TL;DR: There is no single "AI engineer certification." As of September 2026 the credible options split into four groups: foundational exams (AWS AI Practitioner, NVIDIA NCA-GENL), cloud role exams tied to one vendor's stack (AWS ML Engineer Associate, Google Professional ML Engineer, Microsoft's new Azure AI Apps and Agents Developer Associate), a platform exam (Databricks Generative AI Engineer Associate), and vendor-neutral professional exams on LLM engineering itself (NVIDIA NCP-GENL and NCP-AAI). Course certificates such as IBM's on Coursera are a fifth category with no proctored exam. The right pick is the one that matches the stack you will deploy on, plus one that tests the LLM engineering directly. Every fact below was checked on the vendor's own page; the year's big change is that Microsoft retired AI-102 and AWS is mid-transition to MLA-C02.
Picture a developer who has just finished building her first retrieval pipeline and wants a credential to put next to it. She searches for "AI engineer certification" and finds an AI Overview, a Coursera listing, an IBM certificate, three YouTube videos, and no clear answer, because the market is genuinely fragmented. Each vendor certifies its own stack, the vendor-neutral options are new, and two of the best-known exams changed or retired this year.
This guide sorts the options by what they test and who they are for, with the costs and formats as they appeared on official pages in September 2026. It ends with a decision by situation, because the honest answer to "which one" depends on where you will deploy.
Start here
A certification is a resume-screen signal, and the artifacts you have built are what interviews test. If you have not built the systems yet, the AI engineer roadmap is the sequence, and this guide tells you which exam to take after each stage.
What a certification does and does not do for an AI engineer
A certification gets you through resume screens more often, especially at larger companies and for roles that name a vendor stack. It forces breadth across a domain you might otherwise sample lightly, which is its real educational value. And it gives a hiring manager a floor: this candidate at least knows the vocabulary and the reference architecture.
It does not carry you through a technical interview, where you are asked to walk through a system you built and defend its decisions. It does not substitute for having served a model on a GPU or evaluated a retrieval pipeline. And it does not transfer between stacks: an AWS credential tells a Google Cloud shop little.
That is why the recommendation at the end pairs two credentials rather than picking one, and why every exam below is placed after a stage of building rather than before it.
The 2026 landscape in one table
AI engineer certifications, as of September 2026
| Credential | Vendor | Cost | Format | Valid | Best for |
|---|---|---|---|---|---|
| AI Practitioner (AIF-C01) | AWS | $100 | 65 questions, 90 min | 3 years | Foundational, any role touching AI on AWS |
| NCA-GENL | NVIDIA | $125 | 50-60 questions, 60 min | 2 years | Foundational LLM engineering, stack-agnostic |
| ML Engineer Associate (MLA-C01 to C02) | AWS | $150 (C02 beta $75) | 65 questions, 130 min (C02: 85, 170 min) | 3 years | Building and operating ML on AWS |
| Professional ML Engineer | Google Cloud | $200 | 50-60 questions, 2 hours | 2 years | ML and GenAI on Google Cloud |
| Azure AI Apps and Agents Developer Associate (AI-103) | Microsoft | By region | 120 min | Annual renewal | GenAI and agents on Microsoft Foundry |
| Generative AI Engineer Associate | Databricks | $200 | 45 scored questions, 90 min | 2 years | LLM apps on the Databricks platform |
| NCP-GENL | NVIDIA | $200 | 60-70 questions, 120 min | 2 years | Professional LLM engineering: fine-tuning, optimization, serving |
| NCP-AAI | NVIDIA | $200 | 60-70 questions, 120 min | 2 years | Production agents: tools, orchestration, evaluation |
| AI Engineering Professional Certificate | IBM via Coursera | Subscription | 13 courses, about 4 months | No expiry | Structured coursework, no proctored exam |
Foundational exams
AWS Certified AI Practitioner (AIF-C01). Sixty-five questions in 90 minutes, $100, delivered through Pearson VUE at a center or online, valid three years. It covers AI and ML concepts, generative AI fundamentals, foundation model use on AWS, and responsible AI at a level a product manager could pass. For an engineer it is a warm-up rather than a destination; take it only if your employer values AWS badges or you want a gentle first exam.
