Free tools Windows power users keep installed
One-click scans. No signup required.
Use Microsoft Azure for AI by matching the platform to the work first: Microsoft Foundry is the usual starting point for generative-AI applications and agents, while Azure Machine Learning is designed for traditional machine learning, custom training, advanced data preparation, and end-to-end MLOps. A standalone Azure OpenAI resource can be enough when you only need model completions. After choosing the platform, select a model deployment whose processing boundary, capacity model, region, and price fit your workload and compliance requirements.
This guide explains that decision in operational terms, including a reference architecture, deployment steps, production safeguards, recovery planning, and responsible-AI controls. Product names, preview status, regional availability, quotas, pricing, and service-level terms change, so verify them for your model and Azure region before you provision anything.
As an Amazon Associate I earn from qualifying purchases.
Start by classifying the workload
The first Azure decision is not a portal menu; it is the job your system must perform. Microsoft’s Well-Architected guidance separates intelligent applications from training and fine-tuning workloads because their state, dependencies, scaling, availability, security, and responsible-AI requirements differ.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Intelligent applications
An intelligent application uses an existing model to perform a user or business task. Examples include a grounded question-answering assistant, an agent that calls business APIs, document summarization, classification, extraction, and customer-service automation. These systems usually need an inference endpoint, application hosting, identity, retrieval or tool connections, monitoring, and controls for user data and model output.
Training and fine-tuning
Training workloads create or adapt models. They require data preparation, experiment tracking, compute orchestration, model registration, evaluation, reproducible pipelines, and a path to deployment. Fine-tuning can also introduce sensitive-data handling, longer-running jobs, and a need to compare versions before release.
Write down the workload before selecting a service. Record the expected users, request pattern, latency target, model inputs and outputs, data sensitivity, geographic restrictions, evaluation method, and whether the model is built in-house or supplied by Azure. This short inventory prevents a chat application from being forced into a training platform, or a production MLOps program from being treated as a one-off prompt endpoint.
Choose the Azure platform that fits
Microsoft Foundry for generative applications and agents
For a new generative-AI application or agent, Microsoft recommends considering Microsoft Foundry as the development platform. Foundry brings model access, agent development, evaluation, governance, and connected Azure services into a project-oriented experience.
Recommended Free Tools
A Foundry setup is organized around a top-level Foundry resource, projects that isolate development work, and connections to services such as storage, search, and secrets management. A single resource with one project is a sensible default for one developer exploring an idea. Multiple teams, environments, or compliance boundaries may require separate projects or resources, with explicit ownership and policy controls.
Azure Machine Learning for traditional ML and MLOps
Azure Machine Learning remains the better fit for supervised or unsupervised machine learning, custom model training, advanced data preparation, and end-to-end MLOps. It supports the lifecycle around a model rather than only the inference interaction: datasets, experiments, pipelines, registries, managed endpoints, and release processes.
Use it when you need repeatable training runs, feature and data-processing workflows, model lineage, controlled promotion between environments, or custom compute. You can still consume generative models elsewhere in an overall solution; choosing Azure Machine Learning does not mean every component must be trained from scratch.
A standalone Azure OpenAI resource for limited completion use cases
If the requirement is limited to sending prompts and receiving completions, a standalone Azure OpenAI resource may be sufficient. This can be simpler than adopting the broader Foundry project model when you do not need agent hosting, integrated evaluation, or a larger set of connected tools. Reassess the choice if the application grows into retrieval, agents, multi-team governance, or formal evaluations.
| Workload need | Typical starting point | Why |
|---|---|---|
| Generative application or agent | Microsoft Foundry | Project-based development, agent capabilities, model access, evaluation, and service connections. |
| Traditional ML, custom training, advanced data preparation | Azure Machine Learning | Experimentation, pipelines, custom compute, model management, and MLOps. |
| Simple prompt/completion API | Standalone Azure OpenAI resource | Smaller operational surface when agent hosting and broader project features are unnecessary. |
| Open-source or custom-weight model on dedicated GPUs | Foundry managed compute, where supported | Dedicated managed capacity, subject to preview limitations and regional/model availability. |
Set up Foundry with useful boundaries
Create the resource and project
- Choose the Azure subscription, resource group, region, naming convention, and owners. Treat the region as a data-processing and availability decision, not merely a place to put the resource.
