The raise is where displacement shows up first
The letter arrives in March. It thanks you for a strong year. Then it gives you a number, and the number is smaller than last year's. Nothing in the letter mentions software. Nothing in your title has changed. This is what AI and wage compression look like from inside a job: not a redundancy notice, a smaller figure on a page you read alone. The headcount chart stays flat. The raise budget does not. Displacement, in its first years, is a change in price before it is a change in people.
The best evidence for this is new and contested, and both halves matter. A July 2026 paper from Apollo Global Management, by Sania Edlich and Torsten Slok, measured where AI is actually used, through the Anthropic Economic Index, and matched that to federal wage data for 321 occupations from 2015 to 2025. Workers in high-exposure occupations saw their wages grow about 6.7 percentage points slower than other workers after 2023. The paper found no significant effect on employment. The authors' reading is blunt: firms are capturing the productivity gain through wage compression rather than workforce reduction.
That is one study, with a narrow definition of exposure, and it should be held loosely. But the mechanism it describes is the one a worker feels first. A firm that automates a third of your tasks does not need to fire you. It needs to stop paying for the third it no longer buys from you. The cheapest way to do that is to let your pay stand still while prices move. No one has to announce anything. The spreadsheet of raises does the work.
Advertised pay is rising, and that is not a contradiction
Here is the finding that cuts the other way. Indeed Hiring Lab looked at millions of US job postings that state a salary. In the occupations most exposed to generative AI, advertised pay rose about 46 percent from 2021 to 2026. In the least exposed, it rose about 25 percent. Controlling for the mix of occupations, postings in exposed work carried a 5.7 percent premium after ChatGPT. The author's title is deliberate: AI exposure is not squeezing advertised pay, it is boosting it. An honest essay on wage compression has to sit with that sentence.
It can, because the two studies measure different people. A job posting is the price of a worker a firm still wants to hire. The Apollo sample is the paycheck of workers already inside. These can diverge, and in a repricing they would be expected to. When routine output gets cheap, the firm posts fewer junior roles and bids harder for the senior person who can check the machine. Postings tilt toward seniority. Advertised pay climbs. Meanwhile the incumbent whose tasks were the ones automated gets the smaller letter in March.
Indeed's own breakdown points the same way. The gap in cumulative pay growth between exposed and unexposed occupations was about 17 points for senior roles, 12 for mid-level roles and 2 for entry-level roles. The author calls the seniority split suggestive rather than proven. Take it as that. But a premium that is 17 points at the top and 2 at the bottom is not a rising tide. It is a ladder being pulled up from below.
The entry-level number is where both sides agree
Start with the people trying to get in. The Stanford Digital Economy Lab, working from ADP payroll records, found that employment of 22 to 25 year olds in the most AI-exposed occupations was about 19 percent below its trend as of June 2026. There was no comparable gap for experienced workers. The lab's reading is that adjustment is happening mainly through reduced hiring. That is the headcount side of the story, and it fits Indeed's two-point premium at the entry level: fewer seats, and seats that pay little more than they did.
Now the pay side. A September 2026 Census Bureau working paper followed graduates by major. In the decile of majors most exposed to AI, initial employment fell 5 percentage points after late 2022. Full-quarter initial earnings fell 13 percent. The authors compare this to graduating into a large recession. About half of the earnings fall came from graduates moving into lower-wage sectors. Read that again. The degree did not change. The person did not change. The price the market set for a first year of their work fell by about an eighth.
This is wage compression in its purest form, because it never shows up as a cut. No one's salary was reduced. A cohort simply started lower, and will carry that lower base through every percentage raise that follows. A percentage raise on a smaller base is a smaller raise. Compounded over a decade, the gap between this cohort and the one before it would be the quietest displacement in the data.
AI and wage compression land hardest at the bottom of the pay scale
The Apollo paper's most uncomfortable table is the one by wage quartile. In the bottom quartile of pay, workers in exposed occupations saw real wage growth about 10.7 percent lower than their unexposed peers after 2023. The second quartile lost less, the third less again. The top quartile showed no statistically significant effect at all. The authors suggest high earners are better placed to absorb or benefit from the tools. The table says something simpler. The repricing lands hardest on the people with the least room to lose it.
