There is a job that did not exist three years ago and now sits in the middle of a lot of businesses. It has no settled title. Sometimes it is an operations person, sometimes a marketer, occasionally somebody from support who got curious.
What they do is take AI that already works and make it do something specific for the business. They are not building models. They are figuring out that the quote process has four steps a machine could handle, wiring it up, discovering it breaks on Thursdays, and fixing it.
The hiring data has now caught up with this. Half of US AI job postings sit outside IT departments entirely, and the fastest-growing skills are not about building AI. They are about deploying it into work that already exists.
The shift nobody announced
For most of the last decade, an AI job meant a research or engineering role: someone with a technical background training models, usually inside a technology company or a dedicated data science function.
That is now half the picture at most. Half of AI postings in the US are outside IT, with growth concentrated among people putting AI into marketing, sales, support, and paperwork rather than people building the underlying systems. The demand moved from creating capability to applying it, which is a normal maturity pattern for any technology and is happening unusually fast here.
PwC's 2026 Global AI Jobs Barometer describes the labour market splitting into two distinct paths, with human skills being rewarded rather than displaced. That framing is worth holding lightly, since consultancy reports have a house style, and the underlying observation matches what the posting data shows: the valuable combination is not deep technical skill alone, it is domain knowledge plus enough AI fluency to apply it.
This is a genuinely different conclusion from the one implied by the layoff coverage we examined in the AI layoffs piece. Both are true simultaneously: AI is cited in a great many job cuts while creating substantial demand for people who can deploy it. The destruction and creation are happening in different places, which is why they feel contradictory in headlines.
What is actually growing
Two specific signals in the data are worth knowing because they tell you what the market has decided is valuable.
The first is agentic AI as a named skill. It went from 0.06% of US postings to 0.23% in a year, an increase of roughly 280%, reaching around 90,000 postings. Those are small percentages of a very large market, and the growth rate is what matters: employers have gone from not naming this at all to naming it in ninety thousand adverts, which is how a skill becomes a category.
The second is stranger and more informative. Forward Deployed Engineer postings are up more than 1,000% year over year. The role means someone technical who works directly with customers, embedded in their environment, making the software actually function in a real business rather than in a demo. A thousand percent growth in that role is the clearest possible statement that the industry has discovered AI does not deploy itself.
On compensation, senior AI and ML engineers earn between $240,000 and $520,000 base plus equity, time to hire averages eight to twelve weeks, and around 70% of accepted offers face a counter-offer. Those are numbers a small business cannot compete with and does not need to, which is the point of the next section rather than a reason to despair.
The entry-level problem
There is a genuinely troubling pattern in this data and it deserves naming rather than glossing.
Employment of 22 to 25 year olds in AI-exposed occupations now trails the rest of the market by 19%, widened from a 13% gap a year earlier. The entry-level door is closing faster than the overall market is contracting, which is a different and worse problem than general softness.
Within AI roles specifically it is starker. Only 3% of machine learning engineer postings and 2% of AI product manager postings are open to entry-level candidates. A field growing this quickly with almost no junior entry points is building a structural problem for itself, because the senior people commanding $520,000 today came from junior roles that increasingly do not exist.
The mechanism is not mysterious. The tasks that traditionally trained juniors, meaning the well-defined, supervised, repetitive work that teaches the fundamentals, are precisely the tasks AI now handles. That is exactly the dynamic IBM was pushing against when it tripled entry-level hiring, which we covered in the IBM hiring piece. It remains the exception rather than the trend.
What this means for a small business
The headline numbers describe a market you are not competing in, and the useful reading is about shape rather than salary.
You will not hire an ML engineer and you do not need one. Nothing a small business does with AI requires someone who can train models. Every practical thing you need, meaning connecting tools, designing prompts, handling failure cases, checking output quality, and knowing when a task is a bad fit for automation, sits well below that skill level and well above no skill at all.
What the Forward Deployed Engineer explosion tells you is that the scarce skill is deployment rather than creation, and that is a skill you can plausibly buy or grow. The market has priced model-building out of your reach and left application-level work far more accessible, which is the opposite of what the salary headlines imply.
The entry-level collapse also creates a genuine opportunity for a small business willing to act on it. A capable graduate who cannot get a junior role at a large company is available to you, frequently at a reasonable rate, and they have grown up using these tools. The large companies have decided that training juniors is not worth the investment, and that decision leaves capable people on the market.
Who to actually hire
If you are writing a role with AI in it, the profile that works for a small business is fairly specific and it is not the one most job adverts describe.
Hire for domain knowledge first and AI fluency second. Someone who deeply understands your operations and can learn to use AI tools will outperform someone who knows AI tools and has to learn your business, because the hard part is knowing which of your forty repetitive tasks is worth automating and what happens when it goes wrong. That judgement comes from understanding the work, not from understanding the model.
Test for the right things. The useful signal is not whether someone can describe how a language model works, it is whether they have taken a real process and made it run without them. Ask a candidate to describe something they automated, what broke, and how they found out it broke. The last part matters most, because people who have genuinely shipped automation have all been surprised by it and can tell you the story.
And be honest in the advert about what the role is. A great deal of AI work in a small business is unglamorous: mapping processes, writing clear instructions, checking outputs, fixing edge cases. Describing it as cutting-edge AI attracts people who want to build models and will be disappointed, while describing it accurately attracts people who enjoy making things work, who are the ones you actually want.
The cheaper option most businesses skip
Before hiring anyone, the option with the best return is usually the one already sitting in your business.
Somebody on your team is already the person who figures things out. Every small business has one: the person who fixed the booking system, who built the spreadsheet everyone depends on, who worked out the new invoicing software before anyone else. That person plus a few hours a week of protected time plus a paid AI tool is a genuinely credible alternative to a hire, and it starts on Monday rather than in eight to twelve weeks.
The evidence supports this more than it supports hiring. Gallup data indicates employees who actively use AI tools face lower layoff risk, which we covered in the Gallup piece, and the structural reason is that AI fluency spread across existing roles is more durable than concentrating it in one new hire who might leave. Half of AI postings being outside IT says the same thing from the employer side: the value is in applying AI to a job someone already understands.
The honest catch is that this only works if the time is genuinely protected. Giving someone AI responsibilities on top of a full workload produces a tab left open for three weeks and a quiet conclusion that it did not work. Two protected hours a week beats ten hours of good intentions, and that is the actual decision in front of most small businesses rather than whether to hire.