Definition
AI engineering roles in 2026 split into seven distinct jobs that share a vocabulary: AI engineer, agentic AI engineer, ML engineer, LLMOps engineer, applied scientist, data scientist and forward deployed engineer. They differ by what the person does with a model, not by how much AI is in the title.
The vocabulary of AI hiring has turned over twice in three years, and the job adverts have not caught up. A team writes "AI engineer", interviews for an applied scientist, and hires someone who spends their first month building retrieval pipelines. Everybody involved was honest and the search still failed.
This is the map as the roles are understood in 2026. Seven titles, what each one is for, where the lines genuinely blur, and what we can and cannot tell you about what each pays.
The seven roles, in one table
Hold them apart by the question each one exists to answer:
| Role | The question it answers | Core skill |
|---|---|---|
| AI engineer | How do we put a model inside our product? | Applied product engineering |
| Agentic AI engineer | How do we make it act, not just answer? | Distributed systems thinking |
| ML engineer | How do we train and serve our own model? | Modelling and MLOps |
| LLMOps engineer | How do we run this reliably and affordably? | Platform and observability |
| Applied scientist | Can the model itself be made better? | Research and experimentation |
| Data scientist | What is the data telling us? | Statistics and inference |
| Forward deployed engineer | Will it work in the customer's world? | Engineering plus customer judgement |
The table is the short answer. The rest of this page is what sits underneath each row, because the distinctions that matter at hiring time are not the ones in the title.
The four that get confused most
AI engineer
Builds products on top of models somebody else trained. Retrieval, prompting, evaluation harnesses, serving, and the integration work that makes a model feature survive contact with real users. It is a software engineering role first, which is why the strongest candidates often have no machine learning background at all. We have a fuller definition of the AI engineer, including how it sits against the data scientist.
Agentic AI engineer
Also written AI agent engineer or agent engineer. Builds systems where the model plans, calls tools, holds state across several steps, and keeps going until a goal is met or it gives up. The difference from AI engineering is not the model, it is the control flow: an AI feature answers once and you can see whether the answer was good, while an agent takes a sequence of actions against real systems and can be wrong in the middle of one.
That changes what you are hiring for. The hard problems are partial failure, retries that are not idempotent, cost that scales with how confused the model gets, and knowing what the system is permitted to do without a human. Those are distributed systems problems wearing a new hat. The agentic AI engineer has its own page, because it is the newest of the seven and the least well defined.
ML engineer
Trains, tunes and productionises models, and owns the pipelines around them. The title has been squeezed from both sides since foundation models arrived: a great deal of work that used to need a trained model now needs an API call and good evaluation. ML engineering has not gone away, it has concentrated, into teams with proprietary data worth training on and problems a general model handles badly.
LLMOps engineer
The platform underneath everyone else. Model serving, routing between providers, caching, token cost, latency budgets, eval infrastructure, and the observability that tells you an agent started failing three days ago. Roughly half of senior AI work is platform-side rather than modelling, which is why this separated out as a role in its own right rather than staying a hat the ML engineer wears.
An advert that lists model training, agent orchestration, eval infrastructure and stakeholder analysis is not describing a senior hire. It is describing four people, and the strongest candidate for any one of them will read it and move on.
The three that sit at the edges
Applied scientist and research engineer
Improves the model rather than the product around it. Novel architectures, fine-tuning strategy, evaluation methodology at a level that produces new knowledge rather than a dashboard. Almost always needs a research background and almost never needs to exist in a company that is not training its own models. If you are reaching for this title and you use foundation models through an API, you probably want an AI engineer with strong evaluation instincts.
Data scientist
Analysis, experimentation and inference. The oldest title of the seven and the one whose meaning varies most between companies: in some it is a statistician, in others it is an ML engineer with a different reporting line. Scope the work, because the word will not tell you.
Forward deployed engineer
Builds with the customer, in the customer's environment, against the customer's constraints. It is not an AI role by definition, but it has become one in practice, because most companies deploying AI into an enterprise discover the model was the easy part and the data, the permissions and the workflow were not. We have the full explainer and a comparison against solutions and sales engineering.
Three questions that resolve almost every case
When a brief could be two of these roles, these settle it faster than any list of skills:
- Who trained the model? If the answer is a provider, you are hiring an AI engineer, an agentic AI engineer or an LLMOps engineer. If the answer is you, an ML engineer or an applied scientist is in scope.
