Job-market study · Captured 22 July 2026
“AI Engineer” has become a real job title with a real shape, and it is not the same job as machine learning engineer. So I read the postings: 68 application-layer AI engineering roles across 48 companies, from frontier labs to SaaS scale-ups, retrieved and read in full on one day. Research and model-training roles were excluded on purpose. This is what the job asks for — and who the ideal candidate has become.
Scope
“AI Engineer” and “machine learning engineer” are different jobs with different skills. Mixing them produces the confused career advice this study exists to replace.
Roles whose core is building products on top of models: agents and tool calling, retrieval and vector search, prompt and context engineering, evaluation and guardrails, and the backend and frontend around all of it.
Research roles, roles whose core is training or fine-tuning foundation models, classical ML and data science, training infrastructure, and MLOps-only positions — regardless of what the title said.
192 postings were attempted and 124 were rejected on scope — including several whose title said “AI Engineer” but whose work was research or model training.
Findings
Each number is counted from the 68 postings read.
62/68
ask for agents or tool calling
Almost every posting in this study asks for agents, tool calling, or multi-step orchestration. Two years ago this was an experimental line item; it is now the central description of the work. If you build one thing before applying, build an agent that calls real tools, holds state, and recovers from its own mistakes.
67/68
expect backend ownership
Practically every posting expects you to own services, APIs, data access, and the failure paths around them. The model is one dependency in a distributed system, and the system is what you are hired to keep standing. Frontend engineers underestimate how much of this role is ordinary server-side engineering.
51/68
ask for evaluation of model output
Three quarters of these postings ask for evaluation of model output — golden datasets, scorers, regression gates, observability. It remains the least glamorous skill in the field and the most reliable way to stand out, because most candidates still arrive without any of it. This is the single highest-return thing on the list.
21/68
name TypeScript
Python leads, as it always does, but TypeScript is the clear second and appears far more often than Java, Go, or anything else. This is the strongest signal in the study for anyone arriving from web development: your language is in the room, and increasingly the product surface of these systems is written in it.
8/68
mention fine-tuning at all
Fine-tuning appears in a small minority of postings, usually as one option among many rather than the work itself. It is the most over-studied topic relative to demand in the whole field. Understand when it is the right call — which is rarely — and spend the time you save on retrieval and evaluation instead.
24/68
include frontend work
Frontend is not the core of the job, but it appears often enough to matter — and when it does, it is usually the difference between a prototype and something people will use. AI features are judged in their loading, streaming, partial, and failure states, which is exactly the craft a product engineer already owns.
39/68
are customer-facing
Forward-deployed and customer-embedded engineering has become a dominant pattern in this role: you sit with the people whose workflow you are automating, and you ship against what you learn there. Communication is a stated requirement, not a soft bonus, and it is often what separates two technically equal candidates.
25/68
name LangGraph — the most of any tool
The most frequently named framework in the entire study is LangGraph, ahead of LangChain itself. That is the market saying the interesting problem moved from chaining calls to managing state, control flow, and recovery across long-running work. MCP appears next, as the standard way agents reach real systems.
12/68
name a coding agent as expected tooling
A meaningful share of postings name coding agents — Claude Code, Cursor, Codex, Copilot — as expected tooling rather than curiosities. Employers are hiring for demonstrated fluency with agentic development, which means how you work is now part of what you are assessed on, not just what you know.
45/68
never mention a degree
The majority say nothing about formal education. Where a degree does appear, it is usually paired with an explicit equivalent-experience clause. Combined with published salary bands well into six figures, the picture is a field that pays for demonstrable delivery and barely asks where you learned it.
The stack
Note how flat this list is. No vector database dominates, which is the clearest signal in the study that you should learn retrieval as a data problem and stay portable rather than betting on a vendor.
The ideal candidate
Read together, these postings describe one person with striking consistency. Not a researcher, and not a machine learning engineer — a senior product engineer who has learned to build reliably on top of models.
Most postings ask for five or more years of general software engineering. The AI part is treated as a capability layered on top of an engineer who can already ship, own services, and be trusted with production.
Python for the AI ecosystem, evaluation tooling, and data work; TypeScript for the product surface and increasingly for the agent runtime itself. The strongest candidates move between both without ceremony.
Not a demo. Something with tools, state, error recovery, and a story about what went wrong in production and how it was fixed. This is the portfolio piece these postings are implicitly asking for.
Builds evaluation harnesses, sets a quality bar with the product owner, and can prove a change improved things. This is the clearest dividing line between candidates in the current market.
