Breaking into AI in 2026.

The question I get most often is some version of: what do I need to work in AI in 2026? Most advice gives everyone the same answer. Learn Python. Get a certificate. Network. That advice fails because a research scientist and a policy specialist don't share a single day-to-day skill. This article gives ten answers, one for each real AI role. The list is about fit, not permission. Only one of the ten doors requires a specific credential.

How I picked these ten, and how to read them

I built the list from three sources. The Stanford AI Index Report 2025 tracks which AI roles companies actually post and fill.^[1] LinkedIn's Economic Graph data and Indeed's Hiring Lab show which roles are growing fastest by posting volume.^[3][4] My own hiring and consulting experience supplied the two entries that don't have a government occupation code yet: conversation designer and prompt or context engineer.

Each role gets the same three answers, in the same order.Education that helps covers the shortest honest path in.Personality traits that fit covers only claims backed by peer-reviewed research or years of hiring. Foundational tech to have covers the smallest set of tools you can't skip. Each role ends with a first move you could start this month.

One note on personality. The Big Five model measures openness, conscientiousness, extraversion, agreeableness, and emotional stability. It is the most studied personality framework in workplace psychology. The Barrick and Mount meta-analysis found that conscientiousness, the follow-through trait, predicts job performance in nearly every occupation. Openness predicts how well people learn new skills.^[9] Read every trait here, including the "wrong fit" lines, as a matter of taste, not talent. They describe the daily work you'll enjoy. They say nothing about what you're capable of.

1. AI or ML research scientist

Education: PhD is the norm, with side doorsCode: heavyFirst move: reproduce one published result

These are the people who publish the papers, invent the model architectures, and work on AI safety at the frontier. Hiring is concentrated at a small number of labs, includingAnthropic,OpenAI,Google DeepMind, andMeta FAIR. This role goes first because it is the only one of the ten where a doctorate is the norm. The other nine doors are more open.

Education that helps

The standard path is a PhD in machine learning, computer science, statistics, or another quantitative field, with publications at the field's main research conferences, such asNeurIPS andICML. It is no longer the only path. Every frontier lab also hiresresearch engineers. They sit next to the scientists, build the experiments, and don't need a doctorate. Residency and fellows programs offer a third way in.Anthropic runs one, and several other labs do too. These programs exist to bring in strong people without the traditional credential. One habit matters more than the degree: pick a published result and reproduce it from scratch. People who do that a few times stop looking like applicants and start looking like colleagues.

Personality traits that fit

Very high openness to experience, the trait research links to scientific productivity. Very high curiosity. Kashdan and colleagues call the key pair "joyous exploration" and "stress tolerance."^[10] Patience matters too. Most days end with an experiment that didn't work, and the wins arrive months apart. You'll know it's the wrong fit if you need a visible win every week to feel like you're moving.

Foundational tech to have

• Python, fluently.
• PyTorch or JAX, the two main toolkits researchers use to build and train models.
• Linear algebra, calculus, and probability at a working level.
• A conceptual grasp of distributed training. That means knowing how big models get trained across many computers at once.
• An experiment tracker like Weights and Biases, so your results are reproducible.
• The habit of reading a paper on arXiv (the public site where AI research is posted) and running its code the same afternoon.

Your first move: pick one interesting published result and reproduce it end to end. Write up what surprised you. That single artifact says more than a transcript.

2. Machine learning engineer

Education: bachelor's or self-taught, plus a deployed projectCode: heavyFirst move: deploy one small model with monitoring

These are the people who put models intoproduction (the live system real users touch) and keep them there. The closest occupation the US Bureau of Labor Statistics tracks is data scientist. It is projected to grow 34% from 2024 to 2034, the fourth-fastest of any US occupation.^[2] This is the highest-headcount technical AI role.

Education that helps

A bachelor's in computer science or a related field is the norm, and it helps. Many machine learning engineers took another path. They came from software engineering and taught themselves throughfast.ai,DeepLearning.AI, orAndrej Karpathy's free courses. What matters most is putting a machine learning system in front of real users. A notebook of practice exercises doesn't count. A deployed model with a monitoring dashboard does, even a small one.

