Far less wholesale replacement than the coverage suggests, and considerably more quiet contraction than most people notice. Announced layoffs attributed to AI make news. What is doing most of the work is a vacancy that never gets posted, and that leaves no trace anyone reports.
Reading the employment data rather than the announcements produces a much clearer picture of where jobs are actually going.
Key points
- Replacement is rare; non-replacement is common. Roles disappear when someone leaves and nobody is hired.
- Announcements overstate AI’s role. Automation is a more acceptable explanation for a layoff than overhiring was.
- The overall labour market is growing, by about 5.9 million positions to 2035.
- Clerical work is contracting sharply, with tellers down 13 percent and bookkeeping clerks down 6 percent.
- Entry level absorbs most of the impact, because that is where the automatable tasks sat.
How the replacement actually happens
The mechanism matters, because it determines what an individual should watch for.
The vacancy that never appears
This is the dominant pattern. Someone resigns, retires or moves internally. The manager considers whether the position needs refilling, concludes the remaining team can absorb it with better tooling, and the role is quietly removed from the plan. Nobody was replaced. A job simply stopped existing.
There is no announcement because nothing happened that requires one. Aggregated across an economy this accounts for far more reduction than every publicised layoff combined, and it is invisible in exactly the way that makes it hard to plan around.
The absorbed task rather than the absorbed job
More common still. A role loses forty percent of its content to automation and keeps existing, now containing harder work. The person is not replaced; their job got denser. Over time the team needs fewer people to handle the same volume, and that shows up as reduced hiring rather than as departures.
The genuine restructure
Real and less frequent than coverage implies. An organisation redesigns a function around automation and reduces headcount deliberately. This happens most in contact centres, back-office processing and content operations, which is precisely where the task shape suits it.
The convenient attribution
Worth naming because it distorts the public picture. A company that overhired and needs to correct has a choice about how to explain it. Attributing reductions to AI transformation sounds forward-looking. Admitting to poor planning does not. Some proportion of announced AI-driven cuts are ordinary corrections wearing a better label.
This is not a conspiracy theory; it is an observation about incentives. Nobody is required to explain a layoff accurately, the market frequently rewards the technology framing, and there is no mechanism that checks. Anyone building a picture of the labour market from press releases is reading a document written for a different purpose.
What the data says about where jobs are going
The occupational projections are more informative than any announcement, because they cover the whole economy rather than the companies that issue press releases.
Clerical work is contracting
Tellers are forecast to decline 13 percent through 2035, bookkeeping and accounting clerks 6 percent, and information clerks 2 percent. These are large occupations and the declines represent substantial numbers.
Technology roles split rather than shrink
Computer programmers decline 7 percent while software developers grow 10 percent. Support specialists fall 3 percent while information security analysts grow 21 percent. Same sector, opposite directions, divided by whether the work is execution or judgement.
Hands-on work is growing
HVAC technicians grow 11 percent, medical assistants 13 percent, logisticians 18 percent. None of these is being replaced, and several are constrained by recruitment rather than demand.
The total is rising
Employment is projected to increase from 170.3 million to 176.2 million between 2025 and 2035. The composition changes considerably; the total does not fall.
On measurement, the Bureau publishes AI exposure categories for 831 occupations and states plainly that exposure “does not imply job loss, productivity gains, automation probability, or wage effects.” Reports treating exposure scores as predictions of unemployment are misreading the source.
What replacement costs the organisation
The public conversation treats headcount reduction as a straightforward saving, and the accounting is considerably messier than that.
The training pipeline
Junior roles were never justified on their immediate output; they were an investment in having mid-level people later. Removing them saves money now and creates a shortage in about five years, at which point the same capability has to be bought on the open market at a considerably higher cost. Several industries have run this cycle before with apprenticeships, and the pattern repeats reliably enough to predict.
The institutional knowledge
The person who leaves and is not replaced took an understanding of why things are the way they are. Documentation captures a fraction of it. Organisations discover the gap during an incident, which is the worst moment to discover anything.