Can sit with the people whose work is being automated, hear what is actually broken, and turn it into a shipped change. More than half these roles are explicitly customer-facing.
Knows what a request costs, where the latency goes, and which model is the cheapest one that clears the quality bar. Nobody senior is impressed by a feature that works and is unaffordable.
Fluency with Claude Code, Cursor, or Codex now appears in the requirements themselves. The skill being tested is judgement about what to delegate, not enthusiasm.
What it pays
$182K – $290K
Median published band across the 41 postings that disclosed a range in US dollars. Median rather than average because the published ranges include geo-tiered bands and salary-versus-equity choices that run past a million — an average would describe none of these jobs.
What to do about it
It is named in more postings than anything else and it is the cheapest gap to close for someone who already programs. Get to the point where you can read and modify Python comfortably, then move on — nobody is hiring for your third language.
Tools, state, error recovery, and a deployment. This single project answers most of what these postings ask about, and it is worth more in an interview than any certificate.
A golden dataset from real usage, a scorer, a regression gate in CI, and an honest writeup of what your system still gets wrong. Three quarters of postings ask for this and almost no applicants bring it.
Chunking, metadata, permissions, hybrid search, reranking. Vector database choice is the least important decision — the corpus shows no dominant vendor, so learn the concepts and stay portable.
LangGraph is the most-named tool here for a reason: state machines, retries, interruption, and human-in-the-loop are what production agents need. Build the loop yourself once before adopting a framework.
Employers now name these tools in requirements. Be ready to say what you delegate, what you never delegate, and how you review what comes back.
It is the most over-studied topic relative to how rarely it is asked for. Learn enough to know when it is the wrong answer — which is most of the time — and spend those weeks on evaluation instead.
Most of these roles are customer-facing. Being able to describe a trade-off between cost, latency, and quality to someone who does not write code is a stated requirement in this market.
The path through all of it
Python, retrieval, agents, orchestration, evaluation, and the backend around them — in order, with one thing to build in each phase, and free resources for every step. The path was written from the same evidence as this study.
Open the learning path →Method
I searched company careers pages for AI engineering roles, opened every posting I could, and read the requirements in full. Each one was then judged against a written scope rule, and anything whose core was research, model training, classical machine learning, or infrastructure was rejected — including postings whose title said 'AI Engineer'. A separate screening pass re-checked every survivor against the same rule, defaulting to exclusion when the answer was unclear.
Accenture · Amazon Web Services (AWS) · Apple · Baseten · Braintrust · Claude · Cohere · Cursor (Anysphere) · Databricks · Datadog · Decagon · Distyl AI · EPAM Systems · Figma · Future · Glean · Google · Google DeepMind · Harvey · Hebbia · Instrumentl · JUST ADD AI GmbH · LangChain · Machinify · Material Bank · Microsoft · Mistral AI · Nearform · Netflix · Notion · OpenAI · Perplexity · Personio · Ramp · Remote People · Ro · Salesforce · Scale AI · Sierra · Snowflake · Stripe · Tebra · Tenex · Thoughtworks · Unstructured · Vercel · Zuma · deeploi
FAQ
A machine learning engineer trains, tunes, and serves models. An AI engineer builds products on top of models someone else trained: agents, retrieval, tool integrations, evaluation, and the services and interfaces around them. This study deliberately excludes the first job. The skills barely overlap in practice, and conflating them is why so much career advice for this role is wrong.
No. These postings ask for software engineering experience plus demonstrated work with LLM applications. What they want is someone who can ship reliable systems and has learned how models behave — not someone who can derive backpropagation.
TypeScript is the second most-named language here and appears far more than Java or Go, so it is genuinely valuable rather than tolerated. But Python appears in far more postings, and the evaluation and data tooling assumes it. Treat Python as required and TypeScript as your advantage.
An agent that calls real tools, deployed, with an evaluation harness measuring its quality and an honest writeup of its failures. That one project speaks to agents, backend, evaluation, and judgement at once — which is most of what these postings are asking about.
LangGraph is the most-named tool in the study, with LangChain and MCP close behind — but learn the underlying loop before any framework, because the abstractions change faster than the concepts. Vector databases are the clearest case: no single vendor dominates the postings, so learn retrieval as a data problem and stay portable.
The capture date is at the top of the page and the extracted data sits in the repository, so anyone can see exactly what the numbers are based on. Every figure is computed from that data rather than typed in, so when it is re-run the numbers and the date move together.
That gap is where I work — the same agent reliability, retrieval, and evaluation problems these postings are hiring for. Describe what is blocked and I will tell you whether I can help.