Is DP-100 still available to book?

No. DP-100 retired on 1 June 2026, and the Azure Data Scientist Associate certification and its renewal assessment retired with it. Microsoft lists AI-300 as the replacement. If you already hold the data scientist certification, it remains active until its expiry date, but it cannot be renewed.

Is AI-300 only about generative AI?

No. Traditional machine learning still matters: designing the MLOps infrastructure and managing the model lifecycle together make up 40 to 50% of the exam. The three generative AI areas make up the rest. Expect questions on MLflow, endpoints and drift as well as Foundry evaluation and RAG tuning.

Do I need DevOps experience before AI-300?

You need the basics. Microsoft's audience profile asks for an entry-level understanding of DevOps, including GitHub Actions and command-line tools, plus infrastructure as code with Bicep and Azure CLI. If you have never written a CI/CD workflow, budget an extra week or two to learn Git, YAML workflows and the Azure CLI first.

Should I take AI-300 or AI-103?

Choose by the work you do. AI-300 is for people who train, deploy, monitor and evaluate models and generative AI systems in production. AI-103 is for developers who build AI apps and agents with Microsoft Foundry. Data scientists and ML engineers usually fit AI-300; application developers usually fit AI-103.

Can I practise for AI-300 without paying for GPUs?

Yes, for most of it. Experiment tracking, pipelines, endpoints, Bicep, GitHub Actions, evaluation and RAG tuning all work on CPU compute or serverless model endpoints. Set compute clusters to scale to zero, turn on idle shutdown for compute instances, and delete endpoints after each lab to keep costs low.

AI-300 MLOps Engineer Certification: Guide for Indian Data Scientists

AI-300, Operationalizing Machine Learning and Generative AI Solutions, earns the Microsoft Certified: Machine Learning Operations Engineer Associate certification. It replaced DP-100, which retired on 1 June 2026. AI-300 is less about building a model and more about running it well: infrastructure as code, CI/CD with GitHub Actions, MLflow, safe deployment, drift monitoring, and GenAIOps for generative AI apps and agents in Microsoft Foundry.

At a glance

Exam: AI-300, Operationalizing Machine Learning and Generative AI Solutions
Certification: Machine Learning Operations Engineer Associate, intermediate level, renewed free every year
Replaces: DP-100, retired 1 June 2026, together with Azure Data Scientist Associate
Tools named by Microsoft: Azure Machine Learning, Microsoft Foundry, GitHub Actions, Bicep and Azure CLI
Background expected: data science, Python, and entry-level DevOps
Exam time: 100 minutes for role-based exams, or 120 if the exam contains labs
Pass mark: 700

DP-100 is retired: what changes for data scientists in India

Microsoft's June 2026 Partner Center announcements list DP-100 as retired on 1 June 2026 and replaced by AI-300 from 12 May 2026. The Azure Data Scientist Associate certification and its renewal assessment are retired as well. If you hold it, it stays active on your transcript until its expiry date, but you cannot renew it.

The bigger change is the question each exam asks. DP-100 asked whether you could do data science on Azure. AI-300 asks whether you can run machine learning and generative AI in production, repeatably and safely.

Topic	DP-100 (retired)	AI-300
Data exploration, AutoML	A full skill area	AutoML remains, inside training orchestration
Source control	One bullet on Git integration	Bicep, Azure CLI, GitHub Actions, network restrictions
Deployment	Online and batch endpoints	Endpoints plus progressive rollout and rollback
Production monitoring	Light	Data drift, performance metrics, retraining triggers
Generative AI	Prompt flow, RAG, fine-tuning	Foundry infrastructure, evaluation, observability, RAG tuning, fine-tuning

Add up the three generative AI areas in the AI-300 study guide and they come to 40 to 55% of the exam. A DP-100 holder who never deployed a language model to production has real catching up to do.

Who AI-300 suits

Microsoft's audience profile asks for a data science background with Python, an entry-level understanding of DevOps (GitHub Actions and command-line tools), and working knowledge of Azure Machine Learning, Foundry and infrastructure as code. In practice it suits:

Data scientists in global capability centres, product companies and analytics teams who are tired of handing a notebook to engineering and want to own deployment.
ML engineers and MLOps engineers who already run pipelines and want a credential that matches the job title.
DevOps and platform engineers supporting AI teams. You know pipelines; AI-300 adds the model lifecycle and evaluation side.

Pick something else first if:

You are still learning Python and basic machine learning. Build those skills before you touch MLOps.
Your company does its data science inside Microsoft Fabric rather than Azure Machine Learning. The one-day DP-604 Fabric data science course is a closer fit.
You mainly build AI apps and agents rather than operate models. Look at our AI-103 exam guide for India instead.

The five skill areas and how to practise each

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

Workspaces, datastores, compute targets and access control; data assets, environments and components; sharing assets across workspaces with registries; and infrastructure as code: GitHub integration, deploying workspaces with Bicep and Azure CLI, automating with GitHub Actions, restricting network access and managing source control.

Practise: write one Bicep file that creates a workspace in Central India and deploy it from a GitHub Actions workflow. Sign in with OpenID Connect and a federated credential rather than a stored client secret; Microsoft's guide to using GitHub Actions with Azure Machine Learning calls OIDC the more secure option.