NVIDIA Certified Associate: Generative AI LLMs (NCA-GENL). Fifty to sixty multiple-choice questions in one hour, remotely proctored, $125, valid two years. NVIDIA's listed topics run from machine learning fundamentals and prompt engineering through data preprocessing, experiment design, Python libraries for LLMs, and integration and deployment, which is the breadth of the first four stages of the roadmap. It is the better foundational pick for an engineer because it tests the model work rather than one cloud's product names, and the NCA-GENL practice tests on Preporato are built to the same blueprint.
Cloud role exams
These are worth taking when you will deploy on that cloud, and only then.
AWS Certified Machine Learning Engineer, Associate. The current MLA-C01 exam is 65 questions in 130 minutes, $150, valid three years, and results are reported on a 100 to 1,000 scale with 720 to pass. It covers data preparation, model training and evaluation in SageMaker, deployment, and MLOps practices such as registries and drift detection. The transition matters if you are planning to sit it: registration for the updated MLA-C02 opened September 1, 2026, at a beta price of $75 with 85 questions in 170 minutes, and the last day to take MLA-C01 in English is September 28, 2026. AWS recommends at least a year of hands-on experience with SageMaker AI and Amazon Bedrock for the updated exam. If you can wait for the C02 results window, the beta is the cheaper route to the same credential.
Google Cloud Professional Machine Learning Engineer. Fifty to sixty multiple-choice and multiple-select questions in two hours, $200 plus tax, available in English and Japanese, valid two years under Google's professional certification terms. Google recommends three or more years of industry experience including a year on Google Cloud. The exam was updated in 2026 to reflect the transition from Vertex AI to the Gemini Enterprise Agent Platform and to prioritize Google Cloud native solutions, so study from the current exam guide rather than older material.
Microsoft Certified: Azure AI Apps and Agents Developer Associate (exam AI-103). This is the credential that replaced the long-running Azure AI Engineer Associate; Microsoft retired AI-102 and its renewal assessments on June 30, 2026. AI-103 is a 120-minute proctored exam covering planning and managing Azure AI solutions, generative AI and agentic solutions, computer vision, text analysis, and information extraction, built around Python and Microsoft Foundry. Pricing is set by the country where the exam is proctored, and Microsoft role-based certifications renew annually through a free online assessment. If you searched for "Azure AI Engineer Associate" and found guides describing AI-102, they are out of date.
The Databricks platform exam
Databricks Certified Generative AI Engineer Associate. Forty-five scored multiple-choice questions in 90 minutes, $200, valid two years, with six or more months of hands-on experience recommended. The domains are application development (30 percent), assembling and deploying apps (22 percent), design (14 percent), data preparation (14 percent), evaluation and monitoring (12 percent), and governance (8 percent), assessed through Databricks tooling such as Vector Search, Model Serving, MLflow, and Unity Catalog. It is the right exam for teams building on the lakehouse and a poor fit otherwise, because the product specifics dominate.
Vendor-neutral professional exams
These are the two credentials that test LLM engineering itself rather than a cloud's product catalog, and they are the ones this site's AI Engineer path is built around.
NVIDIA Certified Professional: Generative AI LLMs (NCP-GENL). Sixty to seventy questions in 120 minutes, remotely proctored, $200, valid two years. NVIDIA recommends two to three years of practical experience with LLMs. The published domain weights tell you what the exam values: model optimization (17 percent), GPU acceleration and optimization (14 percent), prompt engineering (13 percent), fine-tuning (13 percent), data preparation (9 percent), model deployment (9 percent), evaluation (7 percent), production monitoring and reliability (7 percent), LLM architecture (6 percent), and safety, ethics, and compliance (5 percent). Nearly half the exam is optimization, acceleration, and fine-tuning, which is exactly the senior half of the roadmap. Practice tests are on the NCP-GENL page.
