AI-300 study scenarios: example profiles
How to study for AI-300 depends on which of its two halves you already work in. The exam is 120 minutes, English only, with a scaled pass mark of 700 out of 1000, and it splits into Azure Machine Learning (the first two domains, 40 to 50%) and generative AI operations in Microsoft Foundry (the last three, 40 to 55%). Nobody passes on one half alone, so each profile below is really a question of which half is the gap.
Example profiles, not real people. The profiles below are illustrative planning examples built from the official exam outline. They are not reviews or reports of anyone’s result.
The study plan assumes a candidate who already trains models in Azure Machine Learning, with four weeks at 8 to 10 hours, about 32 to 40 hours. It also says to plan six to eight weeks without Azure Machine Learning or GitHub Actions experience.
The data scientist who trained models for DP-100
Starting point. Comfortable with experiments, MLflow tracking, hyperparameter sweeps and registering models. Has seen less of GitHub Actions, Bicep and federated authentication, and little of Foundry.
Where the time goes. The model lifecycle domain (25 to 30%) is largely familiar, so the hours move to the three generative AI domains: GenAIOps infrastructure (20 to 25%), quality assurance and observability (10 to 15%) and optimisation (10 to 15%). Within the MLOps infrastructure domain (15 to 20%), infrastructure as code is the new part.
The trap. Spending a week polishing training topics because they feel productive. Deployment, progressive rollout, rollback and drift monitoring sit inside the lifecycle domain and are operations questions, not training questions.
Order. MLOps infrastructure with Bicep and GitHub Actions, a short lifecycle refresher centred on endpoints and monitoring, then the three generative AI domains.
Planning assumption. A reasonable planning assumption is 34 hours: 7 on infrastructure, 6 on lifecycle, 9 on GenAIOps infrastructure, 5 on evaluation and observability, 4 on optimisation, and 3 for practice.
The DevOps engineer who has never trained a model
Starting point. Writes workflows, deploys Bicep and thinks in pipelines. Machine learning vocabulary (experiments, components, data assets, drift) is unfamiliar.
Where the time goes. The model lifecycle domain, because it is the largest single area and the one with the most sub-objectives: training orchestration, registration, deployment and monitoring. The outline’s responsible AI evaluation of a model is also new ground.
The trap. Assuming the infrastructure domains will carry the score. The two Azure Machine Learning domains are worth up to half the paper, and a workflow that authenticates correctly does not answer questions about sweep jobs or feature retrieval specifications.
Order. Workspace resources and assets, the lifecycle domain end to end, then Foundry deployment, then evaluation.
Planning assumption. A reasonable planning assumption is six to eight weeks, or about 50 to 65 hours, with around a third on the lifecycle domain and the rest split by the published weights.
The generative AI app developer coming from AI-103
Starting point. Has built agents and RAG on Foundry, understands evaluators and content filters, and has deployed models. Has not trained or registered a classical model, or run a drift monitor.
Where the time goes. The first two domains, which together are worth 40 to 50% and are the half they have not touched. The generative AI domains reward what they already know, with new items in provisioned throughput, prompt versioning and fine-tuning management.
The trap. Reading the generative half as a repeat of the apps exam. This one asks how to version prompts in Git, run evaluation in a workflow and tune retrieval thresholds, which is operations, not building.
Order. Azure Machine Learning workspace and lifecycle first, then a fast pass over Foundry operations with attention to what is new.
Planning assumption. A reasonable planning assumption is 36 to 40 hours, with about 60% on the two Azure Machine Learning domains.
The three AI-300 profiles compared
| Profile | Front-loaded domains | Main risk |
|---|---|---|
| DP-100 data scientist | GenAIOps infrastructure, IaC | Over-studying familiar training topics |
| DevOps engineer | Model lifecycle | Expecting infrastructure skill to carry the score |
| AI-103 app developer | MLOps infrastructure, lifecycle | Treating the generative half as an apps exam |
Which AI-300 profile fits before you book
Sketch from memory how a model gets from a dev workspace to production without retraining, then say which evaluator measures groundedness and what data it needs. The first question tests the Azure Machine Learning half and the second the Foundry half.
Whichever answer is thin points to your profile. Check the exam format for the five weights, try the sample questions, and take a timed practice test. Former DP-100 candidates should read AI-300 versus DP-100, and AI-103 candidates can compare against the AI-103 hub.