Personality traits that fit

Conscientiousness matters most here. The Barrick and Mount meta-analysis found it predicts job performance in nearly every occupation.^[9] In this work, the payoff shows up as carefully labelled data, careful logging, and careful thinking about failure modes nobody has hit yet. You'll know it's the wrong fit if "launch it and see what happens" is the pace you enjoy most.

Foundational tech to have

• Python, and SQL, the standard language for asking questions of a database.
• Docker, the standard tool for packaging software so it runs anywhere, and one cloud platform: AWS, Google Cloud, or Azure.
• PyTorch or TensorFlow. PyTorch is the safer bet. It leads 48% to 44% among Python developers surveyed in 2024.^[6]
• Software engineering hygiene: version control with Git (the system that tracks every change to code), testing, code review.
• Familiarity with the shared systems teams use to store model inputs and track model versions. You'll hear "feature store" and "model registry."
• The judgement to say don't use machine learning for this when it's true.

Your first move: train one small model on data you care about, deploy it anywhere public, and put a monitoring dashboard on it. Then improve it monthly and keep notes. That's the portfolio.

3. Applied AI or software engineer

Education: degree or bootcamp, portfolio decidesCode: heavyFirst move: build one LLM feature end to end

This is where the biggest hiring surge sits in 2026. These engineers don't train models. They build features on top of them. Every software company now needs someone who can wire an LLM (large language model, the kind of AI behind ChatGPT and Claude) into a product without making it slow, expensive, or embarrassing. Indeed's Hiring Lab finds software development among the occupations that mention AI most often in job postings.^[4] This role has the widest set of open doors on the list.

Education that helps

A bachelor's in computer science helps. So does a bootcamp. In my hiring experience, a strong GitHub portfolio settles the question either way. Employers want someone who can read the model documentation from OpenAI, Anthropic, or Google and turn it into a working feature by Thursday. Show them you've already done it.

Personality traits that fit

Openness to new tools, because the model landscape changes every few months. Yesterday's best practice is often today's cost problem. Product taste, meaning you notice when an interaction feels off, not just when a test fails. Emotional stability, because the AI feature you launched last Tuesday might behave differently on Friday after a provider updates its model. You'll know it's the wrong fit if you're happiest with a fixed spec that never moves.

Foundational tech to have

• Python and TypeScript, now the two most-used languages on GitHub. AI work pushed both to the top.^[5]
• REST APIs, the standard way software services talk to each other.
• AI toolkits (SDKs) from at least two model providers, so you understand what varies between them.
• Basic prompt literacy: writing clear instructions a model can follow reliably.
• Basic eval literacy: a small test suite for AI output, so you know whether a change made things better or just seemed better in your demo.
• A monthly habit of reading model release notes.

Your first move: build one LLM feature end to end (a summarizer, a smart search box, a draft generator), deploy it, and write a short post about the tradeoffs you hit. Repeat next month.

4. AI data engineer

Education: bachelor's or respected certificationsCode: heavyFirst move: build one retrieval pipeline over messy documents

The AI data engineer is growing out of the classical data engineer role. The core skills are the same. The difference is where the day goes: building the document sources, evaluation datasets, and data pipelines that feed AI systems.RAG(retrieval-augmented generation) is the technique where an AI looks facts up in your documents before answering. It became a default architecture over the last two years, and it made this role one of the highest-impact seats on any AI team.

Education that helps

A bachelor's in computer science, mathematics, or statistics is the norm, and it helps. Certifications fromDatabricks,Snowflake, ordbt Labscarry weight too, because hiring managers can map them to skills. What matters most is designing data that still makes sense to a stranger six months later.

Personality traits that fit

Conscientiousness again. The difference between a working pipeline and an outage is often a check somebody chose to add. Comfort with ambiguity also matters. "What does clean data mean" is a business question dressed up as a technical one, and the answer is yours to propose. You'll know it's the wrong fit if you'd rather wait for a complete spec than draft one yourself.

Foundational tech to have

• SQL, fluent, not scripted. Plus Python.
• One workflow orchestrator: Airflow, Dagster, or Prefect. That's the tool that runs your data pipelines on schedule.
• One transformation tool (dbt or SQLMesh) and one data warehouse (Snowflake, BigQuery, or Databricks). The warehouse is the central home for a company's analytics data.
• One vector database: pgvector, Pinecone, or Weaviate. A vector database searches by meaning instead of exact words.
• An understanding of chunking, which is how you split documents so an AI can search them well. That choice changes answer quality.
• Enough eval literacy to know when a retrieval change actually helped.

