AI and Machine Learning Microsoft Foundry

Deployment type

Deployment type is the Azure AI or Microsoft Foundry model hosting option, such as standard, global, data-zone, regional, batch, serverless, or provisioned deployment, that determines availability, data processing location, capacity model, cost, and operational behavior.

Source: Microsoft Learn - Deployment types for Microsoft Foundry Models Reviewed 2026-05-13

Exam trap: Changing production before checking the exact Azure scope, owner, identity, destination, and rollback or recovery procedure.
Production check: Can you name the Azure scope, owner, resource ID, dependency, and rollback or recovery path without guessing?

Practice this topic Microsoft Learn

Article details and learning context

Aliases: Microsoft Foundry deployment type, Azure AI deployment type, model deployment type, Foundry Models deployment type
Difficulty: intermediate
CLI mappings: 4
Last verified: 2026-05-13

Related context No curated learning path mapped yet; use the category trail.

Compare related concepts Search related resources Browse the glossary

Understand the concept

In plain English

Deployment type is the model hosting option that determines how an Azure AI or Microsoft Foundry model deployment runs, scales, bills, and processes inference traffic. In Azure, it helps teams select the right model serving pattern for latency, quota, capacity reservation, regional requirements, workload isolation, and cost expectations. Plainly, it is a named part of the architecture that operators can point to when they need evidence, ownership, and a safe change path. A useful glossary entry should explain where it appears, what it controls, what depends on it, and which signal proves it is healthy.

Why it matters

Deployment type matters because a small Azure setting can shape customer experience, security posture, operational visibility, and incident recovery. When it is shallowly documented, teams may troubleshoot the wrong service, queue, device, policy, deployment, workspace, or destination while the real dependency remains hidden. In enterprise Azure work, the value is shared language: application, platform, security, data, finance, and operations teams can discuss the same object without guessing. That reduces incident time, improves audit quality, clarifies ownership, and makes production changes safer because failure modes and graph relationships are visible before change. Treat Deployment type as production owned when customer traffic, regulated data, device fleets, shared infrastructure, or release automation depends on it.

Technical context

Technically, Deployment type appears in Azure AI Foundry model deployment settings, Azure OpenAI deployments, Microsoft Foundry model catalog, SKU names, quota pages, capacity settings, and inference endpoint configuration and interacts with Azure AI Foundry, Azure OpenAI Service, Microsoft Foundry Models, and Azure Monitor. Configuration is reviewed through deployment type, SKU name, model version, and capacity value, while operators validate live state through deployment list, SKU and capacity, model availability, and quota usage. Scope determines which permissions, logs, commands, and dependencies matter.

Exam context

Compare with

Model deployment type Deployment capacity Provisioned throughput

Where it is used

Where you see it

In Azure AI Foundry, deployment type appears when selecting how a model endpoint should serve requests, bill usage, and handle data processing geography during production review.
In quota planning, it appears when provisioned, regional, data-zone, batch, or global options have different capacity and approval requirements during production review when operators collect repeatable evidence.
In incident review, it appears when throttling, latency, data residency, or unexpected model cost traces back to the selected deployment type during production review when operators collect repeatable evidence.

Common situations

Choose a deployment type for a production generative AI endpoint.
Compare provisioned capacity against pay-per-token standard deployments.
Investigate latency, throttling, quota, and data processing location for model traffic.

Illustrative Azure scenarios

These examples show how the concept can affect design and operations. They are illustrative scenarios, not customer claims.

Scenario 01 Deployment type in action for consumer banking Scenario, objectives, solution, measured impact, and takeaway.

Scenario

FinchBank Digital, a consumer banking organization, needed to address a customer service chatbot had unpredictable latency during statement season and unclear data processing expectations. The architecture team used Deployment type as the control point for a measurable production improvement.

Goals

Keep median model response time below 900 milliseconds
Document approved data processing geography
Reduce surprise token spend by 20 percent

Approach using Deployment type

Architects reviewed the available model deployment types in Azure AI Foundry and moved the production chatbot from an exploratory standard deployment to an approved provisioned deployment type with explicit capacity. Private networking, monitoring, content safety checks, and chargeback tags were connected to the deployment record. The team validated Deployment type in a lower environment, captured before-and-after evidence, and promoted the change through a controlled release with rollback ownership. Runbooks were updated so support engineers could identify the correct scope, identity, dependency, telemetry signal, and approval record without asking the original implementer. The final design connected governance with day-to-day engineering work, which made the change understandable to security, operations, finance, and application stakeholders. The team validated Deployment type in a lower environment, captured before-and-after evidence, and promoted the change through a controlled release with rollback ownership.

