Azure Machine Learning
Azure Machine Learning is Azure's managed service for training, deploying and operating custom machine learning models. It covers notebooks, AutoML, distributed GPU training, pipelines, MLflow tracking, model registries and managed online and batch endpoints, with hybrid options through Azure Arc.1
- Intended for
- Data scientists, Machine learning engineers, MLOps and platform engineers, Application developers integrating models
- Runs on
- Azure public cloud regions, Azure Arc-enabled Kubernetes clusters for hybrid training and inferencing1
At a glance
- What is it?
- Azure Machine Learning is a cloud service for the machine learning project lifecycle: preparing data, training models, deploying them and running MLOps. Work is organized in workspaces that hold experiments, compute, data, models and connections. It accepts models built in open-source frameworks such as PyTorch, TensorFlow, scikit-learn, XGBoost and LightGBM, and also supports R and .NET. Microsoft lists it as one of the services integrated in Microsoft Foundry. Its architecture guide positions it for custom training, production MLOps, managed endpoints and hybrid compute, while Foundry is the primary platform for generative AI and agents.123
- What does it do?
- Data scientists write code in managed Jupyter notebooks or VS Code, build pipelines in the drag-and-drop designer, or run AutoML for classification, regression, forecasting, vision and language tasks. Compute clusters and serverless compute run distributed training with PyTorch, TensorFlow or MPI, including GPU options. MLflow tracking, Git integration, scheduled pipelines and CI/CD with GitHub Actions or Azure DevOps support MLOps. Models deploy to managed online endpoints, with traffic splitting between versions, or to batch endpoints. Responsible AI dashboards cover fairness, explainability and error analysis.1
- Who needs it?
- Data science and ML engineering teams that build their own models, whether classical machine learning, deep learning or fine-tuned open-source LLMs, and need an auditable path to production. Microsoft says the service is for teams implementing MLOps in a secure and auditable environment, and that platform developers can build tooling on its Azure Resource Manager APIs.1
- What does it need?4
- An Azure subscription and an Azure Machine Learning workspace
- vCPU quota for the VM families you plan to use, including GPU families for deep learning
- Python with SDK v2, R, or the Azure CLI v2
- Supporting resources such as Blob Storage, Key Vault, Container Registry and Application Insights, which are billed separately
- What it is not
- Azure Machine Learning is not the main tool for generative AI apps and agents; Microsoft's architecture guide names Foundry for that and says a workspace is needed only for training compute, MLOps or managed endpoints for custom models. It is not needed when pretrained APIs are enough (use Foundry Tools). It is not a data engineering platform; when ML is a small part of an analytics workload, Microsoft suggests Microsoft Fabric or Azure Databricks as a starting point. Its prompt flow feature retires on 20 April 2027.13
Availability and licensing. Generally available. Microsoft states there is no extra charge for the service itself: you pay for compute used for training or inference and for supporting services such as Blob Storage, Key Vault, Container Registry and Application Insights (product page, checked 26 Oct 2026). Prompt flow in Azure Machine Learning and Foundry retires on 20 April 2027, and its container images no longer receive updates.4
The problem it solves
Moving a model from a notebook to production needs reproducible environments, tracked experiments, versioned models, scalable training compute, safe deployment and monitoring. Built by hand, these pieces become fragile and hard to audit.
Azure Machine Learning provides them as managed services in one workspace, with lineage from a deployed model back to the code snapshot, data and environment that produced it.
How it works
- Workspace: the central resource for experiments, compute, data assets, environments and models, integrated with Azure Key Vault, Azure Container Registry and virtual networks.
- Compute: compute instances for notebooks; compute clusters and serverless compute for training, including GPU options; Kubernetes and Azure Arc-enabled clusters for hybrid work; serverless Spark or attached Synapse Spark pools.
- Build and track: Python SDK v2, Azure CLI v2, the studio, designer, AutoML, hyperparameter tuning and MLflow-compatible tracking.
- Deploy: managed online endpoints for real-time scoring with traffic splitting, and batch endpoints for asynchronous scoring.
- Govern: registries share models, pipelines and environments across workspaces; lineage links jobs and assets.