An October 2025 working paper for the Washington Center for Equitable Growth helps explain why. Chiara Chanoi and Chris Bangert-Drowns split AI exposure by how the tool is used. Where use is augmentative, the worker checking, iterating and learning alongside the model, a percent more exposure predicts 2.5 percent higher hourly wages. Where use is automative, the model completing the task on its own, it predicts 2.3 percent lower wages. Across all exposure the net effect is a small positive 0.5 percent. The average hides the split. The split is the whole story.
So the question for your own pay is not whether your occupation is exposed. Most knowledge work is. The question is which of your tasks the firm now buys from a model and which it still buys from you. A [bookkeeper](/roles/bookkeeper/) whose reconciliation is automated and whose exception handling is not sits on both sides of that line at once. Our role pages score that split task by task, as an illustrative map rather than a measurement. The map matters because the raise follows the split, not the title.
A flat raise budget is a repricing, not a bad year
Managers rarely say the word compression. They say the budget was tight. The Atlanta Fed's Wage Growth Tracker gives you a yardstick to test that. In January 2026 the median job stayer saw hourly wages grow 3.5 percent over a year. The median job switcher saw 4.7 percent. The overall figure was 3.6 percent, down from 3.7 the month before. If your raise sits below the stayer figure in a year your team shipped more than ever, the budget is not the whole story. Your tasks were repriced.
The tell is in what the review praises. Read the performance review next to the raise letter. If the review praises your speed, your volume, your throughput with the new tools, and the raise does not move, you have been told what happened. The firm counted your extra output and assigned it to the software. The productivity went to the margin. The raise went to the senior person who approves your work. This is the division Indeed's seniority gap describes, seen from one desk.
It could still be a bad year. Firms have bad years. The way to tell is to look at the ladder, not the letter. Did the firm promote fewer people from your level this cycle? Did the junior posting under you close and not reopen? Was the starting band lowered for the next hire? One of these is noise. All three together is a structure. In the scenario where adoption keeps deepening, that structure would spread from exposed occupations to adjacent ones before any headcount chart changed.
The headcount chart is the last chart to move
Wage data move before employment data for a plain reason. Firing is expensive and visible. It triggers legal notices, severance, press, and the morale cost of the people who stay. Letting a raise lag prices triggers nothing. Not hiring a junior triggers nothing. Lowering a starting band triggers nothing. Each is a decision a finance team can make inside a spreadsheet, in one afternoon, with no announcement. The Apollo finding of no significant employment effect is exactly what that path would look like from the outside.
The scale of the quiet path is already measurable, with caveats. Apollo counts 5.8 million US workers in high-exposure occupations under its threshold, about 3.7 percent of the labor force, and puts the labor income shortfall at $28 billion a year as a lower bound. Those are the authors' figures for a narrow definition of exposure, and they say the figure would grow as adoption deepens. Whether it does is a forecast. What is not a forecast is that this sum appears in no layoff statistic. It appears in 5.8 million smaller letters.
So the worker waiting for a layoff announcement to confirm the threat is waiting for the wrong signal. By the time headcount moves, the compression has run for years and set the base from which every future raise is computed. The signal that arrives first is the one most people dismiss as personal: a smaller raise, an unexplained band, a junior role that closed. They are not personal. They are the price of your task stack, set by someone who never has to say so.
This week: price your tasks before your employer does
The move is to separate your pay from your title and attach it to tasks. Take your last two performance reviews and the raise letter that followed each. List the tasks the reviews praised. Mark which of those tasks the firm now does with a model, in whole or in part. Then mark which ones still need you to check, decide, escalate or sign. The first list is where your raise was quietly reassigned. The second list is what you should be paid for. If you cannot name the second list, that is the finding, and it is the one to act on.
Then do one of three things, depending on what you found. If your second list is long, write it in the firm's own words and bring it to the next pay conversation as a task list, not a plea. If it is short, choose one adjacent task on the augmentative side, the review, the exception, the handoff, and start doing it before anyone asks. If the junior posting under you has closed, write that down too. The ladder below you is where your own leverage used to come from.
The point is not to win one negotiation. It is to see the repricing before it compounds. Advertised pay for exposed work is rising. Pay for the tasks a model already does is not. Which side of that line you stand on is, for now, partly a choice. [Test my task stack](/#task-stack-test) and find out which tasks are still yours.
