AI engineer roadmap 2026: skills, certs and the order
An AI engineer roadmap should begin with programming foundations, then move through LLM apps, RAG, agents, evaluation and deployment; leave custom model training until a role needs it. An AI engineer builds applications on foundation models, while an ML engineer concentrates more on training, productionising and operating models. This guide gives you that order, small projects that prove each step, and a sensible point at which a certification helps.
You do not need to collect a certificate from every cloud to follow this path. Choose the platform you can use for a real project, learn the transferable ideas first, then take an exam that matches the work you can already do. The route below is deliberately narrower than a giant skill map: it is for building useful AI applications before specialising in model research.
AI engineer vs ML engineer
The titles overlap, but the starting problem is different. An AI engineer uses foundation-model services to build, evaluate and deploy an application: an assistant, a retrieval system or a tool-using workflow. An ML engineer works further down the model lifecycle, including data preparation, training, productionisation and monitoring of machine-learning models.
That distinction changes the order you learn. If you want to deliver an application that answers questions from changing documents, start with APIs, retrieval and evaluation. If your target role is to train a model, operate training pipelines or detect production drift, you will need the ML and MLOps branch sooner; best MLOps certifications explains that branch in more depth.
There is still shared ground. Both paths need clear code, data handling, deployment discipline and monitoring. The difference is not a status ladder where one title is automatically more advanced; it is the kind of system you are responsible for making reliable.
AI engineer roadmap in six stages
1. Build the programming and data foundation
Learn Python first, then the practical data habits around it: reading inputs, handling failures, keeping configuration separate from code and working with the data your application needs. You do not need to begin with advanced mathematics or distributed training to write and test a small application. The purpose of this stage is to make the rest of the roadmap reproducible rather than dependent on a notebook that only works once.
Build a small Python program that accepts a request, validates it, calls a local function and writes a useful result with an error message when the input is bad. You are ready for the next stage when you can explain the inputs, outputs and failure path, and can change the program without breaking its basic behaviour. This is also the right time to learn only enough cloud identity and configuration to run your own project safely.
2. Make a focused LLM application
Learn how an LLM API, instructions and supplied context shape an answer. Treat prompting as one part of application design, alongside input validation, a defined task and an output your code can use. Avoid trying to solve every task with a general chatbot; a narrow application gives you a clearer way to tell whether a change helped.
Build a small assistant that turns a limited type of input into a repeatable output, such as a structured summary or a draft based only on text you supply. You are ready for retrieval when you can vary the instructions and context deliberately, save representative inputs and explain why an answer changed. If the failure is a stable response style rather than missing information, a practice question on fine-tuning output behaviour helps you separate fine-tuning from prompt changes.
3. Add retrieval before changing the model
Learn embeddings, semantic search and retrieval-augmented generation, or RAG. Retrieval lets an application fetch relevant, changing source material before it answers; it is usually the first answer when the knowledge is not fixed in the model. Start with the concepts in what RAG is and how it works, then compare the choices in RAG versus fine-tuning.
Build a cited question-and-answer feature over a small, changing set of documents. You are ready for agents when you can show which retrieved passages support an answer, update a document and see the answer reflect that change. Test the retrieval decision with a practice question on embeddings for semantic search and a practice question on RAG for changing internal documents.
4. Add one tool-using workflow
Learn agents only after you have a bounded application and a retrieval path to build on. In this context, an agent is a workflow that chooses or invokes a tool to make progress, not a reason to give an assistant unlimited actions. Define what the tool can do, what it returns and when the application should stop instead of guessing.
Build a narrow assistant that selects one retrieval or lookup tool and returns a cited result. You are ready for systematic evaluation when the workflow can handle expected tool failures and you can trace the route it took for a small set of requests. A practice question on selecting an Agent Bricks option is useful for checking whether an application need calls for an agent-building option rather than an unrelated service.
