Key takeaways
- Nearly all hiring volume is in engineering roles that use models, not research roles that invent them.
- Pay tracks proximity to revenue, skill scarcity and outcome ownership — not the job title.
- Most people who get these jobs moved sideways from backend, data or platform engineering.
- One shipped project with an honest evaluation outperforms any stack of certificates.
- These postings are keyword-searched heavily, so vocabulary gaps keep capable people from being read at all.
The roles, and what they actually involve
The label "AI job" covers work that differs enormously day to day. Getting the distinction right matters, because preparing for the wrong one wastes months.
Research roles are the smallest category and the hardest to enter, typically requiring publications and often a doctorate. Nearly all the hiring volume is in engineering roles: building systems that use models rather than inventing them.
That imbalance is the single most useful thing to know before planning a route in. People frequently prepare for the research track — reading papers, reimplementing architectures — while applying to engineering jobs that test none of it, and conclude the market is impossible when they were simply preparing for a different one.
- Machine Learning Engineer — training, evaluating and deploying models in production; heavy software engineering
- AI / LLM Application Engineer — building products on top of existing models: retrieval, tool use, evaluation, guardrails
- Data Engineer — the pipelines everything else depends on; consistently in demand and often undervalued
- MLOps / Platform Engineer — serving, monitoring, cost and reliability of models in production
- Research Scientist — novel methods; smallest headcount, highest credential bar
- Applied Data Scientist — measurement, experimentation and decisions rather than model building
What pay actually tracks
Published salary ranges vary so much by country, company stage and seniority that quoting a single figure would mislead you. The more useful observation is what compensation correlates with, because that is what you can influence.
Three things move it consistently: proximity to revenue, scarcity of the specific skill, and whether you own an outcome or execute a task. An engineer who owns a system that demonstrably makes or saves money is paid differently from one who implements tickets, regardless of job title.
The practical implication is to look past the title in a posting and read the responsibilities. Two roles called "AI Engineer" at different companies can be entirely different jobs at entirely different pay levels.
| Factor | Why it pays | How to move it |
|---|---|---|
| Proximity to revenue | The impact is measurable | Choose teams that own a number |
| Skill scarcity | Few people can do it | Go deep rather than broad |
| Outcome ownership | You carry the consequence | Ask for the outcome, not the ticket |
| Company stage and funding | Budget differs enormously | Compare like with like |
| Location and remote policy | Bands are set regionally | Check the band before the interview |
The skills that come up in real postings
Strip away the buzzwords and hiring for these roles is more conventional than it appears. Strong software engineering underpins nearly all of it — most production AI work is data plumbing, evaluation and reliability rather than modelling.
The interview loops reflect that. Candidates are far more likely to be asked to write clean code against a practical problem, or to design a system that stays within a latency budget, than to explain an architecture. Preparing as though it were a machine learning exam is a common and costly misread.
- Python, and genuinely solid general programming ability
- SQL and data modelling — underrated and asked about constantly
- Working knowledge of a modelling framework; depth matters less than being able to ship
- Retrieval, embeddings and evaluation for LLM application roles
- Cloud fundamentals, containers and CI/CD
- The ability to define what "working correctly" means and measure it
Getting in from an adjacent role
Most people entering these jobs are not new graduates — they are backend engineers, data analysts or platform engineers who moved sideways. That is by far the most reliable route, because the surrounding engineering skill transfers directly.
The thing that converts an application is a shipped, working project you can talk about in depth. Not a tutorial reproduction; something with real data, a real evaluation of whether it works, and an honest account of what broke. One such project outperforms a stack of certificates.
The strongest version of this move happens inside your current job. Volunteering for the AI-adjacent work on your own team gives you production experience with real constraints, a reference who can vouch for it, and a story that does not depend on a side project competing with your day job for attention.
- Build something end to end, deployed, that someone other than you has used
- Be able to explain your evaluation method and its weaknesses
- Move towards AI-adjacent work inside your current job first
- Mirror the posting’s vocabulary on your resume — these roles are keyword-searched heavily
A realistic note on competition
These postings attract very large applicant volumes, which means the automated screen matters more here than in most fields. A resume that does not name the specific tools in the posting frequently never surfaces in the recruiter’s search, however capable the candidate.
Before applying, check your resume against the actual job description and close the vocabulary gaps. It is the cheapest possible improvement to your odds, and for high-volume roles it is often the difference between being read and not being read.
Volume also means referrals matter more here than in most hiring. A referred application is read by a person; a cold one has to survive a search first. That is worth remembering before spending another evening on the fortieth cold application.
Frequently asked questions
Do I need a PhD to work in AI?
For research scientist roles at labs, usually yes. For the large majority of AI engineering jobs, no — strong software engineering plus demonstrable shipped work matters far more than the degree.
What is the difference between an ML engineer and a data scientist?
ML engineers build and run models in production and are closer to software engineering. Data scientists focus on analysis, experimentation and informing decisions. The titles are used inconsistently, so read the responsibilities rather than the label.
Can freshers get AI jobs?
It is harder than for mid-level candidates because entry-level tasks are the ones most affected by automation, but it happens regularly. The candidates who succeed almost always have a real deployed project with an honest evaluation, rather than only coursework.
Which AI role has the most openings?
Engineering roles that build on existing models, by a wide margin. Research is the smallest category and the hardest to enter, yet it absorbs a disproportionate amount of candidate preparation time.
Why do published AI salary figures vary so much?
Because the titles cover genuinely different jobs across countries and company stages. An average across them describes no real role — look at what moves pay instead: revenue proximity, scarcity and outcome ownership.
What is the fastest route in from a backend role?
Take the AI-adjacent work on your current team. It gives you production constraints, a reference who can vouch for the work, and a story that does not compete with your day job for evenings.
Check this against your own resume
Scan your CV against a real job description, or build a parse-safe one from scratch. Your first scan costs nothing.