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

The largest area. It covers MLflow experiment tracking, AutoML, notebooks, hyperparameter tuning, training scripts, distributed training, pipelines, comparing jobs, registering MLflow models, responsible AI evaluation, archiving models, real-time and batch endpoints, progressive rollout and rollback, and monitoring for data drift with retraining or alert triggers.

Start every training script with one line, so each run records what you did:

import mlflow

mlflow.autolog()  # logs parameters, metrics and the model for supported libraries

Then practise a safe rollout with the Azure Machine Learning Python SDK v2. Send a slice of live traffic to the new deployment, watch the metrics, and roll back if needed:

endpoint = ml_client.online_endpoints.get(name="loan-default-endpoint")

endpoint.traffic = {"blue": 90, "green": 10}  # canary: 10% to "green"
ml_client.online_endpoints.begin_create_or_update(endpoint).result()

endpoint.traffic = {"blue": 100, "green": 0}  # roll back if metrics worsen
ml_client.online_endpoints.begin_create_or_update(endpoint).result()

A public loan-default or customer-churn dataset is enough to practise the whole lifecycle. You do not need company data, and you should not use it in a personal lab.

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

Foundry resources and projects, managed identities and role-based access control, private networking, Bicep and Azure CLI deployments, foundation models on serverless API endpoints or managed compute, model versioning, provisioned throughput units for high-volume workloads, and prompt management: designing prompts, comparing variants and keeping prompts under version control in Git.

Practise: store your system prompts as files in a repository, change them only through pull requests, and run an evaluation before each merge. That single habit covers three exam bullets and is exactly what production teams need.

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

Test datasets and data mapping, quality metrics such as groundedness, relevance, coherence and fluency, risk and safety evaluations, automated evaluation workflows, continuous monitoring in Foundry, latency and throughput, token consumption and cost, and tracing for troubleshooting.

Practise: build a test set of questions your users would really ask. If your users write in Hindi, Tamil or another Indian language as well as English, put both in the test set; an app that scores well in English alone has not been tested for India.

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

RAG tuning (similarity thresholds, chunk sizes, retrieval strategies), choosing and fine-tuning embedding models, hybrid search, measuring improvement with relevance metrics and A/B tests, advanced fine-tuning, synthetic data for fine-tuning, and managing a fine-tuned model through to production.

Practise: change one RAG setting at a time, such as chunk size, rerun your evaluation and record the result. It is also the right way to reason through tuning scenarios in the exam: one variable, one measurement.

A lab plan that keeps your Azure bill small

Work on CPU. Most AI-300 skills do not need a GPU. Microsoft sets quotas per VM family, and GPU families (their names start with "N") often need a quota increase on new subscriptions. Request one only when a lab truly needs it, as explained in managing Azure Machine Learning quotas.
Let compute go to sleep. Turn on idle shutdown for compute instances, or schedule them to stop at, say, 11 pm IST. Set compute clusters to a minimum of zero nodes so you pay nothing when no job runs. A stopped compute instance still bills for its disk, so delete what you no longer need.
Delete endpoints after each lab. A managed online endpoint keeps its compute running, and billing, while it is deployed.
Pick your region sensibly. Central India keeps latency low, but if a VM size or model is not available or has no quota there, run that lab in another region.

A ten-week plan for working data scientists

Assumes evenings in IST and one longer weekend block. If you already run MLflow and CI/CD at work, you can merge weeks.

Week	Focus
1	Workspace, datastores, compute, data assets, environments
2	Bicep and Azure CLI; GitHub Actions with OIDC
3	MLflow tracking, AutoML, hyperparameter tuning
4	Pipelines, components, registries
5	Endpoints, safe rollout, rollback, drift monitoring
6	Foundry projects, identity, private networking
7	Foundation model deployment, versioning, prompt management in Git
8	Evaluation metrics, safety evaluations, tracing, cost tracking
9	RAG tuning, hybrid search, fine-tuning
10	Practice assessment, weak areas, book the exam

Exam day and renewal

Microsoft Learn is available during role-based exams, but the timer does not stop. Use it to confirm a CLI flag or SDK parameter, not to learn a topic from scratch.
Online from home: a walled room with a closed door, no one else entering, and a government ID whose name matches your Microsoft profile exactly. A personal laptop is safer than a work laptop, which may block the OnVUE browser.
Retakes: 24 hours after a first failure, then 14 days between later attempts.
Renewal: free, through an online assessment on Microsoft Learn before the one-year expiry. Our renewal guide for India explains the steps.

How CertTulen Academy can help

CertTulen Academy runs the AI-300 Machine Learning Operations Engineer course as live, MCT-led online training with hands-on labs covering MLOps pipelines, GitHub Actions and GenAIOps in Foundry. Professional certification courses include an exam voucher, mock practice tests and lab access; see the course page for details and INR pricing. Data science teams can book a private corporate cohort, online or in person in major Indian cities. Use our contact page or WhatsApp +91 93618 83583 to ask for a batch that suits your time zone.