Breaking into AI? A complete guide to core roles in AI companies

“AI didn’t replace you. The people who learned AI before you did.”

Prologue: A Tale of Two Worlds

It’s 2 a.m., and the lights are still on in a high-rise in Shanghai’s Lujiazui financial district. Chen, a 32-year-old product manager, has just received a “mutual termination” notice from HR, his team has been cut from 28 people down to 6. At that very moment, less than 3 kilometers away in Zhangjiang Science City, a startup building multimodal foundation models has just sent out its 47th offer letter, dangling a ¥1.5 million annual package plus equity in front of a 26-year-old algorithm engineer.

This isn’t an isolated incident. This is the most honest cross-section of the 2026 job market.

Look around. Growth in the traditional internet sector has dried up. Roles in finance, construction, administration, junior translation, and basic customer service are teetering on the edge of automation. Data from Liepin and Maimai shows that starting in the second half of 2025, headcount at traditional internet giants has shrunk by over 30% overall, while demand for AI-related roles has surged 178% against the trend.

Meanwhile, the AI sector is a money magnet: foundation models are migrating from labs into factories, and agents are taking over complex business workflows, from writing code, running financial reports, and handling customer support, to performing end-to-end legal contract reviews. Valuations of OpenAI, Anthropic, ByteDance’s Doubao, DeepSeek, Moonshot AI, and Zhipu have more than doubled in the past 18 months. A million-yuan salary for new graduates is no longer news, it’s the industry baseline.

If you’re standing at a career crossroads and feeling anxious, that’s your intuition trying to tell you something: the ticket to the old world has expired, and AI is the only Noah’s Ark heading into the future.

But take note, this ark isn’t free. The price of admission is your willingness to keep learning.

The Truth About the Cold Front: AI Isn’t Just Coming for Entry-Level Jobs

Many people mistakenly believe AI will only replace assembly-line work. The biggest shift between 2025 and 2026, however, is this: white-collar workers, middle managers, and even highly paid specialists are being systematically squeezed by AI.

• Legal: A top-10 U.S. law firm announced layoffs of 40% of its paralegal staff. The reason: GPT-5 paired with a domain-specific legal model can complete contract comparison reviews in 3 minutes, work that used to take 8 hours.
• Finance: JPMorgan’s “COIN” system already replaces an estimated 360,000 hours of compliance review work. A Chinese securities research institute reassigned half of its junior analysts to “AI trainer” roles.
• Design: A Shanghai-based e-commerce operations agency shrank its visual design team from 60 people to just 12 in Q1 2026. The remaining staff are mostly responsible for “directing the AI” and quality-checking its output.
• Media: A leading new-media company built a fully automated agent pipeline using Coze, covering topic selection, image sourcing, drafting, and distribution, reducing its operations team from 200 people to 35.

The cold, hard truth: AI doesn’t discriminate by job title. It only cares whether the work can be structured. Any task whose output can be defined, evaluated, and given feedback on is one AI can learn to do.

The good news: the more old jobs AI swallows, the more new ones it spawns. For every position AI replaces, 2-3 brand-new “AI collaborator” roles emerge. The only question is whether you’re willing to become one of those collaborators.

The Four Core Roles in AI: A Detailed Learning Roadmap

Don’t let the letters “A-I” intimidate you. The AI industry has moved from pure research into a phase of full-scale industrial engineering. That means experienced professionals can transition in with a real seat at the table, as long as they round out the skill trees below.

The four paths that follow cover more than 90% of high-paying opportunities in the AI era, ordered from highest to lowest technical barrier to entry.

1. Algorithm / Machine Learning Engineer — “Architect of the New World”

Best fit for: Programmers with strong math backgrounds, STEM students with research experience, fresh master’s and PhD graduates from top-tier universities, and those with competitive algorithm experience.

Salary range (Tier-1 cities): ¥500K-1.5M for new grads; ¥800K-3M with 3 years of experience; ¥5M+ for senior experts.

Current state: Pure-theory researchers are oversaturated. What companies hunger for now is engineering talent who can ship. Meta’s headline-grabbing $100M offers to top researchers in 2025 made everyone’s blood race, but those are the apex. The real shortage is in the middle of the market: engineers who understand model fundamentals AND can package those models into stable production systems.

