AI engineer vs. ML engineer: Where the work touches the model
AI engineer vs ML engineer: what each builds, the four places they overlap, which to hire first, and the signals that say it's time to add the second.
Key takeaways
- The AI engineer vs ML engineer question comes down to where the work touches the model. AI engineers build the product system around a model that already exists: prompts, retrieval, tools, evaluation, latency and cost. ML engineers work on the model itself and the machinery that trains, deploys and monitors it.
- Titles lag the work. In Pave's compensation data, 83% of employees in the AI/ML job category carry an ML title and 3% an AI title, so a posting that says "ML engineer" may describe either job.
- Most teams building features on hosted models need an AI engineer first. A team with proprietary data and a prediction problem no off-the-shelf model solves needs an ML engineer.
- The two roles share evaluation, fine-tuning and serving cost, and they mean different things by each. Interview for the specific work, not the shared vocabulary.
- Add the second role when a signal says so: an inference bill that justifies your own model, a quality plateau that retrieval can't fix, or a product built on your own data.
Why this question matters
You're about to write a job description or approve a requisition, and the title you choose decides who applies. Write "ML engineer" for work that's really prompt design, retrieval and evaluation, and you'll attract people who'd rather be training models, interview them on the wrong things, and lose them in a year. Write "AI engineer" for a forecasting or ranking problem on your own data, and you'll hire someone excellent at the wrong half of the stack.
The market doesn't help. The data from Pave shows most people doing AI/ML work still carry an ML title, while the fastest-growing role is the other one: LinkedIn's 2026 Jobs on the Rise list put AI engineer first. This guide compares those two roles. If you also need to place a data engineer in the picture, the full three-way comparison is data engineer vs. ML engineer vs. AI engineer, and it's the better starting point for that question.
The frame: Where the work touches the model
Every AI system has a model in the middle, data going in to produce it, and a product system around it. The two roles split on which of those they own.
An AI engineer works around a model that already exists, hosted or open-weight. The output is a feature or system that behaves correctly for users: retrieval that finds the right documents, an agent that completes a task, an evaluation suite that says whether last week's change made things better, and a cost and latency profile the business can live with. The skill set is software engineering with model behavior as the domain.
An ML engineer works on the model and the machinery that produces and serves it: data preparation for training, training and fine-tuning jobs, offline evaluation against held-out data, deployment, and monitoring for drift. The skill set is software engineering with statistics, optimization and distributed compute as the domain.
Swyx's 2023 essay on the rise of the AI engineer drew the same line: AI engineers productize foundation models and APIs, while ML engineers and researchers train and fine-tune models at scale. It quotes Andrej Karpathy's expectation that there would be far more AI engineers than ML engineers, and that one can do well in the role without ever training a model. Pave's definitions agree: AI engineers typically build complete AI systems, and ML engineers develop systems that learn from and improve with data.
What does each role build?
An AI engineer's work product is a system you can demo and measure. An ML engineer's work product is a model, a pipeline or a platform that other people's systems depend on. The table lays out the two side by side.
AI engineer | ML engineer | |
|---|---|---|
Primary output | A working AI feature or system in production | A trained, evaluated and deployed model, or the pipeline and platform behind one |
Daily work | Prompt and retrieval design, tool and agent orchestration, evaluation suites, guardrails, latency and cost tuning | Training and fine-tuning runs, feature and data preparation, offline evaluation, deployment, drift monitoring |
Typical tools | Hosted model APIs, open-weight models, vector stores, orchestration frameworks, evaluation harnesses, tracing | PyTorch, TensorFlow or JAX, experiment tracking, feature stores, distributed training, model serving stacks |
How they test quality | Task-level evals on real user inputs, regression sets, human review of samples | Held-out metrics, ablations, offline-to-online comparison, calibration |
The failure they prevent | A feature that works in the demo and fails on real inputs, or costs too much per call | A model that overfits, drifts or can't be reproduced |
Typical background | Strong product or backend engineer who learned model behavior | Software engineer or data scientist with deep statistics and ML systems experience |
Title they often get | AI engineer, applied AI engineer, LLM engineer | ML engineer, applied ML engineer, MLOps or ML platform engineer |
For a fuller role definition and the three AI-engineer subtypes, see what is an AI engineer.
