Overview
Domain 3 moves from classical models to generative AI operations, or GenAIOps, on Microsoft Foundry. Microsoft lists three functional groups. The first covers the platform: creating Foundry resources and projects, using managed identities and RBAC instead of keys, applying private networking, and deploying it all with Bicep and the Azure CLI.
The second group covers foundation models in production. You select a model for a use case, deploy it with serverless API endpoints, standard deployments or managed compute, plan version upgrades and deployment strategies, and size provisioned throughput units for heavy, predictable load. The third group treats prompts as code: designing prompts, creating variants and comparing their results, and storing prompt files in Git so changes are reviewed, versioned and released like any other artifact.
How it's tested
Questions here describe a generative AI workload with requirements on data residency, throughput, cost predictability, security or change control. You decide how to configure the Foundry resource and project, which deployment type and model version policy to use, and how prompts should be versioned and compared before release.
-
Pay per token or reserve
Traffic volume and latency needs decide between standard deployments and provisioned throughput units.
-
Keys or identities
An app must call Foundry models securely, and you choose managed identity with RBAC over API keys.
-
Prompt change control
A prompt edit broke production, and you choose Git versioning, variants and evaluation before release.
A retailer runs a chat assistant with steady, high traffic and strict latency targets during business hours. Standard pay-per-token deployments throttle at peak and costs vary month to month. Provisioned throughput units reserve model processing capacity, giving predictable latency and spend, and a reservation lowers the hourly price. A standard deployment can still absorb overflow, so PTUs are the best answer for the steady base load.
Functional groups 3
Official functional groups from the Microsoft outline, in outline order.
3.1 Implement Foundry environments and platform configurationFoundry resources and projects, managed identities, RBAC, private networking and Bicep deployment.
- What it means
- This group covers building the Microsoft Foundry environment that GenAI teams work in. It includes the Foundry resource and its projects, connections to services such as Azure AI Search and storage, identity and access, network restrictions and declarative deployment.
- What to understand
- A Foundry resource is an Azure AI Services account with project management enabled; projects are child resources that isolate team work, files and agents. Use system-assigned or user-assigned managed identities for the resource and apps, and grant built-in roles such as Azure AI User for builders, Azure AI Project Manager for project administration and Azure AI Account Owner for resource management, with key-based authentication disabled where policy requires. Private endpoints and disabling public network access restrict inbound traffic, and dependent services need their own private access. Bicep templates and az commands make the setup repeatable across environments.
- How it shows up
- Expect a requirement such as no API keys or no public access, and choose the role, identity or networking setting that satisfies it.
- Common confusion
- Hub-based projects are the older model; current Foundry projects are created under a Foundry resource.
3.2 Deploy and manage foundation models for production workloadsServerless API and managed compute deployments, model selection, versioning strategies and provisioned throughput.
- What it means
- This group covers getting foundation models from the Foundry model catalog into production. It includes choosing a model for a use case, picking the deployment option, managing model versions over time and sizing provisioned throughput for high-volume workloads.
- What to understand
- Models sold directly by Azure deploy as standard (pay-per-token) or provisioned deployments, in global, data zone or regional scopes that affect where processing happens. Some partner models use serverless API deployments billed per token, while many open models need managed compute, where you pay for VM hours. Model selection weighs capability, benchmarks, context length, cost, latency and region availability. Version upgrade policies either move to the new default version automatically or pin a version until it retires. Provisioned throughput units reserve capacity, return 429 errors when exceeded and can be discounted with reservations.
- How it shows up
- Read for data residency, predictable cost and burst behaviour, because those details separate global, data zone, standard and provisioned options.
- Common confusion
- Serverless API endpoints have no VMs to manage, while managed compute deployments run on dedicated VMs you pay for while they exist.
3.3 Implement prompt versioning and management with source controlPrompt design, variants and comparison, and Git-based version control for prompts.
- What it means
- This group treats prompts as versioned engineering assets. It covers writing system and user prompts, creating alternative versions to compare, and keeping prompt files in a Git repository so changes go through review and can be traced or rolled back.
- What to understand
- Good prompts state the role, task, constraints, grounding instructions and output format, and use few-shot examples where helpful. Prompt files such as Prompty assets keep model settings and template text together, which makes them easy to diff. Variants change one element at a time so evaluation results can be attributed to that change. Store prompts beside application code, use branches and pull requests for edits, run evaluations in CI before merging and tag releases so a deployed app maps to a known prompt version.
