AI Is Taking the Raise Before It Takes the Job
Workers in high-AI-exposure occupations have seen real wage growth run 6.7 percentage points slower since 2023 than workers in low-exposure roles, and Apollo Global Management found no matching job losses to go with it.
Key Takeways
- Apollo Global Management found real wage growth in high-AI-exposure occupations running 6.7 percentage points slower since 2023 than in low-exposure occupations, with no matching job losses.
- Gartner projects that by 2029, 30% of employees laid off due to replacement by AI will need to be rehired, often at significantly higher cost.
- Cutting an experienced team takes a quarter, while rebuilding the judgment that team carried takes years, which puts the two halves of an AI workforce plan on different clocks.
- As competitors rent increasingly similar AI, accumulated human judgment is the hardest advantage to copy and the only one that leaves the moment it is cut.
Workers in the occupations most exposed to AI have seen real wage growth run 6.7 percentage points slower since 2023 than workers in low-exposure roles. So far, AI is costing people raises more than it is costing them jobs. Apollo Global Management's chief economist, Torsten Slok, and researcher Sania Edlich found that number across roughly 300 occupations, and CNBC put it back in front of a national audience on September 13, 2026. The AI layoffs everyone has been bracing for did not show up in the data alongside it. The paycheck moved first.
That is a stranger finding than a layoff wave, and much harder to see. A layoff has a date and a name attached to it. A raise that comes in 2 points light has neither. It arrives in a 10-minute conversation in March, it gets explained as a tight budget year, and the person walks back to their desk and keeps working.
What 6.7 points looks like at the kitchen table
Behind the number is someone who did the work, kept the title, and opened a compensation letter that read about the same as last year's. He kept the job. He lost the raise. Nobody told him AI was the reason, because in most companies nobody decided it in those words. The money for the AI program came from somewhere, and merit pools are the softest place in the budget to find it.
The question I hear most after a keynote comes from someone in exactly that position. Still employed, and doing the math on a career with fewer rungs in it. That is the real shape of the will-AI-take-my-job question in 2026, and the data answers it in the compensation line before it answers it in the org chart.
The rehiring bill comes due in 2029
On September 9, 2026, Gartner published "Four Shifts Shaping the Future of Work." One projection in it: by 2029, 30% of employees laid off due to replacement by AI will need to be rehired, often at significantly higher cost. That is a forecast about competence. In roughly 3 out of 10 cases, the work turns out to need the person after all, and the company buys them back at a markup.
Cutting takes a quarter. Rebuilding that kind of judgment takes a decade. Same line on the headcount plan, two different clocks, which is most of what leaders get wrong when layoffs and workflow redesign arrive in the same memo.
Deferred retraining is a liability that compounds, and it sits on the company whether or not anyone writes it down. Call it the learning debt. The Gartner projection is that debt coming due with interest, on a published schedule. Circuit City let go of 3,400 of its highest-paid, most experienced salespeople in March 2007 to cut costs, and Best Buy hired a number of them. Circuit City filed for Chapter 11 in November 2008. The expertise did not evaporate. It changed employers.
Tori Paulman, a VP Analyst at Gartner, framed the choice this way:
"The competitive advantage will go to the CIOs and business executives who build an AI-shaped organization where AI value compounds by reshaping roles."
Reshaping a role and removing it are different operations with different price tags, and only one of them keeps the people who can tell when the model is wrong. That matters more each quarter, because the capability distance between the AI leaders and the fast followers keeps narrowing, and Stanford's AI Index has been tracking that convergence. If the model your competitor rents performs close to the model you rent, the rented part cancels out of the competition. What is left is whatever each company actually owns. Accumulated judgment is the hardest of those to copy, and the only one that walks out the door the day you cut it.
Two questions worth putting on the agenda this quarter
The first is one I ask myself before any headcount decision, and I have gotten it wrong before. What would it cost, in dollars and in months, to buy back the judgment we are about to remove? Most companies price the cut to the decimal and price the rebuild not at all. Running both numbers side by side changes some decisions and leaves others exactly where they were, which is the whole point of running them.
The second is about pay, and it is uncomfortable. If your AI-exposed roles are getting flat merit increases while your AI budget grows, your people have already noticed. A Writer and Workplace Intelligence survey in July 2026 found 29% of knowledge workers admit to working against their own employer's AI strategy, rising to 44% among Gen Z, and 30% of those who did named fear of losing their job as the reason. They are guarding against the wrong threat, and they are guarding for a rational reason. Samsung's workforce did the same math more directly when workers rejected a payout over how AI-driven profits were being split.
The same story at two scales
A merit pool is a company decision. Several million merit pools are an economy. Wage growth is how a household absorbs tuition and rent, and a decade of AI-exposed work running 6.7 points behind reaches well past payroll. It shows up in what a family can save, and in what a 19-year-old decides to study.
If you are not the person setting that budget, you still get a move. In your next check-in, ask your manager which parts of your work an agent is expected to be doing in 18 months. The answer tells you what to relearn, and it is a better use of the conversation than asking for reassurance.
The steadying part is that none of this is settled. Slok's own read is that AI has also made the economy more inventive, with new business formation at record highs, which sits oddly next to flat wages and says the distribution is still being decided by people who are making the decisions right now.
So the question for your next leadership meeting is narrow and answerable. If we cut this team in Q1, what is the written plan for buying that judgment back in 2029, and what does it cost? Bring the compensation numbers for your AI-exposed roles into the same meeting and read them next to the AI budget. If you want to talk through how that conversation has gone with other leadership teams, I am easy to find.
Which jobs are considered high exposure to AI
Apollo's analysis scored roughly 300 occupations by how much of the actual task content current AI systems can perform. High-exposure work concentrates in roles built mostly on text, code, and analysis, including writing, software development, research, and customer support. Low-exposure work concentrates in jobs that require physical presence or hands-on care. Exposure is a measure of overlap with what the tools can do, not a prediction that the role disappears.
Why would a company rehire someone it replaced with AI
Because the part of the job that was easy to automate and the part that was load-bearing were rarely the same part. A model can produce the output; evaluating whether the output is correct in a specific business context takes someone who has seen the work go wrong before. Gartner's September 2026 projection puts that correction at 30% of AI-driven layoffs by 2029, and the rehire typically costs more because the company is now buying scarce judgment back on the open market, often as contract labor.
What should I ask my manager if my job is exposed to AI
Ask which specific parts of your current work the company expects an agent to handle within 18 months, and what it wants you doing with the time that frees up. That question is answerable, it is not a request for reassurance, and the answer gives you a concrete list of what to relearn. A follow-up worth asking in the same conversation: what budget exists for that retraining, and who owns it.
Here is what makes Alex a credible voice on this topic: over 20 years at Cisco he shaped a $1.1B innovation portfolio, which meant deciding, budget cycle after budget cycle, which capability to buy and which to build inside the workforce, the exact tradeoff sitting underneath every AI headcount decision being made this quarter.
.svg)