- Create the Microsoft Foundry resource and an initial project. Keep development and production ownership distinct; a single-project layout is mainly a convenient exploration default.
- Define role assignments through Microsoft Entra ID. Grant developers, deployment operators, application identities, and auditors only the permissions each role needs.
- Add connections for the services the application actually uses, such as Azure Storage for files, Azure AI Search for retrieval, and a secrets-management service for credentials. Keep secrets out of source code and prompt templates.
- Record the model, deployment type, region, data boundary, quota, and version in an environment document so a later deployment can be reproduced.
Foundry supports networking choices, but isolation requirements vary. Private connectivity, firewall rules, DNS, managed identities, and egress restrictions may require SDK or command-line configuration instead of a single portal setting. Decide whether public access is acceptable before application code is distributed.
Select a model deployment type
Deployment type controls more than throughput. It affects where inference data may be processed, how capacity is purchased, how traffic behaves, and which models and features are available. Verify the exact combination in the target region because the menu differs by model and location.
Rank #2
Serverless API options
Foundry serverless API deployments include standard, provisioned, batch, and developer types.
| Type | Best fit | Payment or capacity pattern | Key caution |
|---|---|---|---|
| Standard | Interactive, variable traffic and ordinary API calls | Pay per token or other metered usage defined for the model | Capacity and latency can vary with demand; check quotas and regional limits. |
| Provisioned | Steady, high-volume traffic with a predictable throughput target | Reserved provisioned capacity | Can cost more than metered service at low utilization; size capacity from measured demand. |
| Batch | Asynchronous jobs that can wait, such as large document transformations or offline scoring | Batch pricing and job-based processing | Not appropriate for interactive response latency; supported models and regions vary. |
| Developer | Development and experimentation | Developer-oriented limits and terms | Do not assume development limits or availability are suitable for production. |
Microsoft’s deployment documentation lists Global Batch and Data Zone Batch as offering a 50% discount for supported large asynchronous jobs. That is a documented relative discount for eligible deployments, not a guarantee that your total bill will be half of another design. Token volume, model, input and output sizes, storage, networking, retries, and other services still affect the invoice.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsManaged compute for open-source or custom-weight models
Managed compute hosts supported open-source or custom-weight models on dedicated GPU capacity. The documentation available for this guide describes managed compute as being in public preview, without an SLA, and not recommended for production workloads. Preview status can change, so check the current service page before relying on it.
Choose the processing boundary deliberately
Ask two separate questions: where is data stored at rest, and where can it be processed during inference? A storage account in one geography does not automatically mean that every inference operation stays there.
Global deployments
Global deployments may process inference data in any Azure region. They can provide broader capacity and routing flexibility, but they are unsuitable when policy requires processing to remain within a defined geography or data zone.
Data-zone deployments
Data-zone deployments constrain inference processing to the specified Azure data zone. This can reduce the allowed processing area while retaining more flexibility than a single-region design. Confirm which zone the particular model supports.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Geography-based deployments
Standard and Regional Provisioned deployments process within the selected Azure geography. Operational processing can occur between regions inside that geography, so document the geography-level boundary rather than promising a single physical datacenter.
| Decision axis | Questions to answer |
|---|---|
| Processing boundary | Can inference occur globally, anywhere in a data zone, or only within the selected Azure geography? |
| Traffic pattern | Is traffic bursty and interactive, steady and high volume, or asynchronous? |
| Capacity and latency | Will metered standard capacity meet the target, or is reserved provisioned throughput justified? |
| Availability | Is the exact model, deployment type, quota, and feature set offered in the target region? |
| Cost | Should you estimate token charges, reserved throughput, or batch pricing for the actual configuration? |
Write the chosen boundary into your data-protection assessment. Also document what is stored, retention periods, logging destinations, and whether prompts, retrieved documents, or outputs are copied into application telemetry.
Deploy a model and connect an application
- In Foundry, open the project’s model catalog and select the model version that matches the task, context length, modality, and quality requirement.
- Check that the model and desired deployment type are available in the chosen region. If not, select another supported region or revise the processing boundary; do not assume a model can be moved after application launch.