- How it shows up
- Look for teams editing prompts directly in production or unable to tell which prompt caused a regression. The answer involves Git and controlled comparison.
- Common confusion
- A prompt variant is an alternative prompt under test, not a model version or a deployment.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Foundry resource
Azure resource that hosts models, projects and connections in Microsoft Foundry.
- Remember
- Projects are created as children of the resource.
- Tested as
- Platform setup questions.
- Trap
- Assuming every new project needs a hub.
-
Foundry project
Workspace for a team's agents, evaluations, files and deployments access.
- Remember
- Isolates work while sharing the parent resource's model deployments.
- Tested as
- Team isolation scenarios.
- Trap
- Treating a project as a separate billing account.
-
Azure AI User role
Built-in role for building and using Foundry projects.
- Remember
- Common least-privilege choice for developers and app identities.
- Tested as
- RBAC selection questions.
- Trap
- Granting Owner when a developer only needs to build.
-
Keyless authentication
Calling models with Microsoft Entra tokens from managed identities.
- Remember
- Disable local key auth to enforce it.
- Tested as
- Security requirements that forbid shared secrets.
- Trap
- Rotating keys instead of removing them.
-
Global, data zone and regional deployments
Deployment scopes controlling where inference is processed.
- Remember
- Data zone keeps processing within a geography such as the US or EU.
- Tested as
- Data residency scenarios.
- Trap
- Choosing global when processing must stay in the EU.
-
Provisioned throughput units
Reserved model processing capacity billed hourly.
- Remember
- Predictable latency; overflow returns 429 unless spillover is configured.
- Tested as
- High-volume, steady workloads.
- Trap
- Using PTUs for small or unpredictable traffic.
-
Serverless API deployment
Per-token hosted deployment for supported catalog models.
- Remember
- No VM management or quota for compute.
- Tested as
- Low-operations deployment questions.
- Trap
- Assuming every catalog model supports it.
-
Managed compute deployment
Model hosted on dedicated VMs in your subscription.
- Remember
- Charged while the deployment exists, even when idle.
- Tested as
- Open model hosting scenarios.
- Trap
- Forgetting to delete unused deployments.
-
Version upgrade policy
Setting that controls automatic model version upgrades.
- Remember
- Pin a version when output stability matters, and plan for retirement dates.
- Tested as
- Change-control questions.
- Trap
- Assuming a pinned version never needs updating.
-
Prompt variant
Alternative prompt compared against a baseline.
- Remember
- Change one element per variant to attribute results.
- Tested as
- Prompt experimentation questions.
- Trap
- Changing several things at once and guessing which helped.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Foundation model deployment options
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Standard (pay per token) | Shared capacity billed by tokens used. | Simple and elastic, but subject to rate limits. | Expecting fixed latency during peak demand. |
| Provisioned throughput | Reserved capacity billed per PTU hour. | Predictable performance and spend. | Buying PTUs for low or bursty traffic. |
| Serverless API | Per-token hosting for supported partner and open models. | No infrastructure to manage. | Assuming all models are available this way. |
| Managed compute | Dedicated VMs hosting a catalog model. | Pay for VM time whether used or not. | Leaving idle deployments running. |
Access to Foundry
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| API key | Shared secret sent with each request. | Easy but must be stored and rotated. | Using keys when policy forbids secrets. |
| Managed identity with RBAC | Microsoft Entra token for an Azure-hosted app. | No secret to manage; access scoped by role. | Forgetting to assign the role on the right scope. |
| User identity with RBAC | Developer signs in with Microsoft Entra ID. | Suited to portal and SDK development work. | Using personal credentials in production apps. |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Global for everything
- What candidates pick
- A global standard deployment for an EU-only data requirement
- Why it looks right
- Global deployments offer the highest default quota.
- Why it's wrong
- Global deployments can process data in any region; a data zone or regional deployment keeps processing within the required boundary.
Clue to watch for Mentions of residency, sovereignty or processing location.
-
PTUs fix all throttling
- What candidates pick
- Move a small, bursty workload to provisioned throughput
- Why it looks right
- PTUs are presented as the premium option.
