TL;DR: An ML engineer builds and operates models trained on the company's own data: feature pipelines, training jobs, deployment, and monitoring for drift. An AI engineer builds products on top of pretrained foundation models: prompting, retrieval, fine-tuning, agents, and serving large models within a budget. The roles share serving, evaluation, and MLOps, and that shared middle is growing. Pursue ML engineering if the company's value comes from models trained on proprietary data; pursue AI engineering if it comes from applying language models to a product. The certification blueprints for each role are the clearest objective statement of the difference, and they are quoted below.
A ride-hailing company has two teams down the hall from each other. One trains the model that predicts arrival times from millions of past trips, retrains it nightly, and watches its error creep up when a city changes its road layout. The other builds the support assistant that reads the company's policies, answers rider questions, and escalates refunds, and its main worries are retrieval quality, hallucination, latency, and the monthly inference bill. Both teams are "doing AI." Their tools, their failure modes, and their hiring bars barely overlap.
The first team is ML engineering. The second is AI engineering, in the sense the title has settled into since 2023. This guide lays the roles side by side by what each one ships, where they genuinely overlap, how to read a job posting that uses the titles loosely, and how to choose.
Start here
If you have already chosen the AI engineer side, the roadmap is the build sequence and how to become an AI engineer is the hiring side. If MLOps is the part that interests you, the MLOps course guide covers the shared middle in depth.
The one-sentence difference
An ML engineer's artifact is a trained model; an AI engineer's artifact is a system built around a model someone else trained. Everything else follows from that. The ML engineer's inputs are data and a training objective, and the hard problems are data quality, feature engineering, training infrastructure, and keeping a model accurate as the world changes. The AI engineer's inputs are a foundation model and a product requirement, and the hard problems are getting reliable behavior out of a probabilistic component, grounding it in the right information, giving it tools safely, and running it at an acceptable cost and latency.
What each role does in a typical week
A week in each role
| Activity | ML engineer | AI engineer |
|---|---|---|
| Monday | Debug a feature pipeline that silently dropped a column | Trace why the assistant cited the wrong policy document |
| Tuesday | Tune a gradient-boosted model and compare against last week | Rebuild the retrieval index with hybrid search and rerank, re-run the eval set |
| Wednesday | Ship a retrained model through the registry with a canary | Cut inference cost with prompt caching and a model-tier router |
| Thursday | Investigate drift alerts after a product change | Add a tool budget and stop condition to an agent that looped |
| Friday | Write the retraining schedule and monitoring thresholds | Run LLM-as-judge with position-bias checks on the new prompt |
Both weeks contain deployment, monitoring, and evaluation. The difference is what is being deployed and what "evaluation" means: a held-out test set with a fixed metric for the ML engineer, a labeled question set plus a judge model plus latency and cost numbers for the AI engineer.
Skills side by side
ML engineer only. Feature engineering and feature stores, classical and deep learning training (gradient-boosted trees, neural networks on tabular and time-series data), distributed training on your own data, hyperparameter search at scale, and the statistics to design a valid offline evaluation. Model architectures are chosen and trained rather than prompted.
AI engineer only. Prompt design and structured output, retrieval-augmented generation with hybrid search and reranking, parameter-efficient fine-tuning of large pretrained models (LoRA and QLoRA), agents with tool calling and orchestration, LLM inference optimization (quantization, batching, KV-cache behavior, vLLM), and defenses against prompt injection. The model is a given; the system is the work.
Shared. Python and one ML framework, containers and CI, model serving, experiment tracking and registries (MLflow), monitoring, and cost accounting. This shared set is exactly what MLOps means, which is why an MLOps skill set is the easiest bridge between the two roles.
Skills by role
| Skill area | ML engineer | AI engineer |
|---|---|---|
| Feature engineering and feature stores | Core | Rare |
| Training models on proprietary data | Core | Occasional (fine-tuning only) |
| Prompting and structured output | Rare | Core |
| Retrieval-augmented generation | Rare | Core |
| Fine-tuning large pretrained models | Occasional | Core |
| Agents and tool use | Rare | Core |
| LLM inference optimization | Rare | Core |
| Serving and deployment | Core | Core |
| Evaluation | Core (held-out metrics) | Core (eval sets, judges, latency, cost) |
| Experiment tracking, registries, monitoring | Core | Core |
What the certification blueprints say
The most objective definition of each role is what its exams test, because vendors write blueprints from job analyses.
