Microsoft Certified: Machine Learning Operations Engineer Associate
Code
AI-300
Length
120 minutes
Pass mark
700 out of 1,000
Price
Set by country, shown when you book
Valid for
12 months, renewed by a free online assessment
Study time, our estimate
60 to 80 hours
AI-300Associate
This replaced DP-100 in 2026, and the change matters. It treats classic MLOps and GenAIOps as one job, with Bicep, GitHub Actions, MLflow, prompt versioning and provisioned throughput all in scope. Good pick if you run models rather than invent them.
Take it if
You own the pipelines and platform behind AI on Azure.
Skip it if
You have no Python or data science background. Microsoft expects both, plus entry-level DevOps.
Before you start
Check yourself against this list first. If something is missing, close that gap before you pay for an exam.
A data science background in PythonMicrosoft expects this.
Entry-level DevOpsGitHub Actions and the command line are assumed.
You have trained a modelTrain one in Azure Machine Learning first. Learn more
None formally. Data science background and Python expected.
If you fail
Wait 24 hours, then 14 days
What the exam covers
Design and implement an MLOps infrastructure15–20%
Implement machine learning model lifecycle and operations25–30%
Design and implement a GenAIOps infrastructure20–25%
Implement generative AI quality assurance and observability10–15%
Optimize generative AI systems and model performance10–15%
What to focus on
Azure Machine Learning workspaces, compute, data assets, environments and registries.
MLflow tracking, model registration, and rollouts with safe rollback.
Deploying workspaces with Bicep, Azure CLI and GitHub Actions, and restricting network access.
Foundry for production: serverless against managed compute, and when provisioned throughput makes sense.
Prompts as code: versioning in Git and comparing variants.
Evaluation metrics by name: groundedness, relevance, coherence, fluency.
Tuning RAG: thresholds, chunk size, hybrid search and A/B tests.
Everything on one mapEvery topic and key idea in this certification.
mindmap
root((AI-300))
MLOps infrastructure
Infrastructure as code
Assets
Network access
Model lifecycle
MLflow
Progressive rollout
Drift
GenAIOps infrastructure
Deployment options
Provisioned throughput
Prompts in Git
Quality and optimization
Evaluators
Tuning RAG
Fine-tuning
The loop AI-300 testsEach arrow is an automation you may be asked about.
flowchart LR
A["Train"] --> B["Track in MLflow"]
B --> C["Register"]
C --> D["Deploy with rollout"]
D --> E["Monitor"]
E -->|"Drift detected"| A
Lesson 1 of 4, 15–20% of the exam
MLOps infrastructure
Environments you can recreate are the base of everything else.
Infrastructure as code
Deploy workspaces with Bicep and Azure CLI, automated through GitHub Actions.
Assets
Data, environments and components can be shared across workspaces through registries.
Network access
Restrict workspaces and use managed identities.
Infrastructure as code for MLThe same template creates every environment.
flowchart LR
A["Bicep in Git"] --> B["GitHub Actions"]
B --> C["Dev workspace"]
B --> D["Test workspace"]
B --> E["Prod workspace"]
F["Registry"] -.->|"Shared assets"| C
F -.-> D
F -.-> E
The largest domain, covering training to monitoring.
MLflow
Tracks runs and registers models.
Progressive rollout
Move a share of traffic and keep rollback ready.
Drift
Compare live data with a baseline and trigger alerts or retraining.
A safe model rolloutMove traffic in steps and keep the way back open.
stateDiagram-v2
direction LR
[*] --> Registered
Registered --> Canary: Small share of traffic
Canary --> Rollout: Metrics hold
Canary --> Rollback: Metrics drop
Rollout --> Production
Rollback --> Registered
Common trap
Replacing a production model in one step.
Try it yourself
Set up an endpoint with a rollout and a drift alert.
This is what makes the exam different from DP-100.
Deployment options
Serverless and managed compute suit different workloads.
Provisioned throughput
Reserved capacity for sustained, high-volume traffic.
Prompts in Git
Version prompts, compare variants, and review changes.
Serverless or provisionedDecide from the traffic shape.
flowchart TD
A{"What does traffic look like?"} -->|"Spiky or unknown"| B["Serverless API"]
A -->|"Steady and high"| C["Provisioned throughput"]
A -->|"Custom model"| D["Managed compute"]
Common trap
Editing a production prompt in the portal with no history.
Groundedness, relevance, coherence and fluency measure different things.
Tuning RAG
Adjust similarity thresholds, chunk sizes and hybrid search, then measure.
Fine-tuning
Use it for behavior, and manage it from development through production.
What each evaluator measuresTwo of these judge truth, and two judge writing.
flowchart LR
A["Answer"] --> B["Groundedness: supported by the source?"]
A --> C["Relevance: answers the question?"]
A --> D["Coherence: logical?"]
A --> E["Fluency: reads well?"]
Common trap
Fine-tuning to fix missing knowledge. That is a retrieval problem.
Try it yourself
Run an automated evaluation, change chunk size, and compare results.
Create a workspace with Bicep, track a training run with MLflow, register the model, and deploy an endpoint.
Week 3: automate it
Build a GitHub Actions pipeline that trains, evaluates and deploys. Add drift monitoring and a retraining trigger.
Week 4: GenAIOps
Deploy a foundation model in Foundry, version prompts in Git, and wire up automated evaluation and tracing.
Week 5: optimize
Tune a RAG system on chunk size and hybrid search, measure the change, and fine-tune a model on synthetic data.
Exam week
Take the free practice assessment and the mock below.
Exam day and after
You can open Microsoft Learn during associate and expert exams, but the clock keeps running. It helps with one lookup and not with twenty, so use it for exact names and limits.
Most questions cover generally available features, though preview features can appear if they are widely used.
Try the free exam sandbox first so the interface is not a surprise on the day.
Older DP-100 notes cover part of this exam, but they leave out GenAIOps, and the last three domains together are 40 to 55% of the marks.
Test yourself8 practice questions, scored by exam topic.
Exam facts checked on October 31, 2026 against AI-300 study guide. Lessons are our own explanations, so use the official docs as the final word, and confirm price and dates before booking.