The residual work
Automating eighty percent of a function leaves twenty percent that is disproportionately difficult, because the easy cases were removed. Staffing models that reduce headcount proportionally leave the remaining people handling only hard problems with no recovery time, and the burnout that follows is a cost that appears under a different heading.
The reversal risk
Where a function was cut too far, rebuilding is more expensive than maintaining would have been. Recruitment, training and the productivity gap during rehiring all land at once, and typically the people who understood the previous arrangement are gone.
None of this means reduction is always wrong. It means the saving is smaller than it looks on the day it is announced, and the difference arrives later under headings nobody connects to the original decision.
What to know before deciding
Several things matter for anyone assessing their own position.
Announcements are a poor signal
They tell you which companies want to be seen as transforming, not where employment is actually moving. The occupational data is slower and considerably more reliable.
Watch your own organisation’s vacancies
The most informative indicator available to an individual is whether departures in your team are being refilled. That is visible months before anything is announced and it is rarely hidden.
The cost of a mistake is what protects a role
Where an error is expensive and someone must answer for it, organisations keep a person. Where output is cheap to check and cheap to correct, they do not. This predicts better than task difficulty.
Contraction is slow enough to be missed
A 6 percent decline over a decade does not arrive as an event. It arrives as a colleague not being replaced, which is easy to sit through for years without making a decision.
Some of this will partially reverse
Organisations that stopped hiring juniors are creating a mid-level shortage roughly five years out, and rebuilding costs more than training would have. Several sectors have been through this cycle with apprenticeships before, and the recovery phase is when the firms that kept training find themselves unusually well positioned.
Sector matters less than task shape
Two people in the same industry can face opposite outlooks depending on whether their day is specified and repetitive or ambiguous and consequential. Reasoning from the industry, which is how most coverage frames this, produces bad predictions in both directions.
Which functions are actually being restructured
Where deliberate redesign is happening, it clusters in a small number of places, and the reasons generalise.
Contact centres
The clearest case. High volume, categorised historical data, verifiable outcomes, and a large share of contacts that follow a recognisable pattern. Deflection to self-service began well before generative AI and has accelerated. What remains is the difficult end, which is a harder job than the mixed workload it replaced.
Back-office processing
Invoice handling, claims intake, document capture and data entry. All three properties that make automation succeed are present, and the work has been shrinking for two decades rather than two years.
Content operations
Producing volume to a brief. Copy variants, resizing, translation, standard reporting. This compressed sharply and it employed a lot of people at the entry level, which is part of why the effect on juniors has been so visible.
Where restructuring has not worked
Attempts to redesign functions requiring physical presence, regulated sign-off or unstructured diagnosis have generally underdelivered. Organisations that cut support staff too aggressively have frequently rehired, because deflection metrics improved while resolution times got worse and the cost surfaced somewhere else on the balance sheet.
That pattern is worth holding onto. The functions where restructuring stuck share the three properties. The ones where it was reversed did not, and the reversal took twelve to eighteen months to become visible.
Decision framework
Five questions for assessing your own exposure.
- Are departures in your team being refilled? This single observation beats any national statistic for predicting your position.
- What share of your work is specified and repetitive? That is the automatable portion, and it can be measured in a week.
- Who answers when your output is wrong? If it is you, by name, the role has structure behind it.
- Is your occupation growing or contracting? Check the specific category rather than reasoning from the industry.
- Could you move toward the exceptions? In almost every declining role there is adjacent work that is growing, and it usually recruits internally.
Understanding where these systems genuinely help and where their output needs checking is itself the skill that moves people toward the growing side. If you want a structured route in, explore Coursiv AI lessons and check current plan details on the official site.
Your next step
Count how many people have left your team in the last two years and how many were replaced. That ratio is the clearest signal you have about your own organisation, it costs nothing to establish, and it is usually knowable from memory.
Then take the most repetitive part of your own role and work out who would notice if it disappeared. If the answer is nobody, that is the part to move away from. If the answer is that an expensive mistake would follow, that is the part worth being known for.