What carries over from software engineering
Most of the job is still software. An AI feature needs APIs, data access, auth, logging, deploys, retries and on-call support, and you already know how to do all of that. AWS's guidance on retrieval augmented generation (RAG) lists the parts of a production system: connectors to data sources, document processing, a vector database, a retriever, guardrails, an orchestrator, the user interface, and identity and access control. Only one of those parts is the model itself. The rest is ordinary engineering, and that is the experience hiring managers want to see.
Your testing habits carry over too, but they need adjusting. Anthropic describes coding-agent evals that grade the result with unit tests, which is a tool you already use. The difference is that a single passing run proves less than it does in normal code, because the same input can give a different output on the next run.
What is new: evaluation, retrieval, model behavior and cost
Evaluation is the biggest change. Anthropic's engineering team says each eval task can have its own success rate and can pass on one run and fail on the next, so teams run several trials and score the share that succeed. Graders can be code, another model, or people. Model graders handle open-ended output, but they cost more than code checks and have to be checked against human judgment. You keep two kinds of suite: capability evals, which start with a low pass rate, and regression evals, which should pass almost every time.
Retrieval is the second new skill. In AWS's description of RAG, documents are cleaned, chunked and turned into embeddings, which are numbers that represent meaning, and stored in a vector database. At query time the system pulls the closest matches and adds them to the prompt. Most answer-quality problems start here, so learn to measure retrieval separately from the model's answer.
Model behavior and cost round out the list. Read transcripts often. Anthropic says the transcript shows whether the agent made a real mistake or the grader rejected a valid answer. Cost now scales with tokens and with how many model calls each request makes. OpenAI's developer docs include sections on cost optimization, prompt caching and batch processing, which shows that cost control is part of the engineering work.
How to get AI work inside your current job
The easiest way in is a feature your team already owns. Look for a task that has a clear right answer and real users, such as search over internal docs, sorting support tickets, or pulling fields out of documents. Offer a small version to your manager with an eval plan attached. That plan is what makes the proposal credible. Anthropic says 20 to 50 simple tasks taken from real failures is enough to start, so you do not need a big dataset to begin.
Ship it internally first, then report pass rates, cost per request and latency the way you would report any production metric. An engineer who can say the feature resolves 85 percent of the eval set at a known cost per call is doing AI engineering, whatever their title says. Once you have one or two of these, ask for the title change or apply for an internal transfer.
How long the move takes
No survey measures this, so treat the following as a rough planning range. An experienced backend or full-stack engineer can usually build a working prototype within a few weeks. Learning to evaluate it honestly, tune retrieval and keep costs predictable takes longer, often several months of work on a real feature. Most people need about 6 to 18 months of shipped, measured work before they can present themselves as an AI engineer with evidence behind it.
Engineers who already work near data pipelines, search or platform infrastructure tend to move faster because more of their work carries over. Engineers whose employers have no AI work may need a side project with a public eval set, though that is weaker evidence than a feature real users depend on.
Pay
USD a year, before tax. US figures. The BLS rows come from BLS OEWS May 2025 for software developers, moved to 2026 with the Employment Cost Index. BLS has no AI series, so this is the closest series and it covers the whole occupation, not only AI specialists. The job ad rows are midpoints of the base pay ranges in 21 live AutomationRoles.ai ads from 11 employers on 5 November 2026. Those ads lean toward large AI and technology employers and leave out stock and bonus. Refer to the AI Developers salary guide for figures from live adverts.
What you need
- Production software skillsAPIs, data access, deploys, monitoring and access control make up most of an AI system. AWS's list of RAG components shows how much of the system is ordinary engineering.
- Evaluation designYou need to build task sets, graders and repeated trials for outputs that change between runs. You also need to keep capability evals and regression evals as separate suites.
- RetrievalChunking, embeddings, vector search and ranking decide what context the model gets. Learn to measure retrieval quality apart from the quality of the answer.
- Reading model behaviorRead transcripts regularly to tell real model mistakes from grader errors. Fix prompts, context or tools based on what you find.
- Cost and latency controlTrack tokens and model calls per request. Know the levers vendors document, such as prompt caching and batch processing.
- Evidence of shipped workA feature in production with reported pass rates and costs counts for more than coursework. No certification is required to do this work.
How to get there
- Weeks 1-4Build one small LLM feature end to endPick a narrow task with a checkable answer, call a hosted model through its API and get it running on real inputs.
- Month 1-2Write your first eval setCollect 20 to 50 tasks from real failures, add code or model graders, and run several trials per task.
- Months 2-4Add retrievalChunk and embed your team's documents, store them in a vector index, and measure whether the right passages come back.
- Months 3-6Ship internally and measureRelease to internal users. Report pass rate, cost per request and latency. Read transcripts every week.
- Months 6-12Own an AI feature in productionKeep a regression suite in CI, handle incidents, and use the results to ask for an AI title or an internal transfer.
- Months 9-18Apply with evidenceTake the shipped feature, eval results and cost numbers to external AI engineering roles if the internal route stalls.
Open now on AutomationRoles.ai
Questions people ask
Do I need to learn machine learning math to become an AI engineer?
For application-level work on hosted models, the daily work is evals, retrieval, prompts, tools and cost. Model math matters more for roles that train or fine-tune models. Pick the role first, then learn what it requires.
Will a certification get me hired as an AI engineer?
No certification is required, and none will get you hired by itself. A cloud certification can give your study some structure. A shipped feature with measured results is stronger evidence.
Do I need a degree?
BLS says software developers usually have a bachelor's degree in computer and information technology or a related field. We did not find a source showing that AI engineering roles add a further degree requirement.
How much does an AI engineer earn compared with a software engineer?
BLS has no AI series, so the closest figure is software developers: a 2026 US median of about $140,200, estimated from May 2025 data. AutomationRoles.ai AI developer ads had a median base pay of $203,000 across 21 ads, and those ads lean toward large AI and technology employers.
What should my first AI project be?
Pick a task with a checkable answer that people at your company already need, such as search over internal documents or extracting fields from forms. Build the eval set before you tune the prompt, so you can show that each change helped.
How do I test something that gives different answers each time?
Run each task several times and score the share of runs that pass, instead of trusting a single run. Use code graders where you can, and model graders for open-ended output, checked against human judgment.
