provisioning models, deployments, and capacity in Azure OpenAI without tripping over quota and region constraints
- Wiring private networking, managed identity, and Key Vault into an AI service so it is not internet-public
- Adding telemetry so token usage, latency, and content-filter triggers are visible in Application Insights

## Key concepts

### Service selection

Azure OpenAI exposes OpenAI models with Azure RBAC and networking. AI Foundry is the orchestration and evaluation surface. AI Search supplies retrieval. Cognitive Services covers vision, speech, and document intelligence.

### Deployments and capacity

Models are exposed as named deployments with provisioned (PTU) or pay-as-you-go capacity. Region and quota choices determine availability and cost more than model choice does.

### Managed identity and RBAC

Authenticate workloads to Azure OpenAI with managed identity plus role assignments instead of API keys; keys end up in source and break the rotation story.

### Private endpoints

Put the AI account on a private endpoint and disable public network access when the service handles regulated data; this changes how clients resolve the endpoint.

### Content filters and abuse monitoring

Default content filters run on inputs and outputs. Abuse monitoring stores prompts and completions for review unless modified data processing is configured.

## API reference

```
npx skills add microsoft/azure-skills --skill azure-ai
```

Install the Azure AI skill bundle.

```
npx skills add microsoft/azure-skills --skill azure-ai
```

```
az cognitiveservices account create / az cognitiveservices account deployment create
```

Provision an Azure OpenAI account and a model deployment from the CLI.

```
az cognitiveservices account create -n my-aoai -g rg --kind OpenAI --sku S0 --location eastus2
az cognitiveservices account deployment create -n my-aoai -g rg --deployment-name gpt-4o --model-name gpt-4o --model-version 2024-08-06 --model-format OpenAI --sku-capacity 10 --sku-name Standard
```

## Gotchas

- Model versions are pinned to a deployment; upgrading a model means creating a new deployment, not editing the old one
- Quota lives at the subscription and region level; the same deployment can succeed in one region and fail in another
- Disabling public network access from the portal does not retroactively close existing TCP connections
- Default abuse monitoring stores prompts for 30 days; if that is unacceptable, apply for modified abuse monitoring before going live
- PTU capacity is reserved capacity, billed even when idle; treat it like compute, not API metering

---
Generated by SkillMake from https://github.com/microsoft/azure-skills/tree/main/skills/azure-ai on 2026-05-25T02:43:51.821Z.
Verify against source before relying on details.

File: ~/.claude/skills/azure-ai/SKILL.md