Exam AI-300: Operationalizing Machine Learning and Generative AI Solutions earns the Microsoft Certified: Machine Learning Operations Engineer Associate credential. It is Microsoft's MLOps and GenAIOps exam and replaces the retired DP-100. It covers two halves: running classic ML with Azure Machine Learning (MLflow, pipelines, endpoints, drift) and running generative AI apps and agents with Microsoft Foundry (model deployment, prompt versioning, evaluations, observability, RAG tuning and fine-tuning).
This guide follows the official AI-300 study guide: all five skill areas with weights, the tools to know, a 7-week plan and 15 original scenario questions.
Quick facts (Microsoft Learn, November 2026)
| Item | Detail |
|---|---|
| Exam | AI-300: Operationalizing Machine Learning and Generative AI Solutions |
| Certification | Microsoft Certified: Machine Learning Operations Engineer Associate (Intermediate) |
| Time | 120 minutes; proctored; may include interactive components |
| Passing score | 700 |
| Language | English |
| Price | Based on the country where you take the exam (Pearson VUE) |
| Renewal | Every year, with a free online assessment on Microsoft Learn |
| Practice | Practice Assessment on AI Skills Navigator (sign-in required) |
Read the official certification page and AI-300 study guide before booking.
Who should take AI-300?
- Data scientists and ML engineers who deploy and maintain models on Azure.
- GenAI engineers who must take Foundry apps and agents from demo to production.
- DevOps engineers moving into MLOps. You need Python, Git, GitHub Actions and the Azure CLI.
Microsoft expects a data science background plus entry-level DevOps. If you're starting out, take AI-901 first, and see AI-103 for building (rather than operating) AI apps and agents.
Skills measured and weights
| Skill area | Weight |
|---|---|
| Design and implement an MLOps infrastructure | 15–20% |
| Implement machine learning model lifecycle and operations | 25–30% |
| Design and implement a GenAIOps infrastructure | 20–25% |
| Implement generative AI quality assurance and observability | 10–15% |
| Optimize generative AI systems and model performance | 10–15% |
1. MLOps infrastructure (15–20%)
- Workspace: workspaces, datastores, compute targets, identity and access.
- Assets: data assets, environments, components; sharing across workspaces with registries.
- IaC: GitHub integration; deploying workspaces with Bicep and the Azure CLI; provisioning with GitHub Actions; restricting network access; Git for ML projects.
2. ML model lifecycle (25–30%)
- Training: MLflow experiment tracking, AutoML, notebooks, hyperparameter tuning, training scripts, distributed training, pipelines, comparing jobs.
- Registration: packaging a feature retrieval spec with the model; registering MLflow models; responsible AI evaluation; archiving models.
- Deployment: real-time and batch endpoints with managed inference; testing endpoints; progressive rollout and safe rollback.
- Monitoring: data drift, production metrics, retraining or alert triggers.
# GitHub Actions step: train in Azure ML, then register only if the job succeeds
- name: Train
run: |
az ml job create -f jobs/train.yml -g rg-ml -w mlw-prod --stream
- name: Register model
run: |
az ml model create -n churn --type mlflow_model \
--path azureml://jobs/$JOB_NAME/outputs/artifacts/paths/model/ -g rg-ml -w mlw-prod3. GenAIOps infrastructure (20–25%)
- Foundry environments: resources and projects; managed identities and RBAC; private networking; Bicep and CLI deployment.
- Foundation models: serverless API endpoints vs managed compute; choosing models; versioning and deployment strategies; provisioned throughput units (PTUs) for high-volume workloads.
- Prompts as code: prompt design, variants compared on performance, and prompts versioned in Git.
4. GenAI quality and observability (10–15%)
- Evaluation: test datasets and data mapping; quality metrics (groundedness, relevance, coherence, fluency); risk and safety evaluations; automated evaluation workflows with built-in and custom metrics.
- Observability: continuous monitoring in Foundry; latency, throughput and response time; token consumption and cost; logging and tracing for production troubleshooting.
5. Optimising GenAI systems (10–15%)
- RAG: similarity thresholds, chunk sizes and retrieval strategies; choosing and fine-tuning embedding models; hybrid search (semantic plus keyword); relevance metrics and A/B tests.