Potential outcomes

Median response time improved 38 percent during peak demand
Data processing choice was approved before launch
Monthly variance in model spend dropped 27 percent

What to learn

Deployment type turns model hosting from a guess into an architecture decision with cost, latency, and compliance consequences.

Scenario 02 Deployment type in action for healthcare diagnostics Scenario, objectives, solution, measured impact, and takeaway.

Scenario

MetroCare Labs, a healthcare diagnostics organization, needed to address clinical summarization pilots were created in several regions without a consistent capacity or residency decision. The architecture team used Deployment type as the control point for a measurable production improvement.

Goals

Standardize model deployment decisions across three workloads
Keep regulated inference traffic in approved locations
Give platform teams repeatable evidence for audits

Approach using Deployment type

The cloud team defined deployment type selection criteria for each AI workload. Development used standard deployments, while regulated production summarization used a regional option with approved monitoring and diagnostic settings. The team documented model version, deployment type, quota, owner, and rollback in the release checklist. The team validated Deployment type in a lower environment, captured before-and-after evidence, and promoted the change through a controlled release with rollback ownership. Runbooks were updated so support engineers could identify the correct scope, identity, dependency, telemetry signal, and approval record without asking the original implementer. The final design connected governance with day-to-day engineering work, which made the change understandable to security, operations, finance, and application stakeholders. The team validated Deployment type in a lower environment, captured before-and-after evidence, and promoted the change through a controlled release with rollback ownership.

Potential outcomes

Audit evidence collection time fell from two days to four hours
Unapproved regional deployments were removed
Production change reviews became consistent across workloads

What to learn

A deployment type is part of the AI governance record, not just a dropdown in the model portal.

Scenario 03 Deployment type in action for ecommerce Scenario, objectives, solution, measured impact, and takeaway.

Scenario

HarborRetail Online, a ecommerce organization, needed to address recommendation testing generated high token usage but did not require low-latency interactive responses. The architecture team used Deployment type as the control point for a measurable production improvement.

Goals

Separate batch experimentation from production inference
Cut test workload cost by at least 25 percent
Preserve capacity for customer-facing recommendations

Approach using Deployment type

Engineers separated interactive and offline scoring paths by deployment type. Production recommendations stayed on a low-latency deployment, while experimentation moved to a batch-suitable deployment pattern with scheduled runs, cost tags, and output validation. Azure Monitor alerts tracked quota pressure and failed jobs. The team validated Deployment type in a lower environment, captured before-and-after evidence, and promoted the change through a controlled release with rollback ownership. Runbooks were updated so support engineers could identify the correct scope, identity, dependency, telemetry signal, and approval record without asking the original implementer. The final design connected governance with day-to-day engineering work, which made the change understandable to security, operations, finance, and application stakeholders. The team validated Deployment type in a lower environment, captured before-and-after evidence, and promoted the change through a controlled release with rollback ownership.

Potential outcomes

Experimentation spend decreased 32 percent
Customer-facing capacity was no longer consumed by batch tests
Recommendation release cycles shortened by five days

What to learn

Selecting the right deployment type prevents nonproduction AI experiments from distorting production performance and cost.

Azure CLI

CLI checks for Deployment type are useful because they turn portal assumptions into repeatable evidence. Start with read-only commands that show scope, state, owner, permissions, destinations, deployment settings, or device records. Run mutating, security-impacting, or cost-impacting commands only after approval, because the wrong scope can affect production availability, spend, access, or telemetry.

Useful for

Choose a deployment type for a production generative AI endpoint.
Compare provisioned capacity against pay-per-token standard deployments.
Investigate latency, throttling, quota, and data processing location for model traffic.

Before you run a command

Run az account show, confirm tenant and subscription, and verify the operator identity has approved read access for the exact Azure scope.
Confirm resource group, service name, resource ID, environment, owner, and change record before collecting evidence or modifying production configuration.
Prefer read-only commands first; review any command that creates deployments, changes policy, alters device access, or routes telemetry before running it.

What the output tells you

Whether the resource, setting, device, deployment, policy, queue, or API Management object exists at the expected Azure scope.
Which state, target, timestamp, SKU, identity, destination, count, property, or compliance result is visible to the operator.
Whether the issue is wrong scope, stale configuration, missing permissions, broken telemetry routing, policy drift, device provisioning failure, or release mismatch.

