Microsoft retired the DP-100 and the Azure Data Scientist Associate certification on the first of June 2026, and AI-300 leading to the Machine Learning Operations Engineer Associate credential is the replacement. That is the first thing to know, because a lot of study material, forum advice and third party course content still targets DP-100 and it is now for a certification you cannot earn. The second thing to know is that the change in scope is substantial and, in my view, an improvement. The DP-100 tested designing and implementing data science solutions: exploring data, training models, evaluating them, deploying them.
AI-300 tests operations. The five assessed domains are designing and implementing MLOps infrastructure, implementing machine learning model lifecycle and operations, designing and implementing GenAIOps infrastructure, implementing generative AI quality assurance and observability, and optimising generative AI systems and model performance. The exam runs a hundred and twenty minutes, is proctored through Pearson Vue, and expects Python fluency, entry level DevOps practice with GitHub Actions and command line tooling, and infrastructure as code with Bicep and the Azure CLI. Azure Machine Learning and Microsoft Foundry are the two product surfaces.
I rate the redesign because it tests the thing that actually goes wrong. Training a model is the easy part and always has been. What breaks is everything after: the model that silently drifts, the pipeline that fails at three in the morning, the prompt change nobody evaluated, the generative feature that started hallucinating after a model version bump and took a fortnight to notice. A certification that puts observability, drift detection, lifecycle governance and generative AI quality assurance at its centre is testing the real job.
The inclusion of GenAIOps as a first class domain is the part I would highlight. Evaluation and observability for generative systems is the least mature discipline in the whole field and the one where most teams are improvising. Having a major vendor certification assert that this is a defined skill set with defined practices is useful even if you never sit the exam, because the study guide gives you a checklist of things you probably are not doing. Preparation is free, which is worth saying clearly.
Microsoft Learn hosts the learning paths, the study guide is public, there is an exam sandbox to see the interface, and a practice assessment sits on AI Skills Navigator. The only unavoidable cost is the exam fee, around a hundred and sixty five dollars in most regions, and whatever you spend on Azure while practising. Now the criticisms. The first is the obvious one for any vendor certification: it is entirely Azure specific.
Bicep, Azure Machine Learning, Microsoft Foundry, Azure CLI. The underlying concepts of MLOps transfer, and the specific skills do not. If your employer runs on AWS or Google Cloud this is the wrong certification and you should look at the AWS Machine Learning Engineer Associate or the Google Professional Machine Learning Engineer instead. Only take a cloud certification for the cloud you actually use.
Second, practising properly requires a live Azure subscription. Free tier credits will cover some of it, and anyone who wants real experience with pipelines, endpoints and monitoring will spend money. Budget for it and set spending alerts, because a forgotten compute instance is a classic and expensive mistake. Third, Microsoft Foundry is new.
That means the study material is thinner than for mature exams, community answers to specific questions are scarce, and the product itself is changing under you. Expect some frustration and expect the exam content to shift. Fourth, the twelve month expiry and annual renewal. Microsoft's renewal process is a free online assessment rather than a resit, which is reasonable, and it is still a recurring obligation on your calendar for as long as you want the credential to be current.
Fifth, and this applies to every certification in this catalogue, passing an exam is not the same as being able to do the job. I have interviewed certified people who could not explain how they would debug a failing deployment. The certification is a credible signal that you have studied a defined body of operational knowledge. It is not evidence that you have operated anything.
Build and run something real alongside the study, and talk about that in interviews rather than the badge. On who should take it. If you work in an Azure shop, particularly one moving generative AI features towards production, this is a good use of six to ten weeks and the low current competition on a new credential is a genuine if temporary advantage. If you are a career changer with no production experience, this is too advanced and too narrow, and you would be better served by fundamentals.
My three point nine reflects a well scoped exam that tests the right things, discounted for the vendor lock in, the practice costs, and the immaturity of the material around a product that is still moving.