Microsoft states that the service does not store or process data outside the region where it is deployed.1
Diagram as a list
Copilot & business apps
- Applications and servicesCall the endpoints for predictionsConnects to Managed online and batch endpoints
AI models & platform
- Managed online and batch endpointsServe models in real time or in batches
Data & analytics
- Workspace (experiments, data, models)Organizes assets and run history
- Model registry and shared registriesVersioned models, pipelines and environments
Developer tools
- Studio, Python SDK v2 and CLI v2Authoring and automation surfacesConnects to Workspace (experiments, data, models)
- GitHub Actions or Azure DevOpsAutomate training and deploymentConnects to Workspace (experiments, data, models)
Security, identity & governance
- Key Vault, virtual network and Container RegistrySecrets, network isolation and imagesConnects to Workspace (experiments, data, models)
Cloud & hybrid infrastructure
- Compute clusters, serverless compute and Arc KubernetesRuns training and batch jobs on CPU or GPU
Capabilities
Training at scale1
Compute clusters and serverless compute run single-node, distributed (PyTorch, TensorFlow, MPI) and embarrassingly parallel training, with GPU options.
Why it matters: Large training jobs run on managed infrastructure instead of hand-built clusters.
Limits: GPU capacity depends on your subscription quota and region.
AutoML and designer3
Automated model, pipeline, hyperparameter and feature generation for tabular, vision and language tasks, plus a drag-and-drop designer.
Why it matters: Speeds up baseline models and lets less code-focused users contribute.
Limits: Generated models still need review for fairness and fit before production.
MLOps13
MLflow tracking, Git integration, scheduled pipelines, Event Grid triggers and CI/CD with GitHub Actions or Azure DevOps, with lineage across jobs and assets.
Why it matters: Makes model changes reproducible and auditable.
Limits: Teams need some familiarity with the workspace and asset model before full MLOps.
Managed endpoints1
Managed online endpoints for real-time scoring with traffic splitting, and batch endpoints for asynchronous jobs, with autoscaling and managed identity.
Why it matters: Deploy and roll out new model versions without running servers.
Limits: You pay for the underlying compute the endpoints use.
Foundry Models access and fine-tuning3
Discover, evaluate, fine-tune and deploy supported foundation models through Foundry Models inside Azure Machine Learning.
Why it matters: Data science teams can fine-tune open models next to their classical work.
Limits: Microsoft says Azure Machine Learning is not the primary platform for generative AI development.
Hybrid compute with Azure Arc3
Training and inferencing on Azure Arc-enabled Kubernetes clusters outside Azure.
Why it matters: Keeps some workloads close to data in other locations.
Limits: You operate the Kubernetes clusters yourself.
Responsible AI tools3
Fairness, explainability and error analysis dashboards.
Why it matters: Supports model reviews and audit evidence.
Limits: Dashboards inform decisions; they do not fix model issues by themselves.
Prompt flow (retiring)1
A tool for prototyping and deploying LLM workflows in Azure Machine Learning and Foundry.
Why it matters: Existing users need a migration plan.
Limits: Retires on 20 April 2027; container images no longer get security or package updates; Microsoft points to Microsoft Agent Framework.
Practical use cases
A retailer needs a demand forecast for each of many stores.
- Approach
- Run embarrassingly parallel training, one model per store, on compute clusters.
- Role of Azure Machine Learning
- Azure Machine Learning schedules and tracks the many training jobs.
- Data, infrastructure and skills
- Historical sales data, a workspace and compute quota.
- Type of benefit
- Planning accuracy
- Caveats
- Compute cost grows with the number of models; monitor forecast drift.
- First step
- Train models for a small set of stores and compare against current forecasts.
Sources 1
A regulated team must prove which data and code produced a deployed model.
- Approach
- Use workspaces, registries, MLflow tracking and pipeline lineage.
- Role of Azure Machine Learning
- Azure Machine Learning records lineage from jobs to deployed models.
- Data, infrastructure and skills
- Git-based code, versioned data assets and a CI/CD pipeline.