Your first move: take a messy pile of real documents (your own notes, a public dataset, a friend's small-business files) and build a clean, searchable pipeline over them. Document your decisions.

5. MLOps or AI platform engineer

Education: ops and DevOps backgrounds map cleanlyCode: heavyFirst move: write one runbook or cost story with numbers

These are the people who keep AI systems alive in production. MLOps is shorthand for machine learning operations. In a 2026 org chart, this team often runs the AI cost dashboards, the monitoring stack, and the on-call rotation for AI incidents. McKinsey's State of AI survey finds most organizations still haven't scaled AI across the enterprise, and the blockers are operational, not model quality.^[7] That's the job description.

Education that helps

A background in site reliability, DevOps, or platform engineering (the disciplines that keep software running day and night) maps cleanly, degree or not. A computer science degree is common, but so is moving over from IT operations without one. What matters is visible infrastructure work: a server setup you maintain in code, a runbook you wrote (the step-by-step guide for when a system breaks), or a cost story you can tell with numbers.

Personality traits that fit

Emotional stability. That's the Big Five term for staying clear-headed when the alerts fire at 2am. Systems thinking, the habit of asking "what else does this touch" before changing anything. Conscientiousness, because the job is checklists and boring correctness, and everyone else's work stands on it. You'll know it's the wrong fit if you'd always rather build the new thing than keep the important thing running.

Foundational tech to have

• Linux at a real level, not just as a place to type commands.
• Kubernetes, the standard system for running software across many machines at once. Terraform or Pulumi, which describe your servers in files instead of clicks.
• GPU basics: the specialized chips AI models run on, and what makes them expensive.
• One observability stack for watching live systems. Prometheus and Grafana are the common pair, and OpenTelemetry is becoming the standard.
• AI cost dashboards you can read and edit, plus request tracing with a tool like LangSmith or Arize.
• A working sense of the serving tricks that keep responses fast and bills small (batching requests, caching repeated work).

Your first move: pick any system you can access, write the runbook it doesn't have or the cost breakdown nobody has done, and share it. Operations people who write clearly are rare and get hired.

6. AI product manager

Education: any degree helps, none requiredCode: read it, don't write itFirst move: write a one-page spec with three evals

The product manager is where AI-literate humanities backgrounds are quietly winning. The OECD finds the occupations most exposed to AI are skilled, white-collar ones, not just technical jobs.^[8] In my experience, the people who gain influence in that shift are the translators between business and machine. AI product management is the translator job.

Education that helps

Any degree that taught you to write and reason clearly helps here. None is required. Strong AI product managers come from journalism, philosophy, operations, and a dozen other fields. What the role does require is fluency in reading code, reading evals, and reading model documentation. A product manager who can't explain what an eval is can't spec an AI feature that survives contact with production. That fluency is learnable in months, not years.

Personality traits that fit

Conscientiousness, for the specs and the follow-through. Agreeableness in the Big Five sense, for working across engineering, design, legal, and leadership. Enough emotional stability to say no with a smile. Barrick and Mount found agreeableness matters most in roles built on team coordination.^[9] That's this job. You'll know it's the wrong fit if writing things down for other people feels like overhead instead of the work.

Foundational tech to have

• The ability to read code, even though you won't be the one writing it.
• SQL at a level where you can pull your own metrics.
• Prompt design at a working level.
• Eval literacy: explaining to a stakeholder, without hedging, how the team measures whether the AI is doing its job.
• A once-a-month habit of reading one AI paper end to end, even if it takes two afternoons.
• A planning tool you actually keep current: Notion, Linear, or your team's equivalent.

Your first move: pick an AI feature you use and wish were better. Write a one-page spec (the short plan a team builds from) for the improvement, including three evals that would prove it worked. That document is your portfolio piece.