NVIDIA Certified Professional: Agentic AI (NCP-AAI). Sixty to seventy questions in 120 minutes, $200, valid two years, with one to two years of experience on production agentic projects recommended. The five domains are agent design and cognition (architecture, reasoning, planning, memory, multi-agent workflows), knowledge integration and agent development (retrieval pipelines, prompt engineering, multimodal agents), NVIDIA platform implementation and deployment, evaluation and monitoring, and human, ethical, and compliance considerations. It is the only widely available exam that tests agent engineering end to end. Practice tests are on the NCP-AAI page.
Build before you book the exam
Assemble a retrieval-augmented generation pipeline with local models, then run perplexity, BLEU, and LLM-as-judge with position-bias detection on a real GPU. Both labs sit on the AI Engineer path right before the NCA-GENL and NCP-GENL checkpoints.
Course certificates
IBM AI Engineering Professional Certificate (Coursera). A 13-course series estimated at four months at ten hours a week, at an intermediate level, covering machine learning fundamentals, deep learning with Keras and PyTorch, transformers and NLP, computer vision, and retrieval-augmented generation with LangChain. It is delivered through a Coursera subscription and there is no proctored exam. Course certificates like this one signal structured effort, and hiring managers read them as coursework rather than as a credential, which is the correct reading. They are a reasonable way to learn and a weak way to differentiate.
Which certification to take, by situation
The decision by situation
| Your situation | Take first | Then | Skip |
|---|---|---|---|
| You will deploy on AWS | AIF-C01 if you want a warm-up, otherwise MLA-C02 beta | NCP-GENL for the LLM engineering depth | Google and Azure exams |
| You will deploy on Google Cloud | Professional ML Engineer (current exam guide) | NCA-GENL then NCP-GENL | AWS and Azure exams |
| You will deploy on Azure | AI-103 (Azure AI Apps and Agents Developer Associate) | NCP-AAI if you are building agents | AI-102 material, retired |
| Your team is on Databricks | Generative AI Engineer Associate | NCP-GENL | Cloud role exams unless required |
| Stack-agnostic LLM engineering | NCA-GENL | NCP-GENL, then NCP-AAI | Course certificates as a differentiator |
| Student or career changer | NCA-GENL plus the five roadmap artifacts | One cloud exam once you know your target employer | Collecting three foundational badges |
Two rules make the table work. Take the stack exam for the cloud you will actually use, because the credential does not transfer. And pair it with a vendor-neutral exam that tests the model work, because the stack exam mostly tests product names.
How to prepare without wasting the fee
Build the stage first. Each exam maps to stages of the roadmap: NCA-GENL to stages 1 to 4, NCP-GENL to stages 3 to 6, NCP-AAI to stage 7, and the cloud exams to stage 6 on that cloud. Sitting the exam after the builds turns it into a confirmation rather than a gamble.
Use the vendor's exam guide as the syllabus. The domain weights above come from the vendor pages, and they tell you where to spend time. Half of NCP-GENL is optimization and fine-tuning; a third of the Databricks exam is application development.
Take timed practice tests. The failure mode on 120-minute exams is pacing, and the fix is rehearsal. Preporato's NVIDIA tests are full length and domain-proportional, with explanations that teach the trade-off behind each answer; the free sampler on each certification page is the place to calibrate.
Watch the transition dates. In 2026 the AWS MLA-C01 to C02 change and the Microsoft AI-102 retirement caught people mid-preparation. Check the vendor page for the exam version the week you book.
Frequently asked questions
Key takeaways
0/6 completedNext steps
Decide which stack you will deploy on, then book the matching exam after the matching build. The AI Engineer path places NCA-GENL, NCP-GENL, and NCP-AAI as milestones after the relevant modules, with practice tests on each certification page. For what the role pays once the credential is on the resume, read AI engineer skills and salary.
Sources:
- AWS Certified AI Practitioner
- AWS Certified Machine Learning Engineer, Associate
- Google Cloud Professional Machine Learning Engineer and Google Cloud exam terms
- Microsoft Certified: Azure AI Apps and Agents Developer Associate and AI-102 study guide (retirement notice)
- Databricks Certified Generative AI Engineer Associate
- NVIDIA NCA-GENL, NCP-GENL, NCP-AAI
- IBM AI Engineering Professional Certificate