- Create the deployment and choose standard, provisioned, batch, or another supported option. Set capacity or throughput parameters only after estimating request volume, token size, concurrency, and peak demand.
- Configure authentication with a managed identity where possible. If an API key is unavoidable, store it in a secrets-management service and rotate it.
- Call the endpoint from a small test client. Verify authentication, model version, maximum input and output sizes, error handling, timeout behavior, and content-filter responses.
- Add application-level controls: input validation, prompt and tool allowlists, retrieval filters, output schema checks, rate limits, and a human escalation path for high-impact decisions.
- Instrument latency, token usage, status codes, throttling, safety events, and quality signals. Exclude sensitive prompt content from logs unless there is a documented need, access control, and retention rule.
For a retrieval-augmented assistant, index approved content in Azure AI Search or another connected retrieval service. Retrieve only documents the caller is authorized to see, then pass the smallest relevant context to the model. Retrieval does not replace authorization: a search index must enforce the same tenant, role, and document permissions as the source system.
Understand the introductory chat architecture—and its limits
Microsoft’s Basic Microsoft Foundry Chat Reference Architecture places a web user interface on Azure App Service. The application calls an agent hosted in Foundry Agent Service. The agent can obtain grounding information from Azure AI Search and other connected tools, then send relevant context to an Azure OpenAI model. Identity and monitoring components are included around the request path.
Microsoft explicitly labels this architecture as a learning and proof-of-concept example and states, “This architecture isn’t for production.” It is single-region. A production design should instead start with Microsoft’s baseline Foundry chat architecture and adapt it to the organization’s identity, networking, data, availability, and recovery requirements.
Changes needed before production
- Separate development, test, and production resources or projects with controlled promotion.
- Use private networking and managed identities where isolation and policy require them.
- Define regional or multi-region behavior, including what happens when a model endpoint, search service, storage account, or identity dependency is unavailable.
- Protect retrieval indexes and application APIs with authorization at every layer.
- Set budgets, quotas, rate limits, alert thresholds, and a procedure for model or deployment rollback.
- Test prompt-injection resistance, sensitive-data leakage, unsafe output, tool abuse, and degraded-mode behavior.
Plan Azure Machine Learning reliability and recovery
The following recommendations apply specifically to Azure Machine Learning workloads. They are not a universal guarantee for every Azure AI service.
Separate exploration from production
Use different workspaces or otherwise strong boundaries for exploratory experiments and production assets. Exploration often permits ad hoc notebooks and changing dependencies; production requires controlled images, registered models, repeatable pipelines, and auditable releases.
Provide enough compute
Size training and inference compute for normal demand plus peaks. Track queue time, utilization, startup time, and failed jobs rather than selecting a machine size from a single test. Keep a documented fallback when a GPU family or quota is unavailable in the preferred region.
Release managed online endpoints safely
Use a release strategy such as blue-green deployment for managed online endpoints. Deploy the candidate alongside the current version, direct controlled traffic to it, compare functional and operational metrics, and switch over only after acceptance criteria are met. Keep the previous version available for rollback until the new release is stable.
Design for regional failure
Consider availability zones or multiple regions where the workload’s recovery objective justifies the added cost and operational complexity. Microsoft states that Azure Machine Learning does not provide automatic failover. Recovery planning therefore must include the workspace and dependencies such as Key Vault, Azure Storage, and Azure Container Registry. Document backups, replication, redeployment steps, DNS or routing changes, and the person authorized to declare a recovery event.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Build responsible-AI controls into the lifecycle
Responsible AI is an application policy, not a switch that makes every model safe. Microsoft advises documenting policies for each lifecycle stage and defining how your organization applies responsible-AI principles in its own context.
Identify potential harms
List plausible harms for the specific users and decisions: inaccurate advice, exposure of confidential records, discriminatory recommendations, unsafe instructions, impersonation, and over-reliance on automated output. Consider vulnerable users and high-impact domains separately.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #4
Measure before release
Create representative evaluation sets, including adversarial and edge cases. Measure factuality or groundedness, refusal behavior, sensitive-data leakage, toxicity, latency, cost, and task success. For a classifier or traditional ML model, include class-level performance and error costs rather than a single aggregate score.
