AI-300 study guide: domains, format and a 4-week plan

AI-300, Operationalizing Machine Learning and Generative AI Solutions, is the exam for the Microsoft Certified: Machine Learning Operations Engineer Associate certification. It tests whether you can run both traditional machine learning (with Azure Machine Learning) and generative AI apps and agents (with Microsoft Foundry) in production, using infrastructure as code, CI/CD and monitoring. You have 120 minutes, the exam is proctored, it is offered in English only, and the passing score is 700.

Microsoft lists AI-300 as the replacement for DP-100 and the Azure Data Scientist Associate certification, which retired on 1 June 2026. The emphasis has moved from building models to operating them: automation, governance, deployment, observability and optimisation.

AI-300 at a glance

Details as of September 2026 — confirm on the official exam page before booking.

Who this exam is for

Microsoft describes the candidate as an AI operations specialist who sets up infrastructure for MLOps and GenAIOps on Azure. The profile asks for a data science background, Python experience and an entry-level grasp of DevOps, including GitHub Actions and command-line tools. It names four areas of hands-on knowledge: Azure Machine Learning, Foundry, GitHub Actions, and infrastructure as code with Bicep and the Azure CLI.

There is no formal prerequisite. The honest test of readiness is whether you have deployed a model to a managed endpoint from a pipeline rather than a notebook. If you have only trained models interactively, budget extra time for the infrastructure domains.

What the exam covers

The AI-300 study guide organises the skills into five domains.

Design and implement an MLOps infrastructure (15–20%)

This covers the Azure Machine Learning workspace and its parts: datastores, compute targets, identity and access, data assets, environments, components and registries for sharing assets across workspaces. The infrastructure-as-code task asks you to deploy workspaces with Bicep and the Azure CLI, automate provisioning with GitHub Actions, connect GitHub securely, restrict network access and manage the project in Git. Know which asset type solves which reuse problem.

Implement machine learning model lifecycle and operations (25–30%)

The largest domain. Training covers MLflow experiment tracking, automated ML, notebooks, hyperparameter tuning, training scripts, distributed training, pipelines and comparing runs. Registration covers registering MLflow models, packaging a feature retrieval specification with the model, responsible AI evaluation and archiving. Deployment covers real-time and batch endpoints, troubleshooting, and progressive rollout with safe rollback. Monitoring covers data drift, production metrics and triggers for alerts or retraining.

Design and implement a GenAIOps infrastructure (20–25%)

The Foundry counterpart to the first domain, with more on models. You configure Foundry resources and projects, managed identities and role-based access, private networking, and Bicep deployments. You deploy foundation models through serverless API endpoints or managed compute, choose models for a use case, version and roll out deployments, and configure provisioned throughput for high-volume workloads. The last task treats prompts as code: designing them, comparing variants, and versioning them in Git.

Implement generative AI quality assurance and observability (10–15%)

Evaluation and monitoring for generative apps and agents. Build test datasets and map their fields to evaluators, apply quality metrics such as groundedness, relevance, coherence and fluency, run risk and safety evaluations, and automate evaluation with built-in and custom metrics. For observability, know continuous monitoring in Foundry, latency and throughput metrics, token and cost tracking, and tracing for production debugging.

Optimize generative AI systems and model performance (10–15%)

Two tasks. RAG optimisation covers similarity thresholds, chunk sizes, retrieval strategies, embedding model choice and fine-tuning, hybrid search, and measuring improvement with relevance metrics and A/B tests. Fine-tuning covers advanced methods, synthetic training data, monitoring a fine-tuned model and taking it from development to production.

A 4-week study plan

The two machine learning domains carry up to about half the marks and the three generative AI domains the rest, so the plan splits evenly.

Week 1: MLOps infrastructure. Deploy an Azure Machine Learning workspace with Bicep from a GitHub Actions workflow. Create a datastore, a compute cluster, a custom environment and a component, then share one asset through a registry. Lock the workspace behind private networking and note what broke.

Week 2: model lifecycle. Build a training pipeline that logs to MLflow, run a sweep, register the best model and deploy it to a managed online endpoint. Practise a blue-green style traffic split and a rollback. Set up drift monitoring and an alert.

Week 3: GenAIOps. Stand up a Foundry project with managed identity and private access using Bicep. Deploy a model serverlessly and another on managed compute, and compare the operational trade-offs. Put your prompts in Git, create variants and compare them with an evaluation run.

Week 4: quality, optimisation and mocks. Run quality and safety evaluations on a small RAG app, connect tracing, then change chunk size and search mode and measure the effect. Take a full-length mock, work through the explanations, repair the weakest domain, and take a second mock.

Where candidates lose marks

• Studying DP-100 material. DP-100 centred on data science tasks. AI-300 asks how you automate, secure and monitor them, and adds a full generative AI half.
• Mixing up the two platforms. Azure Machine Learning and Foundry have similar concepts (workspaces and projects, endpoints and deployments, identity and networking). Read each scenario for which platform it is on before choosing an answer.
• Confusing asset types. Data assets, datastores, environments, components and registries are distinct. Many questions describe a reuse or reproducibility problem and ask which one fixes it.
• Choosing the wrong endpoint type. Real-time versus batch, and serverless versus managed compute versus provisioned throughput, are decisions driven by latency, volume and control. Learn the deciding factor for each.
• Reading evaluation metrics loosely. Groundedness, relevance, coherence and fluency measure different things, and safety evaluators score harm rather than quality. Know which metric exposes which failure.
• Missing the observability dependency. Tracing and monitoring questions often turn on whether the telemetry resource is connected at all.

Practising for AI-300 on CertifyAIx

CertifyAIx has 571 original AI-300 questions, organised by the exam's five domains and explained option by option. See [how our practice questions are written](/blog/how-certifyaix-practice-questions-are-written).

• 3 full-length mock exams, timed to the real exam's 120 minutes
• 26 mixed domain papers and 32 topic papers of up to 25 questions each (60 minutes per 25)
• Case-study, drag-and-drop, hot area and ordering formats as well as single and multiple choice

Go to the [AI-300 practice questions](/certifications/azure-mlops-engineer-ai-300). If you are still deciding between exams, [how to choose an AI certification](/blog/how-to-choose-an-ai-certification) compares the options.

Frequently asked questions

How difficult is AI-300?

It is an associate exam with a wide scope: two platforms, infrastructure as code and generative AI evaluation. Candidates strong in data science but new to DevOps tend to find the infrastructure domains hardest.

How long should I prepare for AI-300?

Plan for five to ten weeks. The range depends mainly on how much Bicep, GitHub Actions and Foundry you already use; the modelling knowledge itself is less of a barrier than the operations work.

Is DP-100 still available?

No. DP-100 and the Azure Data Scientist Associate certification retired on 1 June 2026, and Microsoft names AI-300 as the replacement. Microsoft says certifications already earned remain valid, and on your transcript, until they expire.

Can I retake AI-300 if I fail?

Yes. Microsoft allows a retake 24 hours after the first attempt; later retakes have longer waiting periods under its retake policy.

Are CertifyAIx questions real AI-300 exam questions?

No. They are original questions written against Microsoft's published study guide, never taken from a live or past exam.

CertifyAIx is independent and not affiliated with Microsoft. Microsoft Certified: Machine Learning Operations Engineer Associate is a trademark of its owner. All CertifyAIx practice questions are original.