The AWS Certified Machine Learning Engineer, Associate exam covers data preparation and transformation, model training and evaluation in SageMaker, deployment strategies, and MLOps practices such as model registries and drift detection. The Databricks Certified Machine Learning Professional exam weights model development at 44 percent, ML operations at 44 percent (testing strategies, automated retraining, drift monitoring, rollout management), and deployment at 12 percent. Those are ML engineer blueprints: data in, trained model out, operated over time.
The NVIDIA Certified Professional: Generative AI LLMs exam weights model optimization at 17 percent, GPU acceleration at 14 percent, prompt engineering and fine-tuning at 13 percent each, with the remainder across data preparation, deployment, evaluation, monitoring, architecture, and safety. The NVIDIA Certified Professional: Agentic AI exam covers agent design, knowledge integration, deployment, evaluation, and human and compliance considerations. Those are AI engineer blueprints: pretrained model in, optimized and orchestrated system out. The certification guide compares all of them with costs and formats.
Where the overlap is, and why it is growing
Three things sit in the middle, and each is becoming a larger share of both jobs.
Serving. Both roles put models behind endpoints and care about latency, throughput, and cost. The AI engineer's models are larger, which makes the serving problem harder and the skill scarcer.
Evaluation. Both roles need to know when a change made things worse. The ML engineer inherits a mature toolkit; the AI engineer is building one, with eval sets, judge models, and regression gates in CI.
MLOps. Experiment tracking, registries, reproducible pipelines, monitoring, and retraining automation apply to both. As LLM fine-tuning becomes routine, AI engineers need the ML engineer's operational discipline, and as ML teams adopt foundation models for text and vision, ML engineers need the AI engineer's system patterns.
The practical consequence is that the shared middle is the safest place to invest, whichever title you end up with.
The two skills that bridge the roles, on real GPUs
Wire MLflow experiment tracking from a single run to a registered model with stage transitions, then fine-tune Llama 3 8B with LoRA and evaluate it against the base model: the operational discipline of ML engineering applied to the AI engineer's artifact.
Reading a job posting that uses the titles loosely
Titles lag reality by a few years, so the label on a posting is weak evidence. Read the responsibilities and the stack instead.
If the posting mentions feature stores, training pipelines, gradient-boosted models, A/B testing of model versions, or drift, it is an ML engineering role regardless of the title. If it mentions RAG, vector databases, prompt engineering, fine-tuning an open model, agents, or an inference bill, it is an AI engineering role. If it mentions both at length, it is usually a small team that wants one person to do both, and you should ask in the first call which half is 80 percent of the work.
Two common mislabels: "ML engineer" postings that are entirely LLM application work (common at startups founded after 2022) and "AI engineer" postings that are classical ML with a chatbot attached (common at enterprises renaming existing teams).
Which one to pursue
Choose ML engineering if the companies you want to work for compete on models trained on proprietary data (ranking, forecasting, fraud, recommendations, pricing), you enjoy the statistics and data side, and you want a mature toolchain with established best practices.
Choose AI engineering if the products you want to build are language or agent products on top of foundation models, you are strong on the systems side (APIs, serving, cost), and you are comfortable working where the best practices are still being written. Demand for this role grew faster in 2024 to 2026, and its pay band has a higher top in both the US and India because the senior skills (serving, fine-tuning, agents) are scarce; the skills and salary guide has the sourced numbers.
Choose the overlap if you are unsure. Serving, evaluation, and MLOps are hireable on both sides and give you a year to decide with real information.
Moving between them. From ML to AI engineering, the gap is the system patterns (retrieval, tool use, agents) and LLM-specific inference; from AI to ML engineering, the gap is data and training depth and the statistics of evaluation. In both directions the shared MLOps skills carry over intact.
Frequently asked questions
Key takeaways
0/6 completedNext steps
If the AI engineering side is the one you want, start the roadmap at stage 1, or follow the same sequence with graded work on the AI Engineer path. If the shared middle is what you want to build first, the MLOps course guide lays out the loop and the five things you should have shipped by the end.
Sources:
- AWS Certified Machine Learning Engineer, Associate
- Databricks Certified Machine Learning Professional
- NVIDIA Certified Professional: Generative AI LLMs
- NVIDIA Certified Professional: Agentic AI