AI Professionals Career Guide
Explore AI jobs, salaries, and career paths across research, engineering, product, and governance. Find the specialism that fits you, and stay current on models, agents, evaluation, and AI regulation.
How we research our career guides
The Top AI News, in 5 Minutes.
We handpick the best AI newsletters and distill key news and insights into a 5-minute audio briefing, texted to you every morning.
Try the AI Briefing, FreeFreeSee the 10 newsletters we handpicked

At a glance
AI Careers at a Glance
There is no federal occupation called AI professional. The Bureau of Labor Statistics counts the field in pieces: 275,600 data scientists at a $120,230 median in May 2025, growing 35 percent to 2035; 1,717,800 software developers at $135,980; 38,600 computer and information research scientists at $140,300; 192,900 information security analysts at $129,180; and 685,800 computer and information systems managers at $175,140. Across computer and information technology occupations the projection is about 280,000 openings a year. The common picture of the job, somebody training a model, covers one slice of it. The field extends well beyond training: research scientists who design and probe methods, machine learning engineers who make training and serving reliable, AI engineers who build retrieval and agents on models somebody else trained, product managers who decide what is worth building and what counts as good enough, consultants who judge whether an organisation can carry the promise, and governance and security specialists who answer for what the system does. Demand is real and narrower than the discourse: Lightcast's work for the 2026 Stanford AI Index put AI skills in 2.5 percent of United States postings, up 55 percent in a year. Entry is by evidence in most of the field, by graduate degree in research, and the regulatory layer is becoming part of everyone's job.
Key Industry Statistics
2.5%
US job postings naming AI skills (2025)Lightcast, for the 2026 Stanford AI Index
$140,300
Median pay, computer and information research scientistsUS Bureau of Labor Statistics
+35%
Projected growth for data scientists, 2025 to 2035US Bureau of Labor Statistics
280,000
Annual openings across computer and IT occupationsUS Bureau of Labor Statistics
Common Roles
- AI Engineer
- Machine Learning Engineer
- AI Research Scientist
- AI Product Manager
- AI Consultant
- AI Governance Lead
Where AI Professionals Work
- Frontier model labs
- Software and platform companies
- Banks and insurers
- Healthcare and life sciences
- Government and defence
- Consultancies and agencies
Credentials That Matter
- Bachelor's in a computing field
- Master's or doctorate for research
- A portfolio with measured results
- Cloud and ML platform certifications
- NIST AI RMF fluency
Key Areas in AI Careers
Research & Methods
Designing, training and probing architectures, objectives and the learning methods behind them.
Applied AI Engineering
Retrieval, agents and evaluation built on top of models somebody else trained.
ML Platform & Serving
Pipelines, training infrastructure and inference held inside a latency and cost budget.
Product & Adoption
Deciding what is worth building, and what counts as good enough to ship.
Governance & Policy
Risk frameworks, transparency duties and the audit trail a regulator will ask for.
Security & Red Teaming
Prompt injection, excessive agency and the blast radius of a tool call.
AI Jobs and Salaries
There is no AI occupation in the federal statistics, so this table uses the five adjacent occupations the work is actually counted under. Every figure is the May 2025 median from the Occupational Outlook Handbook, and no blended AI salary is invented here.
Data scientists
$120,230
Information security analysts
$129,180
Software developers
$135,980
Computer and information research scientists
$140,300
Computer and information systems managers
$175,140
Source: US Bureau of Labor Statistics, Occupational Outlook Handbook, May 2025 wage data, accessed October 2026.
AI Career Paths
Working in AI is not one ladder. It is four tracks that share a vocabulary and then diverge sharply in what they reward, who they report to and how much of the day is spent near a model.
- Junior AI Engineer0 to 2 yearsBuild features on an existing stack. Measured on reliability and on whether your own evaluation catches your bugs.
- AI Engineer2 to 5 yearsOwn a surface end to end, including its evaluation set, its cost per resolved task and its failure taxonomy.
- Senior AI Engineer5 to 9 yearsOwn the seams: retrieval quality, agent permissions, routing policy and the budget that constrains all three.
- Staff or Principal Engineer9 or more yearsSet technical direction across teams and decide which model bets the organisation stops funding.
Credential that fits A portfolio of shipped systems with measured failure rates
There is no federal occupation code for an AI engineer, a machine learning engineer or a head of AI. The figures on this page therefore come from the five adjacent occupations the Bureau of Labor Statistics does publish, which is also how an employer's compensation system is likely to file the role.
A Typical Day in AI Jobs
Five roles that between them cover most of the field's headcount, ordered from the applied end toward research and leadership. The schedules are representative rather than universal: company size, sector and whether anything is on fire move all of it.
AI EngineerEntry to Mid · Applied ProductBuilding a user-facing feature on a model somebody else trained, where most of the difficulty is retrieval, evaluation and cost.
- 09:00Check overnight evaluation runs and the cost dashboard before opening anything else. Regressions hide in both.
- 09:30Read ten real traces from yesterday's failures. This is the habit that generates everything else on the list.
- 11:00Change one thing: a chunking rule, a reranker, a tool schema. Re-run the evaluation set before and after.
- 14:00Review a colleague's prompt change the way you would review code, including what it costs per call.
- 16:00Write up what moved and what did not, so the next person does not repeat the experiment.
Machine Learning EngineerMid-Level · Platform and ServingKeeping training and inference reproducible, fast and affordable for everyone else, usually with a pager attached.
