Companies are paying for people who can make AI work inside their product, not just demo it. There are few of those people, because the field is new, and companies feel they are losing time. Scarcity plus urgency. That is most of the answer.
I won't put a number in this note. Packages vary too much by city, company and person, and a figure in an article becomes a promise in the reader's head. What I can do is explain why the market behaves the way it does, with sources you can check.
What companies are actually paying for
A demo takes an afternoon. You call a model, show it ten good examples, and everyone in the room is impressed. Then it meets real users and real data, and it starts to fail in ways nobody planned for. My first AI feature that fell flat was a policy chatbot that answered from the old version of a policy, because the old document matched first. The demo had looked fine.
The skill companies are short of is the one that sits between those two moments: finding where the feature fails, and building the system around the model so it fails less. That is slow, unglamorous work, and not many people have done it yet.
The hiring data points the same way. Lightcast, the labour-market firm behind the job-postings data in Stanford's AI Index, summed up the 2026 report this way: hiring has moved away from experimentation and toward execution. Postings increasingly ask for skills in building and managing systems at scale.
Demand is growing fast
Three public sources, all looking at job postings:
- In the U.S., 2.5% of all job postings now mention AI skills, up 55% on the year before.
- In the same data, postings asking for agentic AI skills went from 0.06% of postings in 2024 to 0.23% in 2025. The demand is moving from chatbots to systems that take actions.
- AI engineer is first on LinkedIn's 2026 list of the fastest-growing jobs in the U.S. The most common skills it lists for the role are LangChain, RAG and PyTorch.
Most of this data is from the U.S., where job postings are measured best. I haven't found an Indian report I would link with the same confidence, so read these numbers as the direction of the market, not its level where you live.
Supply is small because the field is new
In 2023, Shawn Wang (swyx) wrote the essay that gave the job its name, The Rise of the AI Engineer. His argument was about supply. A small number of research labs hold the people who can build the models. Everyone else rents those models through an API. So a new kind of engineer is needed in between: a software engineer who knows how to make those models useful in a product. He estimated about 5,000 LLM researchers in the world against about 50 million software engineers, and predicted it would become the most in-demand engineering role of this decade.
Three years on, that in-between group is still forming. Most of the people I know who can do this work learned it on a real product, because courses were behind the work. That is the scarcity part.
The urgency part is simpler. Companies feel they are losing time. Waiting a year to train their own people feels like falling behind, so they pay more to hire someone who has already done it.
Will it last?
It will settle. Two things will pull it down:
- More engineers will learn this. Every developer who ships one AI feature joins the supply.
- The tools will get easier. Things that took real skill in 2023, like getting a model to return valid JSON every time, are now a setting in the API.
The premium for "can call an API" will go. What stays valuable is what stays hard: production experience, and knowledge of the domain the product serves. Someone who knows how legal documents, hospital records or insurance claims really work, and can make an AI system behave inside that domain, is hard to replace with a better tool.
The data already shows this split. Lightcast's October 2026 update found that only 1.4% of junior postings in the U.S. mention AI skills, against 9.3% of postings that ask for five or more years of experience. Companies are looking for people who have already shipped. And the Stanford AI Index 2026 reports that employment for U.S. software developers aged 22 to 25 has fallen nearly 20% since 2024.
I put that last number here on purpose. A high average package hides a hard entry. The market is paying for experience, and it is not handing that experience out.
What this means if you are deciding whether to switch
Don't switch for the number. Switch if you would enjoy the work.
The premium is for people who get good at it, and you only get good at work you would do anyway. Most of an AI engineer's day is not spent calling models. It is spent figuring out where a feature fails: reading outputs one by one, finding the input that broke it, deciding whether the prompt had enough context to answer at all. If that sounds like a puzzle you would chase on a Sunday, you will probably get good at it, and getting good is what the premium is for. If it sounds tiring, the number won't make it less tiring, and the number is the part most likely to change.
You can find out cheaply which one it is for you.
This week
Take one small AI feature, your own or a tutorial, and spend one evening trying to break it. Feed it the messiest real input you can find. Write down every way it fails and why.
If you finish the evening wanting to fix them, you have your answer. That evening is a normal day in this job.
The number follows the skill. The skill follows the work you would do anyway.