AI-300 Certification Guide 2026: Machine Learning Operations Engineer Associate

Microsoft retired the DP-100 exam and the Azure Data Scientist Associate certification on June 1, 2026 — including the renewal assessment, so the credential can no longer be earned or renewed. The replacement is the AI-300 certification: passing Operationalizing Machine Learning and Generative AI Solutions earns a new credential, Microsoft Certified: Machine Learning Operations Engineer Associate. The exam runs 120 minutes, requires a 700/1000 score, and tests five domains split roughly half-and-half between classic MLOps on Azure Machine Learning and generative AI operations on Microsoft Foundry.

This guide covers what AI-300 actually tests (from the official study guide), what the DP-100 retirement means if you hold or were preparing for the old credential, and how to prepare for an exam new enough that no official practice assessment exists yet.

What Is the AI-300 Certification?

AI-300 is Microsoft's intermediate exam for engineers who run machine learning and generative AI systems in production on Azure. Passing it earns Microsoft Certified: Machine Learning Operations Engineer Associate — the successor credential to the retired Azure Data Scientist Associate.

The framing shift starts with the name: "Data Scientist" is gone. Microsoft's audience profile describes MLOps and generative AI operations (GenAIOps) together as AI operations (AIOps), and expects candidates to carry both: training, deploying, and maintaining traditional models with Azure Machine Learning, plus deploying, evaluating, and optimizing generative AI applications and agents with Microsoft Foundry. Assumed skills are a data science background with Python, an entry-level grasp of DevOps practice — GitHub Actions, command-line tooling — and infrastructure as code with Bicep and the Azure CLI. The role tag on the certification page is AI Engineer, not Data Scientist.

The launch-window caveat is the same one early AI-200 candidates hit: no practice assessment, English only, and Microsoft Learn training paths still being assembled. You're preparing from the study guide and hands-on work, which cuts both ways — harder prep, scarcer credential.

DP-100 Is Retired: What Happens to Azure Data Scientist Associate

DP-100 and the Azure Data Scientist Associate certification retired on June 1, 2026, along with the renewal assessment. The certification page — now hidden from search — carries the warning: "This certification and the renewal assessment are retired."

The renewal detail is the part that matters if you hold the credential. Microsoft associate certifications normally renew annually through a free online assessment, but that assessment was switched off with the exam. Your Azure Data Scientist Associate badge stays on your transcript until its individual expiration date, then lapses with no way to extend it. There's no automatic conversion to the new credential — the forward path is sitting AI-300. This is the fourth time Microsoft has used this pattern in 2026, after AI-102's retirement in June, AZ-204's in July, and the announced AZ-500 retirement in August.

It's not just a Microsoft story either. AWS retired its Machine Learning – Specialty exam on March 31, 2026 — with no direct successor — and pointed candidates at an operations-focused associate cert instead. Both clouds killed their data-scientist exams within about ninety days of each other, and both replaced them with credentials about running models rather than designing them. The certs are following the job market: teams increasingly consume foundation models and operationalize them, rather than training from scratch.

If you were mid-preparation for DP-100, the situation is more abrupt than the AZ-204 crowd faced — there's no runway, the exam is simply gone. The consolation is that your effort wasn't wasted: DP-100's training, pipeline, and deployment material maps almost directly onto AI-300's largest domain. What you're missing is the generative AI half, and that's the half worth learning regardless of the exam.

AI-300 Exam Domains and Weights

The official study guide defines five domains. Weights are ranges, but the shape is clear: the two classic MLOps domains carry 40–50% of the exam, and the three generative AI domains carry 40–55% — meaning roughly half of AI-300 has no DP-100 equivalent at all.

MLOps infrastructure (15–20%)

Azure Machine Learning workspaces, datastores, compute targets, and identity and access management; data assets, environments, and components, with registries for sharing assets across workspaces. Then the part DP-100 never tested: deploying all of it as code — Bicep templates and Azure CLI, resource provisioning automated through GitHub Actions, network access restrictions, and Git-based source control for ML projects. This is a DevOps domain wearing an ML hat, and in my experience it's exactly where data-science-background candidates are weakest. If you've never written a Bicep template, start here.

ML model lifecycle and operations (25–30%)

The largest domain and the recognizable core of old DP-100 ground, reframed for operations: MLflow experiment tracking, automated ML, hyperparameter tuning, distributed training for large models, and training pipelines; model registration and versioning, including packaging a feature retrieval specification with the artifact and evaluating models against responsible AI principles; deployment to real-time and batch managed endpoints; and production monitoring — data drift detection, performance metrics, and retraining or alert triggers.

Two bullets stand out as things people usually learn from incidents rather than courses: "progressive rollout and safe rollback strategies" and drift-triggered retraining. Nobody wires up drift monitoring until the first time a model degrades silently for six weeks. The exam is effectively codifying post-incident checklists, which tells you what Microsoft thinks this role is for.