7. Conversation designer

Education: humanities backgrounds welcomeCode: light, but read logs comfortablyFirst move: redesign one bad bot conversation

The conversation designer shapes how an AI talks to a human. The discipline is distinct from prompt engineering, though the two overlap. A prompt engineer optimizes what goes into the model. A conversation designer shapes the whole interaction: turn structure, tone, how the AI declines, how it recovers from errors, and how it hands off to a human. In my hiring experience, a strong conversation designer movescontainment (the share of conversations resolved without a human) and satisfaction more than any model swap does.

Education that helps

Linguistics, communications, human-computer interaction (HCI), cognitive science, English, and journalism are all strong foundations. Humanities training pays back fast here. The questions of the job are human ones. What does this person actually mean? What did we accidentally teach them to expect? What happens at the least convenient moment? A masters isn't required. A portfolio of designed conversations, with your reasoning attached, is.

Personality traits that fit

High openness, so you'll read what real users actually type instead of what you wish they'd typed. High agreeableness in the Big Five sense, the empathy to design for the person who is frustrated at 2am. High conscientiousness, because good dialog design is careful specification work. You'll know it's the wrong fit if the happy path is the only part of the conversation you enjoy designing.

Foundational tech to have

• Reading conversation logs comfortably, the records of what real users typed and how the AI replied. That's where the real conversations live.
• Basic regex, the pattern-matching mini-language for searching text.
• A working idea of tokens, the word-pieces models actually read, and why they limit what fits in a conversation.
• One dialog design or orchestration tool: Botpress, Kore.ai, Voiceflow, or a custom setup.
• Prompt design at a working level, because the tone of the system message, the standing instructions an AI gets before a conversation starts, is conversation design too.
• Eval literacy, because a conversational feature that isn't measured will quietly regress the minute someone fixes something else.

Your first move: find a chatbot that frustrated you, transcribe the conversation, and redesign it turn by turn with notes on why. Three of those and you have a portfolio.

8. Prompt or context engineer

Education: any background, portfolio-firstCode: mediumFirst move: build a small RAG app this weekend

This role didn't have a name five years ago. Today, AI engineer is LinkedIn's fastest-growing job title for the second year running.^[3] Context work is a large slice of what those postings ask for. One version of the job is fading: memorizing clever phrasings and selling them as secrets. The version that's growing iscontext engineering: assembling the right documents, instructions, tools, and examples so a model has a fighting chance at the task. I've written more about this shift inContext Engineering.

Education that helps

Any technical or writing-heavy background works. This is one of the friendliest roles on the list to enter on a portfolio alone. You can build a smallRAGapplication in a weekend, iterate on it for a month with a small test suite, and have a real story to tell. No masters needed. What you do need is a project that worked, and the ability to explain in plain language why it worked.

Personality traits that fit

Curiosity, in the Kashdan sense of joyous exploration. Most of the job is trying things and reading the results carefully.^[10] Healthy skepticism, the willingness to check whether an improvement was real signal or noise. Conscientiousness, because measurement discipline separates the practitioners who improve from the ones who spin. You'll know it's the wrong fit if you'd rather show the impressive demo than run the boring test.

Foundational tech to have

• Python or TypeScript.
• One vector database, the searching-by-meaning kind described in role 4.
• One eval harness: Ragas, Braintrust, LangSmith, or a hand-rolled one. That's the test suite that scores your AI's answers.
• RAG architecture at a level where you can reason about how document splitting and search choices change answer quality.
• Enough cost and latency (response speed) math to predict what a change does to the bill.
• The hard-won knowledge that just add it to the instructions is often the wrong answer.

Your first move: build a small RAG app over documents you know well, this weekend. Then spend a month making it measurably better, and write down how you measured.

9. AI safety and red-teaming engineer

Education: security or ML background, flexibleCode: medium to heavyFirst move: read the NIST framework, try a public red-team challenge

The people whose job is to find out how an AI system breaks before the world does. Red-teamingmeans attacking your own system on purpose to find the weak spots first. Hiring has grown sharply here for two reasons. Regulators and large customers now ask for safety documentation. And major AI labs published scaling policies that require continuous testing of model behavior against specific risks. The work rewards people who notice what everyone else missed.