- 08:30Check pipeline health and drift monitors. A silent data change upstream is the most common cause of a bad week.
- 10:00Work a latency budget: profile where the time goes, then argue about which part of it is negotiable.
- 12:00Move an offline workload to batch processing, which halves its cost, and verify nothing depended on it being live.
- 14:30Pair with a researcher on getting an experiment reproducible enough to hand over.
- 16:30Capacity planning, because the next model version will not fit the current budget without a decision.
AI Research ScientistSenior-Level · Methods and EvaluationOwning a line of questions rather than a product surface, with a master's degree or doctorate as the usual entry.
- 09:00Read overnight results and kill the variants that are clearly not working, before sunk cost sets in.
- 10:30Design the next ablation so it can only have one explanation. Most of the skill is here.
- 13:00Reading group or paper discussion, which is where the field's actual peer review happens.
- 15:00Write. A result nobody can reproduce from your description did not happen.
- 17:00Negotiate compute with the platform team, which is the real constraint on an agenda.
AI Product ManagerMid to Senior · Product and AdoptionDeciding what is worth building and what counts as good enough to ship, with the evaluation bar written before the work starts.
- 09:00Look at yesterday's user-reported failures, not the dashboard. The dashboard shows what you already thought to measure.
- 10:30Agree the quality bar for a new surface with engineering, in numbers, before anybody builds it.
- 13:00Sit with a customer and watch them use the thing. Half of the roadmap comes from this hour.
- 15:00Work the design question of what the product does when the model is wrong, which is most of the product.
- 16:30Write the decision down with the evidence attached, so the next reversal is informed rather than political.
AI Governance LeadSenior-Level · Risk and ComplianceOwning the approval route, the incident process and the answer a buyer or a regulator receives about a live system.
- 09:00Review deployment requests against the Govern and Map functions: named owner, stated context, affected users.
- 11:00Work a security question with engineering, usually about tool permissions and what untrusted content can reach.
- 13:00Answer a customer's due-diligence questionnaire, which is where governance work meets the sales cycle.
- 15:00Track the regulatory calendar and turn one date into an engineering ticket with an owner.
- 16:30Run or review an incident post-mortem, and change the standing process rather than the individual.
JOB SEARCH
How to Get an AI Job in 2026
Working in AI is not one job. It is six, with different entry evidence, different pay bands and different employers, and the first useful decision is which of them you are applying for. What follows is grounded in federal occupational data, the measurement work the field runs on, and the regulation that is now part of shipping.
Entry-Level
Break Into the AI Field
What the field is actually made of, which of its six jobs you are applying for, what each one expects on paper, and what a hiring manager reads instead of the paper.
What you should know about the field
How the work is counted, what the routes into it are, and what the demand data really supports.
There Is No AI Occupation in the Federal Data
The Bureau of Labor Statistics does not publish an occupation called AI professional, AI engineer or prompt engineer. It counts the field in pieces: 275,600 data scientists at a $120,230 median in May 2025, 1,717,800 software developers at $135,980, 38,600 computer and information research scientists at $140,300, 192,900 information security analysts at $129,180, and 685,800 computer and information systems managers at $175,140.
That matters for a job search. The title on the posting is set by the hiring team, while the pay band is usually set by whichever of those occupations the finance system files you under. Ask which internal job family a role sits in, because it predicts the ceiling better than the title does.
The Six Jobs People Mean by Working in AI
Research scientists design and probe methods, and the federal proxy for them expects a master's degree. Machine learning engineers make training and serving reliable and affordable. AI engineers build retrieval, agents and evaluation on top of models somebody else trained. Product managers decide what is worth building and what counts as good enough. Consultants judge whether an organisation can carry the promise. Governance and security specialists answer for what the system does once it is live.
They share vocabulary and almost nothing else day to day. Choosing between them early is worth more than any certificate, because the evidence each one asks for takes a year or two to build and does not transfer cleanly.
What the Demand Data Actually Says
Lightcast's analysis for the 2026 Stanford AI Index found AI skills named in about 2.5 percent of United States job postings, up 55 percent on the previous year, 72 percent on 2022 and close to 300 percent over the decade. Python remained the most requested specialised skill, appearing in 258,674 postings, roughly 30 percent above 2024.
One posting in forty is real growth and a narrow market at the same time. Read it as a reason to be specific rather than a reason to be confident: the roles exist, but a generic AI enthusiast competes against everyone, while a candidate with a named specialism competes against a much smaller field.
Which Credential Each Route Expects
The federal entry requirement differs by route. Software developers, data scientists and information security analysts are all recorded with a bachelor's degree as typical entry-level education. Computer and information research scientists, the nearest proxy for frontier research work, are recorded with a master's degree, and doctorates are common in practice.
So there is no single answer to whether you need a graduate degree. For applied engineering, product and consulting the answer is usually no and the portfolio decides. For research posts the degree is close to a filter, and the time it costs is the real decision rather than the tuition.
Skills employers look for
What a hiring manager checks when there is no track record to check, beyond the degree itself.
Python and SQL, Still the Baseline
Python was the most requested specialised skill in United States AI postings in 2025, named in 258,674 of them, about 30 percent above 2024. SQL sits beside it because almost every applied problem starts with getting the right rows out of somebody's warehouse. Neither is a differentiator any more, and the absence of either is still disqualifying.