5. Evaluate answers before scaling them
Learn to write representative tasks, define the behaviour that matters and inspect failures. Evaluation is not a final polish: it tells you whether retrieval, instructions or a tool change made the application better. For a grounded RAG application, check whether the answer is supported by the retrieved material as well as whether it sounds helpful.
Build a small evaluation set with expected properties for your application, then compare a change against it. You are ready to deploy when you can name the failures you are watching and explain why the evaluation method fits them. Use a practice question on LLM-as-a-judge evaluation to think about judging helpfulness at scale, and a practice question on a groundedness evaluator when fabricated figures are the risk.
6. Deploy, monitor and govern the application
Learn the operational layer last because it has something real to operate. That includes deploying the application, observing its behaviour, controlling access, recording the version of the model and retrieval assets, and deciding how you will respond when quality changes. Governance is part of this stage: define who may use data and tools, not only how the application returns an answer.
Build a deployed version of your small application with a way to inspect requests and outcomes, then make a deliberate change and check the result. You are ready to specialise when you can connect an observed production problem to a corrective action, rather than only redeploying. A practice question on training-serving skew detection is a useful bridge into the ML-engineering side of production monitoring.
Which certification fits which stage
There is no single AI engineer certification. Choose one after the stage it measures is already part of a project, and choose the vendor that matches the platform you expect to use. For a wider comparison across vendors, see best AI certifications.
AWS Certified AI Practitioner, AIF-C01, is the foundation option in this guide (Microsoft's equivalent is AI-901, Azure AI Fundamentals). As of November 2026, it costs US$100 and has 65 questions in 90 minutes. It fits after the first two stages, when you can discuss AI, ML and generative-AI concepts rather than just repeat terms; use the AIF-C01 cert page for a timed test and the AIF-C01 questions to revise individual topics.
Microsoft AI-103 fits the application, RAG, agent and evaluation stages on Azure. As of November 2026, it is a 120-minute exam with a passing score of 700 out of 1000, priced by country: US$165 in the United States and ₹4,865 in India. It succeeded AI-102, which retired 30 June 2026; prepare with the current path in Microsoft's AI certification path, then use the AI-103 cert page and AI-103 questions. A practice question on fine-tuning versus RAG is a good check that you can choose a technique from the behaviour of the data, not from a label.
Databricks Generative AI Engineer Associate is the fit when your RAG and LLM-chain work belongs on Databricks. As of November 2026, it costs US$200 and has 45 questions in 90 minutes. Build a retrieval application before treating it as a starting credential; the Databricks Generative AI Engineer Associate guide adds the platform detail, while the Databricks cert page and Databricks questions provide the practice here.
Google Cloud Professional Machine Learning Engineer, PMLE, fits the later ML-engineering and production branch. As of November 2026, it costs US$200 and has 50-60 questions in 120 minutes, so it is not the default first exam for an application builder. Take it when you can reason about evaluating and productionising AI solutions; use the PMLE cert page and PMLE questions after working through the relevant stages.
AWS Certified Generative AI Developer – Professional, AIP-C01, is a professional-level option for people with production AWS application experience. As of November 2026, it costs US$300 and has 75 questions in 180 minutes; AWS describes two or more years of experience building production applications on AWS or with open-source technologies as the expected background. It is not a beginner substitute for AIF-C01; see the AWS Generative AI Developer Professional guide when that experience describes you.
Cost and preparation
Use free official material before paying for an exam. AWS lists its preparation resources on the AI Practitioner certification page; Microsoft has an AI engineer career path on Microsoft Learn; Databricks offers Free Edition and an AI Prep Guide through its Generative AI Engineer Associate page; and Google points candidates to Skills Boost and a free tier through its PMLE certification page.
Treat the practice tests here as a check after the official material and your build, not as a replacement for either. Work an explanation, return to the stage where the concept belongs, then change your project or notes until you can explain the decision without memorising an answer. That approach is especially useful when the same question is testing architecture, data freshness and evaluation at once.
