Research scientist
The smallest category and the most demanding. Research scientists develop new methods and architectures, typically at large laboratories or well-funded startups.
A doctorate in machine learning, mathematics, physics or a closely related field is close to standard, along with a publication record at recognised venues. Compensation is among the highest in the industry.
Realistically, this is not a route you enter laterally. It requires a research training pathway and several years of specialised study.
Machine learning engineer
The largest well-paid technical category, and the one most software engineers can realistically move into.
The work is engineering: building training pipelines, deploying models to production, managing inference at scale, monitoring drift and performance. Strong software engineering is the foundation, with solid — not necessarily research-level — mathematics on top.
For an experienced backend or infrastructure engineer, this is an achievable transition over a year or so of focused work, and it is usually easier to make internally than by changing company simultaneously.
Data engineer
The least glamorous role on this list and one of the most consistently in demand, because every AI initiative depends on data infrastructure that mostly does not exist yet.
Data engineers build the pipelines, storage and transformation systems that everything else relies on. The skills are concrete and learnable: SQL to a high standard, a programming language, distributed processing frameworks and cloud data platforms.
Because it is less fashionable than model work, competition is lower relative to demand, which makes it one of the better risk-adjusted entry points into the field.
Applied AI and integration roles
The fastest-growing category by headcount: people who take existing models and build useful products with them.
The work involves prompt and context design, retrieval systems, evaluation frameworks, and the substantial engineering required to make probabilistic systems behave acceptably in production. Evaluation is the underrated skill — knowing whether a system is actually working is harder than getting it to produce output.
This is the most accessible entry point for existing software engineers, because it leans more on engineering judgement than on mathematics.
AI product management
Product managers for AI systems need an unusual combination: enough technical understanding to know what is feasible, plus the product judgement to know what is worth building.
The distinctive challenge is that these systems are probabilistic. Defining acceptable failure rates, designing interfaces that communicate uncertainty honestly, and setting evaluation criteria are all genuinely different from conventional product work.
Experienced product managers can move into this by developing technical depth, and demand currently exceeds the supply of people who have both.
Governance, safety and compliance
A growing category driven by regulation, most notably the EU AI Act and comparable frameworks emerging elsewhere.
Roles cover risk assessment, bias and fairness auditing, documentation and conformity assessment, and policy development. They draw on backgrounds in law, policy, audit and social science as much as engineering.
For people with regulatory or compliance experience in another domain, this is one of the more accessible routes into AI work, and demand is likely to grow as enforcement begins in earnest.
Applying AI inside your existing field
The largest opportunity for most people is not becoming an AI specialist at all.
A clinician who understands where diagnostic tools fail, a lawyer who can evaluate document review systems, a financial analyst who can build and validate models, a marketer who can measure whether generated content actually performs — each combines domain knowledge with tool fluency, and that pairing is scarcer than either skill alone.
It is also a far shorter path. Months of deliberate learning applied to expertise you already have, rather than years retraining into a new profession.
How to get there
A few things consistently distinguish people who make the transition.
Build something real and make it visible. Employers can verify a working project; they cannot verify a certificate. A deployed system solving an actual problem, with honest evaluation of where it falls short, is worth more than a stack of course completions.
Learn evaluation seriously. Most practitioners can get a model producing output; far fewer can tell you rigorously whether it is good enough, which is what production work actually requires.
And move internally where you can. Taking on AI work inside your current role is usually easier than changing company and discipline at once.