Where do the two roles overlap?
They overlap in four places, and in each one the same word means different work.
Evaluation. An AI engineer builds evals that judge model output on the task your users care about, often with a model or a human as the grader. An ML engineer measures a model against labeled held-out data. A candidate fluent in one can still be a beginner at the other, which is why "has experience with evals" tells you almost nothing.
Fine-tuning. An AI engineer may fine-tune through a managed service or an adapter method to fix a format or domain vocabulary. An ML engineer owns the training infrastructure, the data pipeline behind it and the decision about when fine-tuning is worth it at all.
Serving and cost. An AI engineer reduces cost with caching, routing, smaller models for easy requests and prompt changes. An ML engineer reduces it by changing the model: distillation, quantization, custom serving. The first is a week of work, the second a project.
Retrieval and embeddings. An AI engineer chooses and uses embedding models inside a retrieval system. An ML engineer is the one who trains a better embedding model when the off-the-shelf ones plateau.
At a small company one person often covers both sides of each of these. At a larger one the roles split along the lines above.
AI engineer vs ML engineer: Which one do you hire first?
Hire for the work in front of you. If the work is building on a model someone else trains, hire an AI engineer. If the work is making a model of your own, hire an ML engineer. The table maps common situations to the role.
Your situation | Hire first | Why |
|---|---|---|
A product feature on a hosted model: search, summarization, assistant, document extraction | AI engineer | The model exists; the work is retrieval, prompts, evals and reliability |
An agent that takes actions in your systems | AI engineer | Tool design, error handling and state management are product engineering |
A prediction problem on your own data: churn, demand, fraud, ranking | ML engineer | No off-the-shelf model knows your data; someone has to train one |
An inference bill large enough to justify a smaller custom model | AI engineer to prove the demand, ML engineer to cut the cost | Building the feature first shows which calls are worth optimizing |
Fine-tuning at scale or running your own GPU serving | ML engineer | Training and serving infrastructure is their core work |
No clear answer yet | AI engineer | Cheaper to learn what works on hosted models than to build model infrastructure for a guess |
The last row is where most product companies land in 2026. Building on a hosted model is the faster path to learning whether the feature earns its place, and an AI engineer is the person who can run that experiment well.
When do you add the second role?
Add an ML engineer when one of four signals shows up in the first AI engineer's work. The inference bill or latency has grown to the point where a smaller custom model would pay for itself. Prompting and retrieval have plateaued on quality, you have labeled data, and fine-tuning would move the metric. Your product is built on a model trained on your own data. Or a regulator or customer needs a model you can inspect and reproduce.
Add an AI engineer to an ML team when the models work and nobody has built the product system around them: the interface, the evaluation on real user inputs, the guardrails, the monitoring. A strong model behind a weak system often loses to a weaker model behind a good one.
Can one person do both?
Sometimes, and the title is usually "applied ML engineer" or "AI/ML engineer". It's realistic when the team is small, the models are hosted, and fine-tuning is occasional. It stops being realistic when you need distributed training or a serving platform and a production-grade product system at once. The guide on how to hire an AI engineer covers what to test when the role has to stretch.
How do you interview each role?
Test the work, with a sample that looks like what the person will do in their first month.
For an AI engineer, give them a small retrieval or agent task and a handful of messy real inputs, then ask how they'd know it works. The answer you want describes a regression set, a way to grade outputs and a view on what a call costs. Ask what they do when the model gets it wrong in production. People who've shipped answer with specifics: fallbacks, thresholds, human review paths.
For an ML engineer, give them a dataset with a known problem, such as leakage, class imbalance or a train-test mismatch, and ask them to find it and fix it. Ask how they'd detect drift six months after deployment and what they'd do when it appears. People who've run models in production answer with monitoring and retraining triggers, and people who've only trained them tend to answer with algorithms.
In both cases, ask the candidate to describe a time the system failed in production and what they changed afterward. The quality of that story separates the two kinds of experience better than any list of tools.