Reading an Evaluation Before Believing It
The AI Index exists because headline capability claims move faster than the evidence for them. The entry-level version of this skill is modest: know what the benchmark measured, on what data, against what baseline, and whether the comparison was run at the same cost. Candidates who ask those four questions out loud read as practitioners rather than enthusiasts.
Shipping Something a Stranger Can Run
Across every route into this field the entry evidence is the same shape: a working artefact with a readme, a reproducible setup and an honest account of what it gets wrong. The federal occupations behind these jobs all record a bachelor's degree as typical entry, which means the degree gets you read and the artefact gets you hired.
Cost Literacy From Your First Project
Model spend is an engineering variable, not a finance one. Prompt caching bills cache writes at 1.25 times base input tokens and cache reads at a fraction of it, and the Message Batches API halves the cost of work that does not need an immediate answer, with most batches finishing inside an hour. Knowing those two levers early separates a prototype from something anyone will run.
Interview questions and answers
Questions entry candidates actually get, and what the interviewer is listening for underneath them.
Which part of AI work do you want to do, and why that one?
What they are looking for: Whether you have chosen a lane. Generalist enthusiasm is the most common reason an otherwise strong entry candidate is passed over.
I want applied engineering rather than research. I like problems where the model is fixed and the difficulty is retrieval, evaluation and cost, and I proved that to myself by rebuilding a search feature three times before the answers were actually right. Research appeals to me intellectually, but I would rather ship something people use every week.
Make it yours: Name the route and the thing you did that tested the preference. A choice with no evidence behind it reads as a guess.
Walk me through a project you built, including what it gets wrong.
What they are looking for: Honesty and measurement. Interviewers assume every entry project is fragile; they are testing whether you know where.
It answers questions over about four thousand internal documents. It is reliable on direct factual lookups and poor on anything requiring two hops, because my chunking splits tables from their captions. I know that because I labelled sixty real questions and scored them, and the two-hop ones failed roughly half the time.
Make it yours: Bring a number you measured yourself. A candidate who can quantify their own failure rate is instantly above the median.
How would you tell whether a new model is actually better for us?
What they are looking for: Whether you reach for a benchmark leaderboard or for your own task. This is the single most useful entry-level instinct.
I would not start from a public benchmark, because it is not our distribution. I would take fifty real requests we already have correct answers for, run both models at the same cost ceiling, and compare accuracy, latency and spend. If the new one wins on two of three I would shadow it in production before switching.
Make it yours: Say out loud that you would hold cost constant. Most candidates compare quality and quietly ignore the bill.
What would you do if the model is confidently wrong in production?
What they are looking for: Debugging order. They want to hear you isolate the layer before you start changing prompts.
First I check whether the right context was retrieved at all, because a retrieval miss and a generation error look identical to a user. If retrieval was fine I look at whether the model version or the prompt changed recently. Only then do I touch the prompt, and I add the case to the evaluation set before I do.
Make it yours: Mention adding the case to the evaluation set. It is the step that separates a fix from a habit.
Mid-Level
Pick a Lane and Go Deep
Where pay and growth actually diverge between the specialisms, which skill clusters are scarce right now, and how the work changes once other people depend on what you built.
What you should know at this stage
Where the money, the growth and the scarcity sit once you are past the first job.
Pay Rises With the Occupation You Are Counted In
The May 2025 medians run $120,230 for data scientists, $129,180 for information security analysts, $135,980 for software developers, $140,300 for computer and information research scientists and $175,140 for computer and information systems managers. The spread inside each is wide: the lowest tenth of data scientists earned under $67,240 and the highest tenth over $199,130.
The practical reading is that a lateral move between occupations often pays more than a promotion inside one. The gap between a senior individual contributor and a first-line manager in these figures is roughly $35,000, which is the same order as several years of in-grade increases.
The Specialisms Diverge Faster Than the Titles Do
Growth is not evenly spread. Data scientists are projected to grow 35 percent from 2025 to 2035 with about 24,800 openings a year, information security analysts 21 percent, computer and information research scientists 22 percent, and software developers 10 percent from a far larger base. Across computer and information technology occupations the projection is about 280,000 openings a year.
By mid-career the question is no longer whether to work in AI but which of those curves you want to be standing on. The ones with small bases and high growth are competitive to enter and slow to hire; the large base is easier to enter and harder to stand out in.
Agentic Work Is the Fastest Growing Skill Cluster
Lightcast's posting analysis for the 2026 AI Index put the agentic AI skill cluster at 0.06 percent of United States postings in 2024 and 0.23 percent in 2025, an increase of more than 280 percent and roughly 90,000 postings. That is still a small slice of the market, and it is the slice growing fastest.
For a mid-career practitioner the useful consequence is that experience here is scarce by definition. Anyone who has run a tool-using system in production for a year, and can describe its retry, approval and compaction policy, is arguing from evidence that most candidates do not have yet.
Security Became a Named Part of the Job
The OWASP Top 10 for Large Language Model Applications, 2025 edition, lists prompt injection first, then sensitive information disclosure, supply chain, data and model poisoning, improper output handling and excessive agency. Those are design faults rather than operational ones, which is why they land on the people building the system rather than on a separate security team.
The labour market reflects it. Information security analysts are projected to grow 21 percent to 2035 at a $129,180 median, with about 14,100 openings a year, and AI-specific security work is a recognised specialism inside that.
Skills employers look for
What separates a competent practitioner from the one a team is organised around.