Coming from software or data engineering
If you are a software engineer, begin with the application path rather than restarting with model training. Your advantage is already in shaping a small service, handling inputs and failures, and deploying code; add LLM APIs, retrieval, evaluation and the safety boundary around tools. Use the six-stage project to translate those existing skills into evidence that you can build an AI application, and read how AI is changing software jobs for the wider preparation question.
If you are a data engineer, begin with the quality and movement of the source material, then make the retrieval application visible to a user. Your advantage is in data pipelines and reliable inputs; add context design, answer evaluation and the application layer before moving to ML production concerns. Choose the ML-engineer branch when you want responsibility for training, model lifecycle and production monitoring, rather than because the title sounds broader.
In either route, keep the first project small enough that you can explain every decision. A cited document-answering feature with an evaluation set is stronger learning evidence than an interface with many unnamed tools. It also gives you a concrete way to decide whether a platform-specific certification is actually the next step.
What to skip or leave for later
Do not start by training a foundation model from scratch, building distributed training systems or optimising GPU infrastructure. Those are specialist needs, and they hide the application skills you need first: a clear task, fresh context, a bounded tool, evaluation and a deployment path. Defer multimodal work and deep fine-tuning until you can identify a project failure that they solve better than retrieval or a better application design.
Skip stale certification advice. AI-102 retired 30 June 2026; AI-900, DP-100 and AZ-204 (31 July) retired in 2026; and MLA-C01 in English retired 28 September 2026. MLA-C02 is beta, so do not present it as a generally available starting point; it belongs only in a later ML-engineering decision.
Also avoid choosing a new project around a service closed to new customers, such as Amazon Q Business or Bedrock Agents Classic. The roadmap is about portable capabilities, so a standard LLM API, retrieval store, evaluation set and deployment practice will outlast a product name. Stop adding tools when the existing project cannot yet show a useful, tested result.
How long it takes
There is no vendor-supported timetable for becoming an AI engineer, and a fixed calendar promise would obscure the work that matters. Progress through the stages when the build and readiness check are true, not when a course playlist ends. A software or data background may change where you start, but it does not remove the need to test retrieval, tools and answers in a real project.
Use certification preparation as a pacing tool only after you have reached its stage. If an exam objective exposes a gap, return to the smallest build that demonstrates the gap and repair it. That makes the path useful whether you are studying alongside existing work or moving between software, data and ML responsibilities.
Frequently asked questions
Can I complete this roadmap part-time alongside work or studies?
Yes, if you use the stages as checkpoints instead of trying to learn every framework at once. Keep one small project active and finish its readiness check before adding the next capability. The important evidence is a working, evaluated application, not the speed at which you copy a curriculum.
Do I need a degree or advanced math to start learning AI engineering?
You can start with programming, data handling and a narrowly defined application. The early stages are about using and evaluating foundation-model applications, not deriving a training algorithm. Advanced ML and model-training work becomes relevant if you choose the ML engineer branch later.
What project should I build for an AI engineer portfolio?
Build a small cited question-and-answer application over changing documents, then add one bounded tool and an evaluation set. It demonstrates the progression from embeddings and retrieval to agent behaviour and groundedness. Describe the inputs, source updates, failure cases and the way you check output quality; those details make the project understandable.
What is the difference between RAG and fine-tuning?
RAG retrieves current source material at answer time, so it fits changing knowledge and can show citations. Fine-tuning changes model behaviour and can fit a stable style or output requirement. Start with retrieval when freshness or evidence is the problem, then use the project's evaluation results to justify a different approach.
Will this roadmap help me get an AI engineer job?
It gives you a sequence for building and explaining the capabilities the role uses, but it does not guarantee a job. Pair a relevant certification with a project you can walk through: its purpose, retrieval choices, tool boundary, evaluation and deployment decisions. That is more useful than claiming familiarity with every AI engineering tool.