Azure AI and Machine Learning Services

Scope note: The Microsoft AZ-900 skills outline dated July 20, 2026 does not list Azure AI, Copilot, or Responsible AI as a separate exam objective. This section is optional Azure context. Finish the three published AZ-900 domains first; do not infer an AI-heavy exam refresh from this page.

Prebuilt services versus custom models

Azure offers AI services for different levels of control. A prebuilt API can classify sentiment or transcribe speech without training a model. Azure Machine Learning supports teams that prepare data, train, evaluate, deploy, and monitor their own models. Microsoft Foundry is a broader platform for building and operating generative-AI applications and agents. These are useful conceptual distinctions, but the published AZ-900 outline does not assign them a weight or guarantee questions about them.

A product name alone rarely identifies the right design. Ask what input the app has, whether the team needs to train a model, what data-governance constraints apply, and who will operate the result. An API can be the simpler choice for an established task; a custom model may be necessary when the prediction depends on an organization's own labeled history. This is a general architecture judgment, not an AZ-900 scoring shortcut.

Examples of ready-made Azure AI capabilities

These services have different prices, limits, model versions, and region support. A “prebuilt” service still requires application integration, testing, privacy review, and monitoring. It does not guarantee accurate or appropriate results for every population or business process.

Older tutorials may recommend Custom Vision, Anomaly Detector, Content Moderator, or the Bot Framework SDK as default starting points. Product lifecycle changes make that unsafe. Custom Vision is on a retirement path, and the Bot Framework SDK is archived. Check Microsoft's current migration and product-status documentation before selecting any older service. Do not memorize a retired product list as if it were an AZ-900 objective.

Azure Machine Learning

Azure Machine Learning supports the lifecycle of a custom predictive model: prepare data, select an experiment, train and evaluate, register a model, deploy an endpoint, and monitor performance. Automated ML can search candidate algorithms; notebooks and pipelines support more controlled workflows. A custom model also brings responsibilities that an API does not remove: representative training data, security, versioning, drift monitoring, and retraining.

Example: A retailer wants to forecast demand from its own sales, inventory, calendar, and promotion history. Azure Machine Learning can support that data-specific training workflow. A prebuilt sentiment API would not solve the forecasting problem. Conversely, if a team merely needs sentiment from customer reviews, Azure AI Language can be a practical prebuilt starting point.

Generative AI and Microsoft Foundry

Azure OpenAI provides access to supported generative models through Azure. Microsoft Foundry is the current name for the platform previously called Azure AI Foundry; it helps teams select models, build applications or agents, evaluate responses, and operate the system. Model availability and names change, so an exam guide should not depend on a fixed GPT or image-model version. A chatbot is an application pattern that may combine a model, orchestration, retrieval, identity, and a delivery channel; a model endpoint alone is not a complete, governed application.

Enterprise controls may include Microsoft Entra authentication, role-based access, private networking, monitoring, and content safety. They must be configured and tested. Do not assume all prompts and outputs stay inside one tenant or Azure region. Processing location depends on the selected deployment type (regional, data-zone, or global), model, feature, and current service terms. A regulated workload needs a specific deployment and data-flow review, not a general promise based on the product name.

Responsible operation

Responsible AI is a governance practice rather than a single deployable product. Teams should consider fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Human review may be needed where errors affect people or legal rights. Evaluation must test realistic inputs, edge cases, and misuse, then continue after release as data and models change. These principles are valuable context but are not a newly added AZ-900 domain.

How to use this section while preparing

If you are studying strictly for AZ-900, prioritize cloud concepts, Azure architecture and services, and management and governance. Know the published objectives and the purpose of the services actually listed there. Return to this section as optional context if your work involves AI, or when planning a later AI-focused credential. When you do, verify current Microsoft documentation: names, model versions, regional availability, and retirement dates change faster than core cloud concepts.

Study boundary: Prebuilt text analysis, custom model training, and generative-app platforms solve different problems. This is a useful Azure distinction, but no published July 2026 AZ-900 weight or AI expansion supports treating it as a tested fourth domain.