Evaluation Design, Not Evaluation Running
At mid-career the question stops being whether you can run a benchmark and becomes whether you can design one for a problem nobody has benchmarked. That means sampling real failures rather than inventing cases, grading outcomes rather than paths, preferring deterministic graders, and calibrating any model judge against human labels before trusting it.
Inference Economics You Can Defend
Cost per token keeps falling while total spend keeps rising, because richer workloads consume far more tokens per task. The levers you are expected to price are caching, batching and routing: cached reads cost a fraction of fresh input, and asynchronous batch processing is half price. Quoting a cost per resolved task rather than a cost per call is the tell of someone who has run this.
Working Across the Specialisms
The failures that reach production are usually at a seam: a research result that does not survive the serving budget, a product decision made without an evaluation, a governance requirement discovered after launch. Mid-career value comes from being able to hold a useful conversation in two of the six specialisms rather than being excellent in exactly one.
Writing for People Who Will Not Read Code
Information security analysts, systems managers and the compliance function all have to act on your judgement without reading your repository. The EU AI Act's Article 4 literacy duty pushes the same way, obliging providers and deployers to raise AI literacy among the staff operating their systems. The writing is part of the engineering, not an overhead on it.
Interview questions and answers
Mid-career questions, where the subject shifts from your code to your judgement and its cost.
You own cost for an AI feature. Where do you look first?
What they are looking for: Whether cost is an engineering variable to you or a line item you report upward.
I instrument cost per resolved task rather than per call, because the per-call number hides retries and escalations. Then the three levers: caching the stable prefix, since cached reads bill at a fraction of fresh input, moving anything that does not need a live answer to batch processing at half price, and routing easy traffic to a smaller model with a measured escalation rate.
Make it yours: Bring a real before-and-after figure if you have one, and say which lever moved it most.
How do you build an evaluation set for a problem nobody has benchmarked?
What they are looking for: Whether you have built one or read about one. The difference shows in whether you start from data or from a rubric.
I start from logs, not imagination: twenty to fifty real failures, grouped into a taxonomy after reading them rather than before. Deterministic graders wherever the answer has a checkable form, a model judge only where nuance demands it, and that judge calibrated against human labels first. Then I report reliability across repeated trials rather than a single pass.
Make it yours: Describe a grader of yours that turned out to be wrong, and how you noticed.
Tell me about a time you worked across a seam between two specialisms.
What they are looking for: Whether you can operate outside your own discipline, which is what mid-career actually buys.
A research result we wanted to ship needed four times our latency budget. Rather than argue, I sat with the researcher and we found that most of the gain came from one component, which we could run asynchronously. We shipped that part, kept the budget, and wrote down the rest as a known gap with a number attached.
Make it yours: Pick a case where you gave something up. Seam stories where you simply won read as politics, not collaboration.
What security failure worries you most in the system you run today?
What they are looking for: Specificity, and whether you treat security as design rather than as a filter bolted on afterwards.
Indirect prompt injection through retrieved content, because our corpus accepts documents from outside the company and the agent has a write tool. The defence is architectural rather than a classifier: untrusted content cannot reach a tool that changes state, destructive actions are gated behind confirmation, and every tool has its own least-privilege credential.
Make it yours: Name your own system's weakest leg. A candidate with no concerns reads as someone who has not looked.
Senior-Level
Lead, Govern or Specialise
The three directions senior AI work goes: deeper technical specialism, management of people and spend, or the governance and regulatory layer that now gates enterprise deployment.
What you should know at this stage
The senior routes are different labour markets, and the regulatory one is the newest.
Management Is a Different Occupation, With Different Pay
Computer and information systems managers are a 685,800-job occupation with a $175,140 median in May 2025, projected to grow 16 percent to 2035 with about 53,500 openings a year. Heads of AI, directors of machine learning and most engineering leaders are counted here rather than under any technical occupation.
It is worth treating as a change of profession rather than a promotion. The work becomes headcount, vendor contracts, model spend and the governance paperwork, and the hands-on hours fall away. People who move expecting a senior version of their old job usually leave within two years.
Research Pay Depends on the Industry, Not the Title
Computer and information research scientists had a $140,300 median across 38,600 jobs in May 2025, but the industry split is the real story: federal government employs 31 percent of them at a $124,870 median, computer systems design 16 percent at $162,760, colleges and universities 7 percent at $85,460, and software publishers 3 percent at $211,270.
The same credential therefore supports a doubling of pay depending on where it is used. At senior level the decision about sector is worth more than the decision about seniority, and it is far harder to reverse.
Governance Frameworks Are Now Procurement Questions
NIST's AI Risk Management Framework, released in January 2023 for voluntary use, organises the work into four functions: Govern, Map, Measure and Manage. The Generative AI Profile, NIST AI 600-1, followed in July 2024 and names the risks specific to generative systems, and a Critical Infrastructure Profile concept note followed in April 2026.
Enterprise buyers increasingly ask which framework a supplier maps to before they ask about benchmarks. Senior people who can answer that question in the buyer's own vocabulary shorten sales cycles, which is a commercial argument rather than a compliance one.
The EU AI Act Timeline, Stated Correctly
Most summaries get this wrong in one direction or the other. The prohibitions and the Article 4 AI literacy duty have applied since 2 February 2025. The general-purpose model obligations have applied since 2 August 2025, with models already on the market given until 2 August 2027. Article 50 transparency, covering disclosure of AI interaction, machine-readable marking of synthetic content and deepfake labelling, applies from 2 August 2026.