A real story: A Hangzhou-based industrial inspection AI company hired Wang in 2025, an engineer who previously worked on recommendation algorithms at Alibaba. Wang has never published at a top conference, but he can use LoRA fine-tuning to distill a 7B open-source model down to 1.5B and deploy it to edge boxes on factory floors, lifting defect detection accuracy from 89% to 97%. The company offered him ¥1.8M annually plus 0.5% equity.

Skill tree:

• L1 Foundations: Advanced Python (generators, decorators, async), linear algebra (matrix decomposition, eigenvalues, the only language for reading model formulas), probability theory (Bayesian inference, MLE). Recommended resources: 3Blue1Brown’s linear algebra series, the CS229 lectures.
• L2 Models: Deep dive into the Transformer architecture (attention mechanisms, positional encoding, KV cache, the soul of modern AI), mastery of deep learning frameworks (PyTorch fluency is non-negotiable; writing custom operators is a plus), and reproducing classic papers (GPT, BERT, LLaMA).
• L3 Advanced: Fine-tuning, LoRA, QLoRA, DPO, RLHF; model distillation, compressing knowledge from large models into smaller ones; quantization, INT8/INT4 to slim models down. The art of adapting models to vertical domains using minimal compute.
• L4 Deployment: Inference optimization, vLLM, TensorRT-LLM, SGLang. Getting large models to run on phones or edge devices, instead of being chained to expensive H100 server rooms.
• L5 Frontier (bonus): MoE (Mixture of Experts) architectures, multimodal fusion, RL-based alignment, long-context optimization.

Pitfall to avoid: Don’t fall into “paper-reading anxiety.” Getting 5 open-source projects running on GitHub is more valuable than reading 50 arXiv papers.

2. AI Product Manager — “The One Who Makes the Elephant Dance”

Best fit for: Traditional PMs, domain experts with deep industry experience, consultants, and senior operations professionals.

Salary range (Tier-1 cities): ¥350K-800K is common; senior AI PMs can reach ¥1M-2M.

Current state: PMs whose only skills are wireframes and spec docs are dead. Today’s AI PM must understand technical boundaries, what the model can do, what it can’t, and what it can do but isn’t worth doing.

Why is this role so hot? Because at AI startups, the biggest bottleneck isn’t the algorithm, it’s the question “whose problem should my model actually solve?” Someone who can translate vertical-industry know-how into AI task specifications is rarer than ten algorithm engineers combined.

A real story: Lisa, a Beijing-based product manager with 8 years in healthcare SaaS, jumped to a medical AI startup in 2025 and saw her annual salary leap from ¥600K to ¥1.3M. Her core advantage isn’t coding, it’s knowing exactly “at which second a radiologist hits a pain point reading a CT scan, and which subtle details cause misdiagnoses.” That kind of business intuition takes years of immersion in the field. The product she now leads with her algorithm team has been deployed in over 30 top-tier hospitals.

Skill tree:

• L1 Technical literacy: Understand token-based pricing (GPT-4o: $2.5/M input tokens, $10/M output tokens; Claude Opus 4: $15/M input tokens, these numbers must be second nature), context windows (the cap on how much a model can “read” in one shot), the mechanics of hallucination (why AI fabricates), and the trade-off between RAG and fine-tuning.
• L2 Interaction redesign: Master the blend of LUI (conversational interfaces) with traditional GUIs. The interaction is no longer click-click-click but chat-chat-chat. Study how ChatGPT, Perplexity, Claude Artifacts, and Cursor have rewritten the playbook.
• L3 Business modeling: Decompose complex business processes into AI-executable tasks. Learn to draw “agent workflow diagrams”, where to use an LLM, where to use a rule engine, and where humans must remain in the loop.
• L4 Cost and compliance: Learn to do the math. If a customer service bot handles a million queries, how much does each model call cost? Are the margins viable? How do you protect data privacy? How do you comply with the EU AI Act and China’s Interim Measures for Generative AI Services? These are life-or-death questions for the product.
• L5 Evaluation systems: AI products have no absolute standard for “good”, only “the score on a given evaluation set.” Learn to design evals, run A/B tests, and close the loop with user feedback. This is where AI PMs most clearly differentiate themselves from traditional PMs.