Education that helps

The three most common paths are offensive security (the attacker's side of cybersecurity), adversarial machine learning research, and an alignment-focused degree. The field is young enough that determined outsiders still get in. A bachelor's is common, and a doctorate helps in research labs. Neither is required if you can show real red-team work, and public challenges exist for building that record. The National Institute of Standards and Technology'sAI Risk Management Frameworkis the shared vocabulary of this role. Reading it end to end is the fastest way to sound fluent.^[11]

Personality traits that fit

Comfort delivering bad news. Constructive contrarianism, meaning you're willing to be the person in the room who says "this isn't safe yet" and can say it in a way people can hear. Very high conscientiousness, and the emotional steadiness to imagine worst cases all day without carrying them home. You'll know it's the wrong fit if keeping the peace matters more to you than naming the problem.

Foundational tech to have

• Python.
• A working knowledge of jailbreaks and prompt injection, the attacks that trick an AI into ignoring its instructions.
• One model-behavior testing framework: Inspect, HELM, or a hand-rolled one.
• Enough security fundamentals to build a threat model. That's a structured list of what could go wrong and who would exploit it.
• Fluency in the vocabulary of the NIST AI Risk Management Framework and the EU AI Act, because those two documents define "compliant" when a customer asks.^[11][12]

Your first move: read the NIST framework, then try one of the public AI red-teaming challenges online (Gandalfis a free one) and write up what you found. A documented break is a portfolio piece.

10. AI ethics, policy, and governance specialist

Education: law, policy, philosophy, or similarCode: none, but read it fluentlyFirst move: write a model card for a product you know

Not the person who writes the "responsible AI" blog post. The person who writes the risk register (the running list of what could go wrong and who owns it), themodel card (the short document explaining what a model can and can't do), and the answer to the customer's compliance questionnaire. The EU Artificial Intelligence Act entered into force in 2024. It created a regulatory floor that requires this role inside any organization deploying AI at scale in the European market.^[12] Demand is going up, not down. This is one of the clearest doors into AI for people who think in arguments rather than algorithms.

Education that helps

Law, public policy, philosophy, science and technology studies, or an interdisciplinary AI ethics program. A growing number of universities offer AI policy tracks. The tracks with a technical component, such as a required semester of machine learning fundamentals, produce the strongest graduates. A law degree helps in regulated industries. A degree in any argument-driven field transfers directly.

Personality traits that fit

Verbal fluency, written and spoken, because most of the job is translating between legal, engineering, product, and leadership people who don't share a vocabulary. Patience for ambiguity. Agreeableness in the Big Five sense, because the role only works if the other teams want to talk to you. You'll know it's the wrong fit if "it depends" is an answer you can't sit with.

Foundational tech to have

• Reading a model card without flinching, and following an eval report.
• A plain-English grasp of what a training set is and how models learn from one.
• The ability to explain the difference between GDPR (Europe's data privacy law), the EU AI Act, and the NIST framework to a non-lawyer.
• Threat-modelling basics, the structured what-could-go-wrong exercise from role 9.
• Familiarity with at least one AI incident database (the AI Incident Database is the most useful).
• A comfortable relationship with spreadsheets.

Your first move: pick a product you know well and write the model card it should have. You'll learn the vocabulary by using it, and you'll have a work sample no course certificate can match.

The pattern under the ten

Only one of the ten roles expects a doctorate, and even that one now has side doors. Seven of the ten share the same core tech stack: Python, SQL, one cloud platform, and one AI provider's toolkit, plus role-specific tools. The other three (product, design, and policy) don't ask you to write code for a living. They do ask you to read it comfortably and to understand evals as more than a testing step.

The personality picture is simpler. Conscientiousness helps in every role. Barrick and Mount predicted that decades ago, and every replication since has confirmed it.^[9] Openness and curiosity predict how fast you'll learn the parts of the job that change, and in AI that is most of the job.^[10] The role-specific traits are a matter of fit, not talent. Be honest with yourself about fit before you spend a year retraining. If the day-to-day of a role sounds like a chore in advance, pick a different door on this list.

If I could give one piece of advice to someone starting from scratch in 2026, it would be this. Pick one role. Build one small thing that role would do at work. Write about what you learned. Repeat, monthly, for a year. That will do more for your career than any certificate or bootcamp. The AI industry is not short on people who took the course. It's short on people who finished the thing.