The high-risk obligations are later: Annex III systems from 2 December 2027 and Annex I products from 2 August 2028. Anyone planning a roadmap on a single August 2026 date is planning against a timeline that does not exist.
Skills employers look for
What an organisation is buying when it hires a senior lead rather than an experienced engineer.
Portfolio Judgement Across Model Bets
Senior people are paid to decide which bets to stop. That means holding a view on which capabilities are improving quickly enough to wait for, which are plateauing, and which problems should never have been given to a model. The AI Index is useful here precisely because it is a measurement project rather than a vendor roadmap.
Hiring Across Specialisms You Did Not Come From
A lead who arrived through engineering has to hire researchers, and one who arrived through research has to hire platform people. The federal data is a useful corrective to internal folklore about what each is worth: research proxies sit at a $140,300 median, developers at $135,980, data scientists at $120,230 and the managers above them at $175,140.
Risk Governance as a Standing Function
NIST's framework splits the work into Govern, Map, Measure and Manage, and the first of those is the one organisations skip. At senior level the deliverable is a standing process: who approves a deployment, how incidents are recorded, what gets re-measured after a model version changes, and who owns the answer when a regulator asks.
Regulatory Translation
Article 50 requires that synthetic content be marked in a machine-readable form and that deepfakes be disclosed, from 2 August 2026, while high-risk duties arrive in December 2027 and August 2028. Turning those dates into an engineering backlog, rather than a legal memo nobody actions, is a senior skill with very few people in it.
Interview questions and answers
Senior questions, where the subject is the portfolio, the process and the regulator rather than the model.
How do you decide which AI projects to stop?
What they are looking for: Portfolio judgement. Senior leaders are measured on what they killed as much as on what they shipped.
I set a measurable bar before the work starts and a date to check it. If a project is not near the bar by then, the question is whether the gap is the method or the data, and the data answer is usually fatal. I would rather close three and fund one properly than keep six alive on hope, and I say that to the team at the start so the decision is not a surprise.
Make it yours: Give a real project you ended and what you did with the people on it. That second part is the test.
What is your governance process for putting a model-backed system live?
What they are looking for: Whether you have a standing process or a series of heroic one-off reviews.
We map to the four functions in NIST's framework: a named owner and an approval route under Govern, a documented statement of context and affected users under Map, a pre-launch evaluation and a post-launch monitor under Measure, and an incident route with a rollback under Manage. The part most organisations skip is Govern, and it is the part a buyer asks about first.
Make it yours: Say what your process rejected. A governance process that has never blocked anything is documentation.
Your product ships into the EU. What is on your compliance roadmap and when?
What they are looking for: Whether your regulatory knowledge is current. Most candidates are working from a summary written two years ago.
Several separate dates, not one. AI literacy and the prohibitions already apply. General-purpose model obligations have applied since August 2025. Article 50 transparency, which means disclosing AI interaction, marking synthetic output in a machine-readable way and labelling deepfakes, is August 2026. High-risk duties are December 2027 for Annex III and August 2028 for Annex I, so they are a roadmap item rather than a fire.
Make it yours: Separate the dates explicitly. Collapsing them into one is the mistake the question is designed to find.
How do you hire for a specialism you have never worked in yourself?
What they are looking for: Humility and method. A lead who hires only in their own image builds a team with one shape of failure.
I borrow a calibrated interviewer from that discipline and let them own the technical signal, while I own the judgement and collaboration signal. I write down in advance what evidence would convince me, so I cannot be talked round by fluency. And I benchmark the band against published occupational data rather than internal folklore, because that is where the inequities come from.
Make it yours: Name who you borrowed and what they caught that you would have missed.
Why AI Professionals Need to Stay Informed
No other field changes its own working assumptions this fast. A method published this quarter can be the default next quarter, a price change can rewrite what is affordable to build, and a regulatory date can move a roadmap. Staying current is not background reading here. It is the part of the job that decides whether your judgement is still worth asking for.
Read more
The second reason is commercial. Enterprise buyers now ask which risk framework you map to and which compliance dates apply before they ask about accuracy, and those answers change more often than the models do.
The bottom line
AI Careers Are Splitting Into Specialisms. Choose Yours.
The federal projections point in five different directions at once: data scientists up 35 percent to 2035, research scientists 22 percent, security analysts 21 percent, systems managers 16 percent and software developers 10 percent from a far larger base. The opportunity is real, and it rewards being specific.
Stay Informed with Hark's AI Briefing
How it works
Curated from the best newsletters
Hark's editorial team handpicks the sources worth following. See how we rank
The news that matters in 5 minutes
Hark distills the most important stories into a concise audio briefing.
Texted to you every morning
You set the time and listen when you want.
Go deeper if you want
Upgrade to the 15-minute briefing and select up to 50 newsletters.

The 10 AI Newsletters We Handpicked
The Rundown AI
Rowan Cheung
- Free
- Daily
- 2M readers
Rowan Cheung's daily five-minute rundown of AI news and tools, teaching 2 million+ readers what just shipped and how to actually use it.
Why AI professionals read it
The largest AI newsletter, a five-minute daily on what launched and how to use it. Even for a specialist it is the quickest read on what the wider market now expects AI to do.
More about this newsletter
Rowan Cheung's The Rundown AI is one of the largest AI newsletters, a daily digest of model launches, tools, and industry news built for a broad audience. Each issue turns the day's biggest release into a workflow you can try the same morning. Breadth over depth, delivered fast.
