What the route looks like
Most AI engineering jobs are software jobs with a model in the middle. You write the code that calls a model, feeds it the right data, checks what comes back, and keeps the whole thing running for users. So the order matters. Learn to build and ship normal software, then learn the parts that are specific to models: calling APIs, retrieval, and evaluation. Finish with one project that real people use, because that is the evidence a hiring manager can actually inspect.
No license or regulator controls the job title. BLS data on software developers, the closest occupation it tracks, says they typically hold a bachelor's degree in computer and information technology or a related field, and some employers prefer a master's. That describes the whole occupation. It is not a rule for AI roles, and many teams hire on demonstrated work.
Retrieval: the skill most AI products depend on
Most business AI apps need the model to answer from company data it was never trained on. AWS describes retrieval augmented generation (RAG) as giving a large language model external data, such as internal documents, so it has the context to answer a specific use case. The pattern has four parts: turn documents into embeddings stored in a vector database, take a user query, run a similarity search to pull relevant chunks into the prompt, then send query and context to the model.
The demo version takes an afternoon. The real work starts with what AWS lists for production systems, such as connectors to different data sources. AWS notes that building the vector database needs data cleaning, formatting and chunking, and that production systems need processing for PDFs, scanned images and presentations, so spend time there.
Evaluation: what separates engineers from tinkerers
Model output varies from run to run, so you need tests that measure quality. Anthropic's engineering team defines an eval as giving an AI system an input and applying grading logic to its output. Graders come in three kinds: code-based checks such as string matching and unit tests, model-based graders that score against a rubric, and human review. Code checks are fast and cheap but brittle. Model graders handle open-ended answers but cost more and are less consistent.
You do not need a huge test set to start. Anthropic's advice is that 20 to 50 simple tasks drawn from real failures is a good first eval set, and that tasks should be clear enough that two experts would agree on pass or fail. Build that early. Showing an eval suite in your portfolio project tells a reviewer you know how to ship model features safely.
Which steps a working developer can skip
If you already write production code, you can skip the software basics. You know version control, testing, APIs, databases and deployment, and those carry over directly. Go straight to building with model APIs, then retrieval, then evals. Your fastest path is often inside your current job: find a feature on your team where a model could help, build it, and measure it.
What you should not skip is evaluation. Developers who are new to models tend to judge output by eye and ship. That works for a demo and fails in production, where fixing one failure can quietly break another. Career changers without a coding background cannot skip the first step. Learn to program and ship small apps before touching models.
Pay
USD a year, before tax. US figures. BLS OEWS May 2025, moved to 2026 with the Employment Cost Index, for software developers: BLS has no AI series, so this is the nearest match and covers the whole occupation. Job ad rows are midpoints of pay ranges in 21 live AutomationRoles.ai ads from 11 employers on 5 November 2026; they lean toward large AI and technology employers and show base pay without stock or bonus. Refer to the AI Developers salary guide for figures from live adverts.
What you need
- ProgrammingYou need to write, test and deploy working software on your own. Most model API work happens in ordinary application code.
- Model API skillsKnow how to call a hosted model, structure prompts, handle errors and limits, and parse output into something your app can use.
- RetrievalUnderstand embeddings, vector databases, chunking and similarity search, and how to clean messy source documents before indexing them.
- EvaluationWrite automated evals with clear pass or fail criteria, and choose between code-based, model-based and human grading for each check.
- A shipped projectOne project with real users, a retrieval layer and an eval suite says more than a list of courses.
- DegreeBLS says software developers typically hold a bachelor's degree. That is an occupation-wide pattern, not a requirement set by any regulator for AI roles.
How to get there
- Months 1-6Learn software basicsLearn one language well, plus version control, testing, HTTP APIs, a database and deployment. Working developers skip this step.
- Months 6-8Build with model APIsBuild small apps that call a hosted model: a summarizer, a classifier, a tool that returns structured output. Learn prompts, errors and cost.
- Months 8-10Add retrievalBuild a question-answering app over a real document set. Spend most of your time on ingestion, cleaning and chunking.
- Months 10-11Write evaluationsStart with 20 to 50 test tasks from real failures. Mix code checks with rubric grading, and rerun the suite on every change.
- Months 11-14Ship a real projectPut one app in front of real users, log what fails, turn failures into new eval tasks, and write up what you changed.
- Month 12 onwardApply or move internallyApply to AI engineer and AI developer roles with the project as evidence. Developers can often move by leading an AI feature on their current team.
Open now on AutomationRoles.ai
Questions people ask
Do I need a machine learning background to become an AI engineer?
Not for most AI engineering roles that build on hosted models. The daily work is software: calling APIs, retrieval, evaluation and deployment. Training models from scratch is closer to machine learning engineering or research.
Do I need a degree?
No regulator requires one. BLS says software developers, the nearest occupation it tracks, typically hold a bachelor's degree in computer and information technology or a related field. Many employers weigh shipped work heavily, but a degree still helps at some companies.
What should my portfolio project include?
A working app with real users, a retrieval layer over real documents, and an eval suite with clear pass or fail tasks. Add a short write-up of what broke and how you fixed it.
How long does it take a working developer to switch?
Usually less time than a beginner, because the software foundation is already there. The work is learning model APIs, retrieval and evaluation, then shipping one project. Leading an AI feature at your current job is often the quickest route.
Will a certification get me hired?
No certification guarantees a job. Cloud AI certifications can help you learn a vendor's services and signal effort, especially if your work history is thin. A shipped project with evals usually counts for more.
How much do AI engineers earn?
BLS has no AI series. For software developers as a whole, the 2026-adjusted US median is $140,200. AI developer ads on AutomationRoles.ai show a median midpoint of $203,000, from ads that lean toward large tech employers and exclude stock and bonus.