- Fine-tuning: advanced methods, synthetic training data, monitoring tuned models, and taking a tuned model to production.
| Problem | First thing to try |
|---|---|
| Answers ignore the right document | Tune chunk size and overlap, add hybrid search, adjust the similarity threshold |
| Answers are fluent but made up | Measure groundedness; tighten retrieval and the system prompt |
| 429 throttling at peak | Provisioned throughput units, or spread load across deployments |
| Token bill doubled | Track tokens per request; trim context, cache answers, use a smaller model where quality allows |
| Model must learn a fixed output style | Fine-tune, with synthetic data if real examples are scarce |
7-week study plan
| Week | Focus | Lab |
|---|---|---|
| 1 | Azure ML workspace, datastores, compute, environments, registries | Create everything with Bicep and the Azure CLI |
| 2 | MLflow tracking, AutoML, hyperparameter sweeps, pipelines | Train and compare three runs; build a two-step pipeline |
| 3 | Model registry, managed online and batch endpoints, blue/green traffic | Deploy two versions with a 90/10 split, then roll back |
| 4 | Drift monitoring, alerts, GitHub Actions CI/CD | Pipeline that retrains when drift crosses a threshold |
| 5 | Foundry projects, model deployments, PTUs, RBAC and private networking | Deploy a model to a private Foundry project |
| 6 | Prompt variants in Git, evaluations, safety evaluations, tracing and cost | Automated evaluation run on every pull request |
| 7 | RAG tuning, hybrid search, fine-tuning; Practice Assessment | Improve a RAG app's relevance score; book the exam |
15 scenario questions with answers
These are original scenarios written for this guide, not exam questions.
1. Several teams' workspaces must reuse the same approved environment and model. What do you use?
An Azure Machine Learning registry.
2. Every workspace must be created the same way in dev, test and prod. How?
Infrastructure as code with Bicep (or the Azure CLI), run from GitHub Actions.
3. Data scientists must compare metrics from 30 training runs. What do you set up?
MLflow experiment tracking in Azure Machine Learning.
4. You need a good baseline model quickly without hand-picking algorithms. What do you use?
Automated machine learning (AutoML).
5. Ten million records are scored once a night. Which endpoint type?
A batch endpoint.
6. A new model version should take 10% of live traffic and be easy to roll back. How?
Deploy it alongside the current version on a managed online endpoint and shift traffic gradually.
7. Input data in production starts to look different from training data. What detects this?
Data drift monitoring, with an alert or retraining trigger at a threshold.
8. Foundry project traffic must never cross the public internet. What do you configure?
Private networking (private endpoints) with managed identities and RBAC.
9. A chatbot has steady, very high traffic and is hitting rate limits. What do you buy?
Provisioned throughput units (PTUs).
10. Two prompt versions must be compared and the winner tracked. How?
Create prompt variants, evaluate them on the same test dataset and keep prompts versioned in Git.
11. Which metric shows whether answers are supported by the retrieved documents?
Groundedness.
12. The app must be tested for harmful content before release. What do you run?
Risk and safety evaluations.
13. Finance wants to know why the monthly AI bill rose. What do you monitor?
Token consumption and resource usage, per application or deployment.
14. Product codes like "SKU-4471" aren't found by semantic search. What helps?
Hybrid search that combines keyword and vector retrieval.
15. There are only 50 real examples for fine-tuning. What can you add?
Synthetic data generated and reviewed for the fine-tuning set.
After AI-300
- Build on Foundry with the Microsoft Foundry tutorial and Foundry RAG with Azure AI Search.
- Compare with Google's Professional ML Engineer and see what's changed in Microsoft certifications in 2026.
- See live AI and ML jobs.
Cloud Soft Solutions in Ameerpet, Hyderabad teaches machine learning, MLOps and generative AI on Azure, AWS and Google Cloud with real-time projects. See the APEX AI, ML, Cloud and Cyber Security program and all courses.
📞 Call +91 96660 19191 · WhatsApp · Flat No. 513, 5th Floor, Nilagiri Block, Aditya Enclave, Ameerpet, Hyderabad
Frequently Asked Questions
What certification does AI-300 earn?
Microsoft Certified: Machine Learning Operations Engineer Associate.
How long is the AI-300 exam?
120 minutes. It is proctored and may include interactive components.
What is the passing score for AI-300?
700.
Does AI-300 replace DP-100?
Yes. DP-100 is retired and AI-300 is Microsoft's current exam for operating ML and generative AI solutions on Azure.
Which AI-300 skill area has the most weight?
Implement machine learning model lifecycle and operations, at 25–30%.
Which tools does AI-300 expect?
Azure Machine Learning, Microsoft Foundry, MLflow, GitHub Actions, Git, Bicep and the Azure CLI.