Best for: Anyone who wants to keep up with AI in five minutes a day. It's the largest AI newsletter for a reason: every issue turns the day's biggest model release or tool launch into something you can actually try that morning, no technical background needed.
Superhuman AI
Zain Kahn
- Free
- Daily
- 1.5M readers
Zain Kahn's daily briefing on AI news, tools, and how-tos, engineered for busy professionals and read by 1.5 million+ subscribers.
Why AI professionals read it
Zain Kahn's daily on AI news, tools and how-tos, built for professionals rather than researchers. Good for tracking the tooling layer that turns a model release into something a team can actually adopt.
More about this newsletter
Zain Kahn's Superhuman AI is built around using AI at work: practical tool tips, prompts, and news framed for productivity in a quick daily read. It is the fastest scan of the big AI dailies. Best for professionals who want practical how-to alongside the headlines.
See the other 8 newsletters
TLDR AI
TLDR
- Free
- Daily (weekdays)
- 1.1M readers
TLDR's daily AI edition compresses research papers, product launches, and industry news into a five-minute weekday read for over a million subscribers.
Why AI professionals read it
The most technical of the big AI dailies, compressing the day's research papers, model launches and industry news into a five-minute scan written for engineers. The fastest way to be sure nothing that shipped overnight got past you.
More about this newsletter
TLDR AI compresses the day's AI research, product launches, and industry news into one-sentence summaries you can scan in minutes. It is the most technical of the big AI dailies, written with engineers in mind. Efficient, dense, and no-nonsense.
The Neuron
The Neuron (TheurgyOne)
- Free
- Daily (weekdays)
- 700k readers
A daily tour of AI news and tools written for working professionals, translating the industry's firehose into practical takeaways for 700,000+ readers.
Why AI professionals read it
A weekday tour of AI news translated into practical takeaways for working professionals. It is the version of the firehose you can forward to a colleague who does not follow the field.
More about this newsletter
The Neuron explains AI news and tools in plain English with a light touch, for readers who want to understand what matters without a computer science degree. It is the most approachable of the daily AI digests. Friendly, clear, and genuinely useful for beginners.
TLDR
TLDR Newsletter
- Free
- Daily
- 8M readers
Five-line summaries of the day's engineering, startup, and science links.
Why AI professionals read it
Five-line summaries of the day's engineering, startup and science links. It keeps an AI specialist attached to the rest of technology, which is easy to lose sight of.
More about this newsletter
Five-line summaries of the day's tech, startup, and programming links, chosen for engineers rather than investors. No commentary and no filler; the TLDR homepage claims more than eight million readers.
The Download
MIT Technology Review
- Free
- Daily
- 250k readers
MIT Technology Review's daily dose of what matters in emerging technology, written with institutional rigor.
Why AI professionals read it
MIT Technology Review's daily on emerging technology, sober and hype free. It puts your field in the context of the ones it is about to collide with.
More about this newsletter
MIT Technology Review's daily briefing distills what matters in emerging technology, drawing on the magazine's own reporting on AI, biotech, climate, and computing. It is institutional, sober, and reliably free of hype. One of the few tech dailies with a real newsroom behind it.
Best for: Readers who want emerging tech coverage from a real newsroom: sober, hype free daily reporting on AI, biotech, climate, and computing from MIT Technology Review.
TechCrunch Daily News
TechCrunch
- Free
- Daily
TechCrunch's flagship roundup of the day's top startup and technology stories, delivered every weekday and Sunday.
Why AI professionals read it
TechCrunch's weekday roundup of funding, launches and industry news. It tells you which AI companies just raised and which are quietly consolidating.
More about this newsletter
The best of TechCrunch's daily coverage (funding rounds, product launches, and tech industry news) compiled by the site's editors. Arrives every weekday plus a Sunday edition. Free, alongside TechCrunch's other newsletters (Startups Weekly, Mobility, Week in Review, StrictlyVC).
The Pragmatic Engineer
Gergely Orosz
- Free
- Frequent
- 1.1M readers
Deeply reported dispatches on how Big Tech and startups actually build software, from an ex-Uber engineering leader.
Why AI professionals read it
Deeply reported dispatches on how Big Tech and startups actually build software, from an ex-Uber engineer. The best account anywhere of the engineering reality your models ship into.
More about this newsletter
Gergely Orosz, formerly an engineering manager at Uber, reports on how software actually gets built inside Big Tech and startups: engineering culture, compensation, layoffs, and deep dives into real systems. His reporting on tech companies is read industry-wide. The deep dives are paid, but the free weekly issues keep casual readers current.
TLDR Dev
TLDR (TLDR Media)
- Free
- Daily
- 470k readers
Daily software engineering news, tools, and web dev links in a five minute read.
Why AI professionals read it
Daily software engineering news, tools and web dev links in a five minute read. The practitioner's feed, and a useful counterweight to research-heavy AI reading.
More about this newsletter
TLDR Dev is the software engineering edition of the TLDR network: one daily email of programming news, new tools, and web development links, each with a one-paragraph summary. It is aggregation, not commentary, tuned for working developers across the stack. Best for engineers who want the day's dev news skimmed in five minutes.
ByteByteGo Newsletter
Alex Xu (ByteByteGo)
- Freemium
- 1M readers
System design newsletter by Alex Xu, co-author of the System Design Interview books, explaining complex architectures in plain language with heavy use of diagrams.
Why AI professionals read it
Alex Xu's system design newsletter, explaining complex architecture in clear diagrams. Serving models is a systems problem, and this is the clearest teaching on systems.