- Why it's wrong
- Provisioned capacity is billed whether used or not and suits steady high volume; bursty small loads fit standard deployments.
Clue to watch for Traffic described as occasional, spiky or low.
-
Owner for developers
- What candidates pick
- Assign Owner on the Foundry resource to developers
- Why it looks right
- It guarantees nothing is blocked.
- Why it's wrong
- Least privilege points to Azure AI User for building, keeping management rights with fewer people.
Clue to watch for The question stresses minimal permissions.
-
Managed compute is serverless
- What candidates pick
- Deploy an open model to managed compute to avoid infrastructure costs
- Why it looks right
- Both options are listed side by side in the catalog.
- Why it's wrong
- Managed compute runs dedicated VMs that incur charges while the deployment exists.
Clue to watch for Cost minimisation with idle periods.
-
Auto-upgrade for stability
- What candidates pick
- Let the deployment auto-upgrade to stay current
- Why it looks right
- Newest model sounds safest.
- Why it's wrong
- Automatic upgrades can change outputs; pin versions and test upgrades through evaluation when stability matters.
Clue to watch for Regulated outputs or strict regression requirements.
-
Prompts as configuration only
- What candidates pick
- Store prompts in app settings and edit them live
- Why it looks right
- It allows fast changes without redeploying.
- Why it's wrong
- Live edits lose history and skip review; Git versioning with evaluation gates gives traceability and rollback.
Clue to watch for A prompt change caused an unexplained regression.
Answer framework
Apply these steps when a scenario asks how to run a generative AI workload on Microsoft Foundry.
- Create a Foundry resource and projects through Bicep or Azure CLI so every environment is built the same way.
- Assign managed identities and least-privilege roles, disable keys if required, and add private endpoints for the resource and its dependencies.
- Select the model from capability, cost, latency, context and regional availability, then choose global, data zone or regional scope.
- Pick standard, provisioned, serverless API or managed compute deployment based on traffic shape and operations effort, and set the version upgrade policy.
- Keep prompts in Git, compare variants with evaluations in pull requests, and tag the prompt version released with each deployment.
Practice question
Want more like this? The AI-300 practice test has questions for every domain.
Checklist
0 of 12
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Begin by building, not reading. Deploy a Foundry resource and a project with Bicep, assign yourself Azure AI User, give a test app a managed identity and call a model without any key. Then disable key authentication and confirm what still works. Add a private endpoint and note which connected services also need private access.
Next, study deployment types side by side. Deploy one model as global standard and as data zone standard, read the quota and pricing pages for provisioned throughput, and write down when each fits. Browse the model catalog and practise selecting models from benchmark, cost and region data. Look up the version upgrade policy settings and current retirement schedule for one model family. Finally, put two prompt variants in a Git repository as Prompty files, run a small evaluation on each, and merge the better one through a pull request. Practice questions in this domain are mostly selection problems, so build a comparison table of deployment options and roles and review it before the exam.
FAQ
How much of AI-300 is Domain 3, Design and implement a GenAIOps infrastructure?
Design and implement a GenAIOps infrastructure is weighted 20–25% of AI-300. Assuming about 50 scored items, plan for roughly 10 to 12 questions.
Is Microsoft Foundry the same as Azure AI Foundry?
Yes. Microsoft renamed Azure AI Foundry to Microsoft Foundry. Current exam wording uses Foundry resources and Foundry projects.
Do I need to know hub-based projects?
Focus on Foundry projects under a Foundry resource, which the study guide describes. Knowing that hub-based projects are the older model helps you spot outdated answer options.
When are provisioned throughput units the right choice?
When traffic is high and steady and you need predictable latency and cost. For small or spiky workloads, standard pay-per-token deployments are usually cheaper.
What is the difference between serverless API and managed compute?
Serverless API deployments bill per token with no infrastructure to manage. Managed compute deploys a model to dedicated VMs that are billed while they run.
Why put prompts in Git?
Git gives prompts history, review and rollback. Combined with evaluations in pull requests, it shows which prompt version caused a change in quality.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 3 knowledge
Test whether you can pick the right Foundry deployment type, access model and prompt workflow for production GenAI scenarios.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the AI-300 AI tutor Get any concept explained your way