ISSUE 003
The AI wage premium is 23, 28, 56 or 62 percent depending who you ask
Every few weeks someone sends me a headline about what AI skills are worth. The numbers are never the same and nobody ever says why. So I went and lined them up. The gap between them turns out to be more useful than any single one of them, particularly if you are the person who would be writing the paycheck rather than cashing it.
An idea worth thinking about
The most requested AI skill is knowing how to use ChatGPT
Lightcast catalogues over 300 distinct AI skills across job postings — machine learning, AI ethics, autonomous driving, robotics, natural language processing. Then they looked at which ones employers actually ask for most. The answer is proficiency with ChatGPT or Microsoft Copilot.
The distribution is just as telling. By 2024, 51 percent of postings requiring AI skills sat outside IT and computer science, and demand for generative AI skills in non-technical roles rose by roughly 800 percent after ChatGPT's launch in late 2022. Lightcast found demand higher in non-tech sectors than in the technical ones AI is usually associated with, and named customer and client support, sales, and manufacturing and production among the fields carrying the largest premiums. Different work needs different skills — transportation roles ask for autonomous driving, maintenance roles for robotics.
Which means the thing being paid for, most of the time, is not expertise. It's ordinary competence with tools anyone can open today. That is either encouraging or unnerving depending on where you sit, and for a small business it mostly means the capability you're trying to hire is one you could plausibly build in someone you already employ.
Read the skills outlook
A tool I disappeared into
What a job actually pays in southeast Missouri, from the federal survey
The Occupational Employment and Wage Statistics program publishes annual employment and wage estimates for over 800 occupations, broken out nationally, by state, and by metropolitan and nonmetropolitan area. Southeast Missouri exists in it as its own nonmetropolitan geography, which is rarer than it should be in federal data. In Missouri the survey is run by MERIC in cooperation with BLS, and MERIC publishes a dashboard on top of it.
Download the XLS rather than reading the web tables. The spreadsheet carries the 10th, 25th, 75th and 90th percentile wages alongside the mean, and the percentiles are the whole point — a mean tells you almost nothing useful about what you'd need to offer, while the 75th percentile tells you what it costs to be competitive.
For scale, BLS reported the Springfield metro area averaging $25.96 an hour in May 2024 against a national $32.66. Read that next to any national salary premium percentage and the percentage starts meaning something different.
Open the wage dashboard
Learn something
How a wage number is actually made, from the people who make one
BLS publishes a Handbook of Methods explaining how each of its statistics is constructed, and the OEWS chapter is the antidote to every wage headline you'll read this year. It's free, it's the primary account, and an hour with it changes how you read all of them.
Three things in it that immediately matter. The estimates exclude self-employed workers entirely — so in a region thick with sole proprietors and contractors, the survey is describing payroll jobs only. Annual wages for hourly occupations are calculated by multiplying the hourly mean by 2,080 hours, a year-round full-time assumption that quietly inflates the annual figure for any job that isn't. And every estimate ships with a relative standard error, which is the published admission of how uncertain it is.
None of that makes the data bad. It makes it legible. A number you know the construction of is worth more than a larger number you don't, which is more or less the argument of this entire entry.
Read the methods handbook
What we’re seeing
Four credible studies, four different AI wage premiums, and at least one article crediting the wrong one to the wrong firm
Lightcast, analysing over a billion job postings, found that listings requiring at least one AI skill advertised salaries 28 percent higher on average — roughly $18,000 more a year — rising to 43 percent for listings asking for two or more. The Oxford Internet Institute, studying more than ten million UK vacancies, found 23 percent, a premium they noted exceeded the one for a master's degree and approached a PhD. PwC's Global AI Jobs Barometer has been reported in the mid-to-high fifties, with a 2026 reading some sources put at 62 percent, up from 57 the year before.
They disagree because they measure different things. Lightcast reads what employers advertise in postings; Oxford read a different national labour market under different conditions; PwC compares roles rather than postings. Advertised salary is not paid salary. A UK vacancy is not a US one. None of these is wrong and none of them substitutes for another.
Then there's the failure mode that made me want to write this down. I found an article headlined as a Lightcast report announcing a 56 percent premium — which is PwC's figure, not Lightcast's. Two firms, two methods, two numbers, welded into one sentence that reads as authoritative and is simply wrong. That is how a statistic becomes untethered, and it happens within a few months of publication.
The turn worth making is whose question this is. Every one of these studies answers "what can a worker earn." If you run a business here, the operative question is what you would have to pay, and that answer is local, not national. Which is why the useful response to a 56 percent headline isn't alarm — it's opening the wage file for your own area and reading the 75th percentile for the role you actually need.
Lightcast figures as reported by CNBC
One question
What would you have to pay to hire someone good with these tools?
A number, for your county and your kind of work. Reply with it — I want to see how far the local reality sits from the national headline.
More in this pillar as it's written.
— Shana
Sources link to the original publisher, never to a summary.