More about this newsletter
Issues break down how large systems work (Docker internals, network protocols, the tech stacks behind companies like Netflix) in the same diagram-first style as Xu's bestselling system design books. Hosted on Substack with a free tier and paid deep-dive content; the site claims over 1,000,000 readers.
You May Also Like
- AI EngineersThe applied end of the field in full depth: retrieval, agents, evaluation harnesses and inference cost, with the interview loop that now tests all four.
- Machine Learning EngineersTraining infrastructure, model serving, MLOps and drift monitoring, which is the track that decides whether anything the rest of the field builds actually stays up.
- AI Research ScientistsArchitectures, reinforcement learning, interpretability and benchmarks, plus an honest account of how much the pay depends on sector rather than seniority.
- AI Product ManagersWhere the quality bar gets set and the roadmap gets argued, covering evaluation, trust and safety, and the two routes in from product and from engineering.
Frequently Asked Questions
Entry-Level
What does it actually mean to work in AI in 2026?
It means one of about six jobs, not one. Research scientists design and probe methods. Machine learning engineers make training and serving reliable and affordable. AI engineers build retrieval, agents and evaluation on models somebody else trained. Product managers decide what is worth building and what counts as good enough. Consultants judge whether an organisation can carry the promise. Governance and security specialists answer for what the system does once it is live. They share vocabulary and very little else day to day, and the first useful decision in an AI job search is choosing which of them you are applying for.
Is working in AI a good career in 2026 and beyond?
On the federal numbers, yes, with a caveat about breadth. There is no AI occupation, so the field is counted as data scientists, growing 35 percent to 2035 at a $120,230 median, software developers at $135,980 growing 10 percent from a 1,717,800-job base, computer and information research scientists at $140,300 growing 22 percent, information security analysts at $129,180 growing 21 percent, and computer and information systems managers at $175,140 growing 16 percent. Lightcast's work for the 2026 Stanford AI Index puts AI skills in 2.5 percent of United States postings, up 55 percent in a year. Real growth, narrow market.
Do you need a degree, and does it have to be in computer science?
It depends which route you want. The Bureau of Labor Statistics records a bachelor's degree as typical entry-level education for software developers, data scientists and information security analysts, and a master's for computer and information research scientists, where doctorates are common in practice. For applied engineering, product and consulting the degree gets your application read and the portfolio decides the outcome. For frontier research the graduate degree is close to a filter, and the years it costs are the real decision rather than the tuition.
How do AI engineers, machine learning engineers and research scientists differ?
By what they own. An AI engineer builds products on a model somebody else trained, so their problems are retrieval quality, agent design, evaluation and cost. A machine learning engineer owns the pipelines, training runs and serving infrastructure, and is usually on call for them. A research scientist owns a line of questions and is judged on results that survive someone else reproducing them. Our AI engineers, machine learning engineers and AI research scientists guides each cover one of those ladders in full depth.
What skills and tools should a beginner learn first?
Python and SQL remain the baseline rather than a differentiator. Python was named in 258,674 United States AI postings in 2025, about 30 percent more than in 2024, and almost every applied problem starts with getting the right rows out of somebody's warehouse. After that, learn to read an evaluation before believing it: what was measured, on what data, against what baseline, and at what cost. Then learn the two cost levers, caching and batch processing, because a prototype nobody can afford to run is not a portfolio piece.
How do you build an AI portfolio that gets you hired without experience?
Build one artefact a stranger can run, not five demos. It needs a readme, a reproducible setup and an honest account of its failure rate measured on cases you labelled yourself. Fifty real questions scored by hand beats any benchmark number you can quote. Say what it gets wrong and why, because interviewers assume every entry project is fragile and are testing whether you know where. A candidate who can quantify their own error rate is immediately above the median.
What are the best entry-level jobs for getting into AI work?
The reliable routes in are data or platform engineering, where pipeline work is permanently under-supplied and teaches you the data layer everything else depends on; junior software engineering on a team that already ships AI features; and analyst or associate product roles where you write the evaluation criteria. Research internships are the route into research and are far more competitive, with only 38,600 computer and information research scientist jobs in the whole country. Going in sideways through the data layer is underrated and works.
Mid-Level
How much do AI professionals earn, and which route pays most?
There is no blended AI salary in the federal data, so read the adjacent occupations. The May 2025 medians are $120,230 for data scientists, $129,180 for information security analysts, $135,980 for software developers, $140,300 for computer and information research scientists and $175,140 for computer and information systems managers. The spread inside each is wide: the lowest tenth of data scientists earned under $67,240 and the highest tenth over $199,130. A lateral move between occupations often pays more than a promotion inside one.
Which AI specialism should you commit to by mid-career?
Decide by the variables you want rather than the subject you like. How much of your week do you want near a model versus near a customer. How tolerant are you of being on call. Do you want to be judged on reproducibility, on uptime, on a product metric or on a governance outcome. Growth also differs sharply: data scientists are projected to grow 35 percent to 2035, research scientists 22 percent, security analysts 21 percent and software developers 10 percent from a much larger base. Small base plus high growth means competitive to enter and slow to hire.
How is agentic AI changing what employers hire for?
It is the fastest-growing skill cluster in the market and still a small one. Lightcast's analysis for the 2026 AI Index put agentic AI skills at 0.06 percent of United States postings in 2024 and 0.23 percent in 2025, an increase of more than 280 percent and roughly 90,000 postings. Because the experience is at most a couple of years old anywhere, anyone who has actually run a tool-using system in production, and can describe its retry policy, approval gates, compaction strategy and cost per resolved task, is arguing from evidence most candidates do not have.