3. Data Engineer — “The Master Alchemist of the AI Era”

Best fit for: Database developers, data analysts, backend engineers, and ETL engineers.

Salary range (Tier-1 cities): ¥400K-900K; AI data architects ¥1M+.

Current state: Garbage in, garbage out. The bottleneck in AI today isn’t the algorithm, it’s high-quality data. A line that circulated through Silicon Valley in 2024 says it best: “Whoever owns the data owns the next generation of models.” OpenAI spending a billion dollars to license data, Anthropic hiring thousands of annotators, every one of these moves puts data engineers center stage.

A real story: A Shenzhen fintech company couldn’t find an algorithm engineer capable of building its risk-control model. The project was ultimately rescued by Li, a veteran data engineer with 10 years of ETL experience. In three months, Li unified customer behavior data scattered across 17 systems into a clean, traceable, time-stamped feature store. The model the algorithm team trained on top of this data set lifted AUC by 0.15. Li’s annual salary jumped from ¥450K to ¥950K.

Skill tree:

• L1 Data fundamentals: SQL at a master level, window functions, CTEs, query optimization at your fingertips. Python data processing (Pandas/Polars; Polars runs 5-10x faster than Pandas and is the darling of 2025).
• L2 Distributed systems: Spark, Flink for real-time processing; data lakes (Iceberg, Hudi, Delta Lake); the cloud-native data stack (dbt, Airflow, Dagster).
• L3 AI-specific: Vector databases like Milvus, Pinecone, Qdrant, and Weaviate, the cornerstone of RAG (Retrieval-Augmented Generation). Learn embedding model selection, HNSW index tuning, and hybrid retrieval (vector plus keyword).
• L4 Data governance: Data masking, lineage tracking, automated annotation pipelines (combining LLMs for semi-automated labeling can boost efficiency tenfold), and synthetic data generation. The discipline of producing high-quality corpora to “feed” models cheaply and at scale.
• L5 Frontier (bonus): Knowledge graph construction, multimodal data pipelines (handling text, image, and audio in a unified flow), and Feature Store platforms.

Industry trivia: A high-quality instruction-tuning dataset of 100,000 entries goes for ¥300K-800K on the open market. Anyone who can build a self-running “synthetic data generation + quality evaluation” pipeline is essentially printing money for their company.

4. Prompt Engineer / AI Implementation Specialist — “The Fastest Path to Cashing In”

Best fit for: Operations specialists, copywriters, consultants, HR, sales, talented humanities graduates, and laid-off mid-career professionals.

Salary range (Tier-1 cities): Entry-level ¥200K-400K; senior AI implementation consultants ¥600K-1.5M; independent consulting day rates ¥5,000-20,000.

Current state: The core of this role isn’t “writing prompts”, that’s just surface-level entry. The real essence is business process automation. To put it another way: you’re not chatting with AI; you’re rewriting an entire business pipeline using AI.

Why is this the fastest path to cashing in? Because it doesn’t require you to write code, but it does demand that you understand business, processes, and communication. Those are precisely the moats that mid-career professionals already have.

Real story 1: Wei, a 38-year-old former trade manager, was let go in 2024 and skipped the job hunt entirely. Instead, he spent 4 months mastering Coze and Dify. He now sells “AI private domain + customer follow-up” automation packages to small and mid-sized export business owners across the Pearl River Delta. Each package goes for ¥30K-80K. In 6 months he closed 11 clients, banking nearly ¥600K in cash, double what he made as an employee.

Real story 2: A regional court brought on an “AI implementation consultant”, originally a court clerk. Using Dify, he built a workflow that automatically organizes courtroom audio recordings, generates draft transcripts, and retrieves similar past cases. Clerk productivity went up 60%. He didn’t write a single line of code, and his salary climbed from ¥120K to ¥450K.