- Does the system take actions, or produce answers? Actions against real systems means agentic. Answers a person reads and acts on means applied AI engineering. The word "agent" in the brief is not evidence either way; half the adverts using it describe a chatbot.
- Who is the user sitting with? If the engineer will be in a customer's environment, in their meetings, against their data, that is forward deployed work whatever the discipline underneath it.
What they pay, and where we stop
We publish bands for the roles we place enough of to have real numbers for. Sydney, base only, 25th to 75th percentile of accepted offers:
| Role | Senior (Sydney) | Principal |
|---|---|---|
| AI / ML engineer | AUD 180-220k | AUD 220-260k |
| Forward deployed engineer | AUD 180-240k | Tech lead AUD 230-270k |
| Software engineer | AUD 160-190k | AUD 190-225k |
Three of the seven roles have a published band here, and four do not. Agentic AI engineer, LLMOps engineer, applied scientist and data scientist are priced against the AI and ML engineer band when we run those searches, because we have not placed enough of any of them to publish a separate number with a straight face. Anyone showing you a precise agentic AI engineer band for Australia in 2026 has built it from job advert text rather than accepted offers.
What the AI premium looks like is clearer. Senior AI and ML engineers sit 12 to 18 per cent above senior software engineers in like-for-like roles. On the midpoints above, AUD 200k against AUD 175k, that gap costs about AUD 30,000 a year once superannuation, payroll tax and other on-costs are added. The cost to hire calculator does that arithmetic by state, and the salary checker has every band in the matrix.
Writing the advert so the right person applies
The strongest candidates in all seven categories self-select hard, and they do it from the responsibilities rather than the title. Three things decide it:
- Name the system, not the stack. "You will own the retrieval and eval layer behind our support product" tells an AI engineer more than a list of six frameworks.
- Say what it does when it fails. Whether a human approves actions, whether it retries, whether it can spend money. Agentic candidates read this first and nobody else cares.
- Pick one role. If the honest answer is that you need two, say so and hire two, in order. The composite advert is the single most reliable way to lose all four candidate pools at once.
If you want the advert checked before it goes out, the JD grader scores it against the things that cost you applications and names the phrase responsible for each point lost. It is rule-based, so the same advert always scores the same, and it runs entirely in your browser.
FAQ
What is the difference between an AI engineer and an agentic AI engineer?
An AI engineer builds product features on top of models: retrieval, prompting, evaluation and serving, where the model produces an answer a person then acts on. An agentic AI engineer builds systems where the model plans, calls tools and takes multi-step actions against real systems until a goal is met. The distinction is control flow rather than the model itself, and it changes what you screen for: agentic work is mostly distributed systems judgement, covering partial failure, retries, cost control and what the system is allowed to do without a human.
How many distinct AI engineering roles are there?
Seven are in common use in 2026: AI engineer, agentic AI engineer, ML engineer, LLMOps engineer, applied scientist, data scientist and forward deployed engineer. They are distinguished by what the person does with a model rather than by seniority or stack. Titles are used loosely across companies, so the reliable approach is to scope the actual work before choosing the title.
Is an AI agent engineer the same as an agentic AI engineer?
Yes. Agentic AI engineer, AI agent engineer and agent engineer describe the same job. The variation is regional and company-specific rather than meaningful. Use whichever form your candidates are likely to search for, and mention the alternatives in the advert body so the role is findable either way.
Do I need an ML engineer if I use foundation models?
Usually not. If you access models through a provider API, the work is applied AI engineering, agentic engineering and platform reliability rather than training. ML engineers are needed where there is proprietary data worth training on, or a problem that general models handle badly. Hiring an ML engineer for an API-based product is a common and expensive mismatch, because the work will not use what they are best at and they will leave.
What do AI engineering roles pay in Australia in 2026?
Senior AI and ML engineers in Sydney sit at AUD 180 to 220k base, with principals at AUD 220 to 260k, from accepted offers at the 25th to 75th percentile. That is 12 to 18 per cent above senior software engineers in like-for-like roles. Agentic AI engineer, LLMOps engineer and applied scientist have no separately published band, because there are not yet enough accepted offers in Australia to build one honestly; those searches are priced against the AI and ML engineer band.