Mapped commands

Deployment type operational checks

direct

az cognitiveservices account deployment list --name <account> --resource-group <resource-group> --output table

az cognitiveservices account deploymentdiscoverAI and Machine Learning

az cognitiveservices account deployment show --name <account> --resource-group <resource-group> --deployment-name <deployment>

az cognitiveservices account deploymentdiscoverAI and Machine Learning

az cognitiveservices account list-skus --name <account> --resource-group <resource-group>

az cognitiveservices accountdiscoverAI and Machine Learning

az cognitiveservices account deployment create --name <account> --resource-group <resource-group> --deployment-name <deployment> --model-name <model> --model-version <version> --model-format OpenAI --sku-name <sku> --sku-capacity <capacity>

az cognitiveservices account deploymentprovisionAI and Machine Learning

Architecture context

Deployment type belongs to AI and Machine Learning architecture decisions where identity, monitoring, cost ownership, reliability, and production support need shared evidence.

Security: Security for Deployment type starts with least privilege, trusted configuration, and evidence that access matches workload risk. Review data processing location, model access permissions, private endpoint usage, managed identity access, and responsible AI controls before approving production use. A common failure is assuming that a working feature, successful deployment, connected device, or populated log destination proves the configuration is safe. Use Microsoft Entra groups, managed identities, RBAC, private connectivity, diagnostic logging, source-controlled definitions, and approval records where applicable. Keep exceptions ticketed, time-bounded, and owned. For regulated workloads, align the term with classification, retention, break-glass, and incident-response procedures. Remove broad access, stale keys, unreviewed contributors, and undocumented exception paths before Deployment type becomes an incident path.
Cost: Cost for Deployment type appears through compute capacity, transaction volume, diagnostic retention, policy remediation, storage consumption, API exposure, message retries, device fleet operations, and the human effort required to recover from mistakes. Review pay-per-token usage, provisioned capacity, batch processing cost, unused quota, and benchmarking spend before expanding production use. Some costs are direct, such as retained logs, provisioned capacity, storage transactions, or queue processing; others are indirect, such as failed releases, duplicated troubleshooting, emergency restores, and support escalation. Tag related resources, monitor usage, and separate exploratory work from production. A cost review should connect spend to a real owner and measurable value.
Reliability: Reliability for Deployment type depends on repeatable configuration, tested dependencies, and clear failure signals. Watch quota availability, regional model availability, failover option, throttling behavior, and capacity reservation because drift often appears later as failed releases, missing telemetry, stuck messages, failed device provisioning, unavailable APIs, or confusing support evidence. Use lower environments, source-controlled definitions where possible, deployment validation, monitoring, and recovery notes before changing production. Operators should know which resource, endpoint, queue, policy, workspace, device, or downstream application fails first and which metric or log proves the failure. The goal is predictable recovery: detect Deployment type drift, preserve service, restore safely, and explain the incident without guessing.
Performance: Performance for Deployment type depends on workload shape, service limits, data volume, network path, diagnostic destination, policy evaluation, device scale, queue behavior, deployment capacity, and the monitoring path used to confirm success. Review request latency, throughput units, rate limits, region proximity, and model warm-up before increasing capacity or retrying blindly. The better fix might be correcting partitioning, reducing log noise, warming an endpoint, tuning queue visibility, selecting a different deployment type, or moving telemetry to a better destination. Measure under representative production conditions. Operators should connect symptoms to evidence: latency, throttling, backlog, failed operations, dropped logs, or stale state. Good performance work ties Deployment type measurements to user impact and avoids hiding design issues behind larger resources.
Operations: Operations for Deployment type should focus on ownership, observability, and safe repeatability. Standardize names, tags, owner groups, environment labels, diagnostic destinations, runbook links, approval records, and change windows so support teams do not reverse-engineer the platform during incidents. Use read-only CLI, API, policy, diagnostic, or portal checks first, then compare live state with intended configuration. For production, connect alerts, audit events, cost records, graph links, and release notes to the same term. The support question should be simple: who owns it, what changed, and what proves the current state?. Capture owner, scope, evidence, and recovery procedure before changing Deployment type in a production environment.

Common mistakes

Changing production before checking the exact Azure scope, owner, identity, destination, and rollback or recovery procedure.
Treating a portal screenshot as sufficient evidence when CLI output, activity logs, diagnostics, and source-controlled configuration are repeatable.
Assuming a name match proves the correct resource when subscriptions, regions, products, device IDs, queues, and workspaces can look similar.