What does each role cost?
Pay depends on seniority, company, location and engagement model more than on the title. For AI engineer compensation by level and engagement model, see the AI engineer rates guide.
Pave's December 2024 analysis put generalist and applied AI/ML engineers at a premium of roughly 10 to 20% over the core software engineering family, so expect either role to price above a general software engineer.
What goes wrong when you hire the wrong one?
An ML engineer hired for product work tends to over-build. Asked to add retrieval to a support tool, an ML engineer may reach for a custom embedding model and a training pipeline before anyone has seen whether off-the-shelf retrieval was good enough. The system gets sophisticated slowly, and the work you needed waits.
An AI engineer hired for model work tends to hit a wall in the middle of the project. They can wire a fine-tuning service together but can't diagnose why a trained model underperforms, and the project stalls at the point where the training data needs real statistical judgment.
The third mistake is the job description that asks for both. A posting that wants distributed training, a production retrieval system and a customer-facing interface describes two senior roles and a product engineer. Each hire is possible. The combined candidate is rare and priced accordingly, which usually means the hiring manager hasn't yet decided which problem comes first.
What to do next
Write down the first piece of work in one sentence and ask where it touches the model. If the sentence mentions a model you'd call through an API, you're hiring an AI engineer. If it mentions training, fitting or deploying a model on your own data, you're hiring an ML engineer. If you can't write the sentence, start with the AI engineer and let the first three months tell you whether a second role is coming.
If the work is on the AI engineer side, see senior AI engineers from a network of 11,000+ vetted builders with under 2% acceptance, and get a matched shortlist within 72 hours of the scoping call. For the case that includes data engineering, the three-way comparison guide is the place to start.
People also ask
An AI engineer builds products and systems around models that already exist, covering prompts, retrieval, agents, evaluation, latency and cost. An ML engineer works on the model itself and the machinery around it: data preparation for training, training and fine-tuning, deployment and drift monitoring. The practical test is whether the work calls a model through an API or trains and serves one on your own data.
Compensation varies by seniority, company, and location more than by role title. The AI engineer rates guide breaks out AI engineer compensation by level. Title pools differ in company mix and seniority, so published title medians are a poor basis for comparing the two roles.
They need working literacy: how embeddings behave, what evaluation metrics measure, why models overfit, and what fine-tuning changes. They don't need training-infrastructure depth. Swyx's 2023 essay quotes Andrej Karpathy saying one can succeed as an AI engineer without ever training a model. The depth that matters for the role is software engineering plus rigorous evaluation of model output.
Usually an AI engineer. Building on a hosted model shows whether the feature earns its place before you spend on model infrastructure. Hire an ML engineer first if your product is a prediction on your own data, such as demand, fraud or ranking, where no off-the-shelf model applies. Add the second role when inference cost, a quality plateau or proprietary data calls for it.
Often, with a ramp. Many ML engineers have the software depth and pick up retrieval and agent patterns quickly, but they tend to over-build and underinvest in task-level evaluation and product constraints. The reverse is harder: a product-focused AI engineer rarely has the statistical training for model diagnosis. Interview for the specific work and judge the candidate on that sample.
Data engineer vs. ML engineer vs. AI engineer
Where data engineering ends and ML engineering begins, with AI engineers next to both. What each role builds, where they overlap, and which one you're missing.
How to hire an AI engineer
A practical guide to scoping, evaluating, and onboarding a senior AI engineer. What the role actually involves in 2026, how to evaluate for production judgment, and what the first 30 days should look like.
What is an AI engineer
A clear definition of what an AI engineer does, how the role differs from data scientists and ML engineers, the three main subtypes in 2026, and what to evaluate when hiring one.
Hire expert talent through A.Team
A.Team's network of 11,000+ vetted senior builders, with under 2% of applicants accepted. Engagements are time-and-materials with transparent per-builder pricing; your team manages day-to-day, and a dedicated Team Success contact runs the kickoff and stays close throughout. Describe the work and get a matched shortlist within 72 hours of the scoping call.
Talk to A.Team