Skill tree:

• L1 Prompting techniques: Master Chain-of-Thought (CoT), the ReAct framework, few-shot learning, role-playing, and structured output (JSON Schema constraints). Learn to coach an AI the way you’d coach an intern, give it a clear role, goal, constraints, and examples.
• L2 Workflow orchestration: Become fluent in Dify, Coze, n8n, and Make.com, or low-code tools like LangChain and LlamaIndex. Build workflows where multiple agents collaborate.
• L3 API integration: Know how to use webhooks to make AI send emails, query CRMs, read Feishu docs, populate Feishu Bitables, even modify code. Learn the differences and best practices for calling APIs from OpenAI, Anthropic, Tongyi Qianwen, and Kimi.
• L4 Solution delivery: Build complete cost-saving and efficiency-boosting loops for specific use cases (AI customer service, AI-assisted writing, AI sales lead cleaning, AI résumé screening, AI data analysis reports). Know how to tell the ROI story, what clients care about most is “How much do I invest? How much do I get back? How fast do I break even?”
• L5 Advanced (bonus): MCP (Model Context Protocol), agent evaluation methods, AI security defenses (against prompt injection attacks).

The most common beginner mistake: Obsessing over “the magic prompt.” There is no magic incantation. The real moat is your ability to deconstruct business processes. A good AI implementation consultant spends 70% of their time talking business with clients, 20% building workflows, and only 10% writing prompts.

Who Will Rise Through This Storm? Three Survivor Profiles

1. The “Refuse-to-Age” Career Veteran

Your deep understanding of an industry is something AI can’t replicate. AI is just a new paintbrush. Law, healthcare, finance, manufacturing, education, senior practitioners in these high-barrier industries will see staggering multiplier effects when they pair their expertise with AI.

For example: a partner at a law firm with 15 years of practice, once equipped with AI tools, can have a 3-person team produce the contract review volume of a 15-person team. Client billing stays the same; his marginal profit jumps 5x.

2. The “Don’t-Worship-Tech” Pragmatist

The best AI solution is the one that solves problems and saves the company money. Don’t chase “the latest, coolest model.” Chase “the right tool for a specific problem.” Those who can embed AI into Excel, into customer service systems, into WeChat groups will make money far more easily than those chasing AGI.

3. The Self-Driven Learner

The half-life of information in AI is just 3 months. GPT-4 to GPT-5, Claude 3 to Claude Opus 4, Sora to Veo 3, the pace of iteration is breathtaking. If you can keep iterating yourself (2 hours a week learning new tools, getting one new project running each month, reflecting on your methodology each quarter), you’ll be the era’s favorite child.

Action Checklist: A 4-Week Transition Plan Starting Tonight

Remember: The job market doesn’t test how many AI concepts you understand, it tests what you’ve actually built with AI. One small working project beats 100 certificates.

A Set of Sobering Numbers (Q1 2026)

• AI-related job postings in China are up +178% year-over-year, but applicants have grown +520%, competition is fierce, but good roles are perpetually understaffed.
• The hire ratio for AI algorithm roles is roughly 1:140; for AI product managers and AI implementation consultants, it’s about 1:30, sidestepping the front line is often the easier breakthrough.
• Less than 8% of résumés in China list “hands-on AI project experience”, one working project already puts you ahead of 92% of your peers.
• Among the four role categories, AI Implementation Specialist is the most undervalued track, demand is growing 3.2x faster than supply.

Closing: Choosing Clearly Amid the Flood

The end of every old era is accompanied by a redistribution of wealth.

The Industrial Revolution put hand-loom weavers out of work, but it minted railroad tycoons. The internet revolution gutted traditional newspapers, but it made ByteDance possible. Every “transition phase” of a technological revolution opens a 5-10 year window, those who seize it eat meat, those who miss it sip soup, and those who never see it coming go hungry.

The AI wave of 2026 is squarely in the middle of that window. The wildest early dividends have already been split up by elite algorithm researchers, but the mid-to-late stage dividends, application dividends, vertical-deployment dividends, AI-meets-traditional-industry dividends, are just beginning.

While others lose sleep over layoff notices, you should be at your computer, getting your first agent up and running. While others complain that “AI took my job,” you should be asking yourself “whose job can I take with AI?” While others wait for “the market to recover,” you should understand, this is the new normal. Recovery isn’t coming. The only thing coming is being replaced or being surpassed.

The old ship has sunk. A new shore is in sight. Pick up the oars.