- Type of benefit
- Auditability
- Caveats
- Lineage covers what runs through the platform; external steps need their own records.
- First step
- Register one production model with full lineage as a template.
A team wants to fine-tune an open-source LLM on domain data.
- Approach
- Use custom training on GPU compute or Foundry Models fine-tuning inside Azure Machine Learning.
- Role of Azure Machine Learning
- Azure Machine Learning provides GPU compute, tracking and deployment.
- Data, infrastructure and skills
- GPU quota, curated training data and evaluation criteria.
- Type of benefit
- Model customization
- Caveats
- GPU capacity and cost are the main constraints; evaluate safety before release.
- First step
- Run a small fine-tuning job and evaluate it on held-out examples.
Who uses it
Make-A-Wish America (national office and chapters)
In productionMake-A-Wish America: unified data, Copilot and an IT help bot
roughly two-thirds of staff (measured)
Nonprofit, United States
Premier League
In productionPremier League: an AI fan companion built on Microsoft Foundry
up approximately 20% (measured)
Media and entertainment, United Kingdom (global audience)
Works with
Optional integration
- Microsoft FoundryThe Foundry product page lists Azure Machine Learning as an integrated service for model training.
- Microsoft Foundry ModelsFoundry Models can be discovered, fine-tuned and deployed inside Azure Machine Learning.
- Azure ArcHybrid training and inferencing run on Azure Arc-enabled Kubernetes.
- Azure Synapse AnalyticsAttached Synapse Spark pools provide Spark processing.
- Microsoft PurviewPurview can discover and catalog Azure Machine Learning data assets.
- Azure DevOpsAzure DevOps pipelines are a supported CI/CD tool.
Complementary tools
- Azure DatabricksMicrosoft suggests Azure Machine Learning managed endpoints alongside Databricks for cross-team endpoint standards.
- Microsoft Foundry ModelsManaged compute builds on Azure Machine Learning and uses its registries.
- Azure GPU virtual machinesAzure Machine Learning compute clusters and serverless compute offer GPU options for training.
- Foundry ToolsMicrosoft suggests Foundry Tools for preprocessing and retrieval and Azure Machine Learning to train and govern the custom model.
- Microsoft FabricMicrosoft suggests pairing Fabric with Azure Machine Learning for broader model governance.
Alternative approaches
- Azure DatabricksFor lakehouse ML with MLflow and Unity Catalog, Microsoft points to Azure Databricks.
- Microsoft FabricWhen ML is a small extension of analytics, Microsoft suggests Fabric or Databricks as a starting point.
Optional integration for
- Microsoft SentinelSentinel notebooks run in Azure Machine Learning workspaces.
- Azure Database for PostgreSQLThe azure_ai extension can call Azure Machine Learning online endpoints.
- Azure Synapse AnalyticsSynapse Spark integrates with Azure Machine Learning.
- Microsoft FoundryThe product page lists Azure Machine Learning as an integrated service for model training.
- Azure ArcAzure Machine Learning can deploy workloads to Arc-enabled Kubernetes clusters.
Relationship labels follow Microsoft's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.
Getting started
Create a workspace
Create an Azure Machine Learning workspace in the region where your data lives.
Check: The workspace opens in Azure Machine Learning studio.
Arrange compute
Use serverless compute or create a compute cluster, and request GPU quota if you need it.
Check: A test job starts on the chosen compute.
Train and track
Run a training script with MLflow logging.
Check: Metrics and artifacts appear in the job details.
Deploy
Register the model and deploy it to a managed online endpoint.
Check: A test request to the endpoint returns a prediction.
Automate
Build a pipeline and trigger it from GitHub Actions or Azure DevOps.
Check: A scheduled or triggered pipeline completes without manual steps.
Plan the prompt flow exit
If you use prompt flow, follow the migration guide to Microsoft Agent Framework before 20 April 2027.
Check: No production workload depends on prompt flow images.
Official resources
Deciding whether you need it?
Three free tools apply Microsoft's own published guidance to your situation, and say when Azure Machine Learning or any new technology is not needed.
Sources
Each statement above links to the source it comes from. Labels say who reported it.
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.