Mitigate and operate
Use content filters, grounding, constrained tools, output validation, human review, rate limits, and clear user disclosures. Azure OpenAI guidance frames the process as identifying harms, measuring them, mitigating them, and operating the system responsibly. Monitor after launch because data, prompts, users, and model versions change.
Respect user data rights
Tell users what data is collected and why. Configure consent and opt-out mechanisms where required. Define a procedure for handling a right-to-be-forgotten request across application databases, search indexes, caches, logs, evaluation sets, and backups. Do not claim deletion is complete until each copy and retention exception has been checked.
Control cost without guessing
There is no responsible single price for “an Azure AI workload.” The estimate depends on model, deployment type, input and output tokens, concurrency, provisioned capacity, batch eligibility, storage, search, compute, networking, monitoring, and region.
- Estimate requests per minute, peak concurrency, average input tokens, average output tokens, and retry rates.
- Choose a candidate model and deployment type, then use the current Azure pricing page or calculator for that exact configuration.
- Model separate development, test, and production consumption. Development traffic can be small but persistent, while production may have sharp peaks.
- Compare standard metered service with provisioned capacity using expected utilization, not the maximum theoretical throughput.
- Move work that does not need an immediate response to a supported batch deployment. The documented 50% Global Batch and Data Zone Batch discount applies only where Microsoft supports those deployments and does not establish your final bill.
- Add non-model costs: App Service or other application hosting, search indexes, storage, managed identities, networking, monitoring, GPU or CPU compute, backups, and data transfer.
- Set budgets and alerts, then review token usage and failed or repeated requests after launch.
Pre-provisioning verification checklist
- Workload: Is this a generative application, an agent, a completion-only integration, traditional ML, custom training, or a combination?
- Platform: Does Foundry, Azure Machine Learning, or a standalone Azure OpenAI resource provide the required lifecycle features?
- Model: Is the exact model version, modality, context size, tool support, and deployment type available in the target region?
- Processing: Does global, data-zone, or geography-based processing satisfy contractual and regulatory requirements?
- Data at rest: Where are prompts, documents, indexes, outputs, logs, backups, and secrets stored, and how long are they retained?
- Capacity: Are standard, provisioned, batch, developer, or managed-compute terms appropriate for the traffic pattern?
- Security: Are identity, least privilege, private networking, secret rotation, tenant isolation, and egress controls implemented?
- Reliability: Are quotas, dependencies, zones or regions, backups, rollback, and recovery procedures tested?
- Responsible AI: Have harms been identified, measured, mitigated, disclosed, and assigned to an operating owner?
- Cost: Has the current configuration been priced, with budgets and alerts configured?
When to revisit the design
Re-evaluate the architecture when the application adds agents or tools, moves from experimentation to production, begins serving multiple teams, handles regulated data, requires a tighter processing boundary, or experiences sustained throttling. Also revisit it when Microsoft changes model availability, deployment types, quotas, preview status, pricing, or service-level terms. A design that was appropriate for a single developer and a proof of concept may be the wrong boundary for a regulated, multi-team production system.
Frequently Asked Questions
Which Azure service should I use for a generative-AI application?
Microsoft recommends considering Microsoft Foundry for generative applications and agents. Use Azure Machine Learning when the core requirement is traditional ML, custom training, advanced data preparation, or end-to-end MLOps. A standalone Azure OpenAI resource can cover a completion-only integration.
Where can Azure process my AI inference data?
It depends on the deployment type and model. Global deployments may process data in any Azure region; data-zone deployments constrain processing to a specified zone; standard and Regional Provisioned deployments process within the selected Azure geography, potentially between regions in that geography. Verify the exact model and region.
Is the basic Foundry chat architecture production-ready?
No. Microsoft labels the Basic Microsoft Foundry Chat Reference Architecture as a learning and proof-of-concept, single-region example and states that it is not for production. Use the baseline architecture as a starting point for production design.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Does Azure Machine Learning automatically fail over to another region?
Microsoft states that Azure Machine Learning does not provide automatic failover. Recovery planning must include the workspace and dependencies such as Key Vault, Azure Storage, and Azure Container Registry.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