GenAIOps infrastructure (20–25%)

Microsoft Foundry environments and project configuration, RBAC with managed identities, private networking, and Bicep-based deployment; foundation model deployment choices — serverless API endpoints versus managed compute, and provisioned throughput units for high-volume workloads; and prompt management treated as a software discipline: designing prompt variants, comparing their performance, and version-controlling prompts in Git. That last block deserves to be taken seriously — prompts-in-Git with measured comparisons between variants is how production teams actually stop regressions, and it's now exam material.

Generative AI quality assurance and observability (10–15%)

Building test datasets and data mappings, implementing AI quality metrics — groundedness, relevance, coherence, fluency — configuring risk and safety evaluations for harmful content, and automating evaluation workflows. On the observability side: continuous monitoring in Foundry, latency and throughput metrics, tracing and debugging, and tracking cost metrics including token consumption.

The token-cost bullet is one I'd underline. I've had a production generation pipeline starve because of token limits nobody was watching — the symptom was silent output degradation, not an error. Cost and consumption observability for LLM workloads is the checklist item that catches that class of failure, and it's the kind of scenario question this domain is built for.

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

Retrieval-augmented generation tuning — similarity thresholds, chunk sizes, retrieval strategies, embedding model selection and fine-tuning, hybrid semantic-plus-keyword search, and A/B testing with relevance metrics — plus advanced fine-tuning methods, synthetic data generation, and managing fine-tuned models from development to production. If you want grounding on the retrieval side before drilling the Azure specifics, our vector database comparison covers the trade-offs these bullets assume you already understand.

AI-300 vs DP-100: What Actually Changed

DP-100 certified a data scientist who designs and trains models; AI-300 certifies an engineer who runs them in production — and roughly half the new exam covers generative AI operations that DP-100 only gestured at with a late-addition language-model domain.

What was cut is as telling as what was added: exploratory data analysis and model design — the "which algorithm and why" material that defined DP-100 — are gone. What replaced them is evaluation pipelines, deployment safety, cost observability, and RAG tuning. Microsoft is betting the scarce skill in 2026 isn't training a model; it's keeping an AI system healthy after it ships. Having spent the last couple of years operating AI features in production, I think that bet is correct.

How to Prepare for the AI-300 Certification

With no official practice assessment until after GA, hands-on work is the only reliable readiness signal. The exam's structure suggests a preparation sequence:

1. Deploy the infrastructure as code from day one. Stand up an Azure Machine Learning workspace with Bicep and the Azure CLI, wire provisioning into a GitHub Actions workflow, and restrict network access. Doing this once covers most of domain one and builds the habit the rest of the exam assumes.
2. Take one model through the full lifecycle. Track experiments with MLflow, tune hyperparameters, build a training pipeline, register the model with a feature retrieval spec, deploy it to a managed endpoint with a progressive rollout, then configure drift detection with a retraining trigger. That single chain touches nearly every bullet in the largest domain.
3. Repeat the loop for a generative AI app on Foundry. Deploy a foundation model on a serverless endpoint, version two prompt variants in Git, run an automated evaluation comparing them on groundedness and relevance, and stand up tracing plus a token-consumption dashboard.
4. Build and tune one RAG system end to end. Vary chunk sizes and similarity thresholds, add hybrid search, and A/B the configurations with relevance metrics. This is domain five in miniature.
5. Walk the exam sandbox before test day — AI-300 is flagged as potentially including interactive components, and Microsoft's newer formats reward interface familiarity.

Where to focus depends on which side you're coming from. Data scientists: your gaps are domains one and three — the Bicep, GitHub Actions, and networking material. Platform and DevOps engineers: the mechanics will feel familiar, but the evaluation-metric and RAG-optimization content in domains four and five is specialized knowledge you can't improvise on exam day.

Where AI-300 Fits in the 2026 Azure Path

Microsoft's 2026 refresh now has a coherent shape. AI-901 covers fundamentals, then the associate level forks by role: AI-103 if you build the AI application layer, AI-200 if you build the services underneath it, and AI-300 if you operate the whole thing in production. SC-500 extends the track into security. Build, run, secure — pick the fork that matches your day job, not the one with the most familiar exam code. (Data engineers get their own new track: DP-750 covers Azure Databricks and Unity Catalog, and is a different role entirely.)

Every domain on AI-300 is learn-by-doing material — you can't read your way to safe rollout strategies or token-cost debugging. CloudaQube's hands-on AI labs put you in real cloud environments building the same training pipelines, RAG systems, and evaluation workflows this exam tests, which beats discovering your gaps in a proctored session with no practice assessment to warn you first.