What certifications or further study are worth it in AI?
Less than the marketing suggests, with a few exceptions. Cloud and ML platform certifications are genuinely read for platform and operations roles, because they map to infrastructure an employer already runs. A graduate degree is close to mandatory for research posts, where the federal data records a master's as typical entry. Everything else competes with a portfolio and usually loses. The one piece of non-technical study that pays disproportionately is fluency in NIST's AI Risk Management Framework, because enterprise buyers now ask about it directly.
How do you prepare for an AI job interview at mid-career?
Expect three themes. Cost: be able to say how you measure it, and say cost per resolved task rather than per call, since the per-call figure hides retries and escalations. Evaluation: expect to design a set for a problem nobody has benchmarked, starting from real logged failures rather than invented cases. Security: expect to name the weakest leg of your own system, usually untrusted retrieved content reaching a tool that changes state. Bring one seam story where you gave something up, not one where you won.
How much does AI security knowledge matter for a non-security role?
More than it used to, because these are design faults rather than operational ones. The 2025 OWASP Top 10 for Large Language Model Applications lists prompt injection first, then sensitive information disclosure, supply chain, data and model poisoning, improper output handling and excessive agency. Every one of those is decided when the system is built, so they land on the builder rather than on a separate team. The job market agrees: information security analysts are projected to grow 21 percent to 2035 at a $129,180 median.
What does a typical day look like across the different AI roles?
They diverge more than the shared job titles suggest. An AI engineer starts with overnight evaluation runs and the cost dashboard, then reads real failure traces before changing anything. A machine learning engineer starts with pipeline health and drift monitors and spends the day on latency budgets and capacity. A research scientist kills failing variants, designs the next ablation and negotiates compute. A product manager watches a customer use the thing. A governance lead reviews deployment requests and answers due-diligence questionnaires.
Senior-Level
What skills do employers expect from senior AI leaders?
Four, and only one is technical. Portfolio judgement: deciding which model bets to stop, and saying so before the sunk cost argument starts. Hiring across specialisms you did not come from, which usually means borrowing a calibrated interviewer and benchmarking bands against published occupational data rather than internal folklore. Risk governance as a standing process rather than a series of heroic reviews. And regulatory translation: turning a compliance date into an engineering ticket with an owner, which is a skill with very few people in it.
Should you move into management, and what does it cost you?
It is a change of occupation rather than a promotion. Heads of AI, directors of machine learning and most engineering leaders are counted as computer and information systems managers, a 685,800-job occupation with a $175,140 median in May 2025, growing 16 percent to 2035 with about 53,500 openings a year. The pay step is real, roughly $35,000 above the software developer median. So is the trade: the work becomes headcount, vendor contracts, model spend and governance paperwork, and the hands-on hours go. People who expect a senior version of their old job usually leave inside two years.
What AI career options exist outside building models?
Most of them, in fact. Product management of AI surfaces, consulting on readiness and use-case selection, governance and policy, AI security and red teaming, evaluation and measurement, technical sales and solutions engineering, developer relations, and research management. Several of these are not separately counted in federal statistics, which is part of why they are less competed for than engineering posts. The common requirement is the ability to hold a technical conversation and a commercial one in the same meeting.
How is AI changing the work of AI professionals themselves?
It moved the scarce skill from production to proof. Generating a working system is cheaper than it has ever been, so the differentiated work is demonstrating that it behaves: evaluation design, error analysis from real traces, measuring reliability across repeated trials rather than a single pass, and instrumenting cost per resolved task. The roles that grew out of that shift, evaluation engineering, context engineering and forward-deployed engineering, barely existed three years ago and are now named directly in postings.
What regulation do senior AI leaders need to track, and by when?
Separate the dates, because most summaries collapse them. Under the EU AI Act the prohibitions and the Article 4 AI literacy duty have applied since 2 February 2025. General-purpose model obligations have applied since 2 August 2025, with models already on the market given until 2 August 2027. Article 50 transparency, covering disclosure of AI interaction, machine-readable marking of synthetic content and deepfake labelling, applies from 2 August 2026. High-risk duties come later: Annex III from 2 December 2027 and Annex I from 2 August 2028.
How do you build an AI governance process buyers will accept?
Map it to a framework they already recognise. NIST's AI Risk Management Framework, voluntary since January 2023, splits the work into Govern, Map, Measure and Manage, and the Generative AI Profile released in July 2024 names the risks specific to generative systems. The function organisations skip is Govern, which is exactly the one procurement asks about: who approves a deployment, how incidents are recorded, what is re-measured after a model version changes, and who answers a regulator. A process that has never blocked anything is documentation.
What is the biggest structural risk to a career built on AI work?
Being counted in the wrong occupation, and being specific about the wrong thing. The first is fixable: understand which job family your employer files you under, because the band follows the filing rather than the title. The second is harder. The field's half-life is short, so expertise in a particular framework, provider or tool depreciates fast, while the durable skills, evaluation design, cost reasoning, risk governance and the ability to explain a system to someone who will not read code, do not. Invest accordingly.
Get Tomorrow's AI Briefing
TLDR AI alone reaches more than a million engineers who scan the day's models, papers and launches before work, and The Pragmatic Engineer is read by the leaders who decide what ships. In a field moving this fast, being last to hear is a real disadvantage.