This is the question everyone asks and nobody can answer with one year. So we don’t give one. We give a median year and an 80% range, and we say exactly what “replaced” means.
What “replaced” means here
A job counts as replaceable in a given year only when all of these are true:
- AI can do at least 90% of the job’s weighted task time,
- at least as well as a typical qualified professional, and
- at least half of employers have adopted it.
That is a high bar on purpose. Long before a job meets it, the job changes: fewer hires, different tasks, new tools. The year we give is not when the job starts changing. It is the earliest point at which most of it could be done without a person.
We never say a job “will be eliminated”. A replacement year is an estimate of when it could be, if the trends hold.
How the estimate is made
We run 10,000 simulated futures for each job. Each one draws its own values for the things nobody knows for sure, then works out the year the job would cross the bar. The median of those years is the headline, and the middle 80% of them is the range.
Each simulated future combines five things.
1. Where the job starts
The job’s coverage today, with a little noise to reflect measurement error. Hands-on tasks start low, because coverage already caps them at what robots can do.
2. How fast AI improves
We base the pace of progress on METR‘s research into the length of tasks AI agents can complete, which has been growing steadily. Each simulated future draws its own pace:
- In three futures out of four, AI’s capability on desk work doubles every 4 to 12 months. The hands-on share of the job progresses more slowly, doubling every 12 to 48 months.
- In one future out of four, progress slows down: desk work doubles only every 18 to 48 months, and hands-on work every 48 to 120 months.
The slow scenario is there because trends can stall. Without it, every future assumes today’s pace continues forever.
3. The quality gate
Being able to do a task is not enough; the work has to be good enough. For jobs with grade A or B quality evidence, we project the quality trend forward to find when AI reaches parity. For other jobs, we add a lag of 1 to 4 years after capability arrives. No job has grade A or B evidence yet, so every job currently takes the lag.
4. Friction
Some jobs are slow to change even when the technology is ready. The model has room for friction from:
- Licensing: the work legally needs a licensed person.
- Regulation: sector rules slow adoption.
- Liability: someone has to carry the risk if it goes wrong.
- Client preference: people want a person, as with care work and therapy.
Each of these, plus physical work and gaps in the evidence, is scored for every job and shown on the page under “What’s stopping it”. They do not yet change the timeline. In the current release every job gets the same friction, so a job’s replacement range does not yet respond to its own blockers.
5. Adoption
Even proven technology takes years to spread. We model adoption as an S-curve: slow, then fast, then leveling off. The curve starts from the share of US businesses using AI today, 17% to 20% in the US Census Bureau’s Business Trends and Outlook Survey, plus half of how much the job’s own tasks already show up in AI use.
The typical lag from “AI can do it” to “half of employers use it” is drawn between 5 and 15 years. Jobs that start further along the curve get a shorter lag, and jobs that start further back a longer one (from 0.3 to 2 times as long).
How to read it
You will see something like:
Median 2041 (80% range 2035–2052)
That means half the simulated futures cross the bar before 2041, and 8 in 10 cross it between 2035 and 2052. (Illustrative example, not a real job.)
- A wide range means the evidence is thin or the job depends on things that are hard to predict, such as robotics.
- “Not foreseeable before 2060” means the median falls after 2060. We do not forecast further out, because nobody can do that honestly.
- A decade band, such as “2040s or later”, appears when evidence for the job is thin: quality grade C or D, or fewer than two studies. The year is also pulled halfway toward the median for the job’s family, and we show the band rather than a precise-looking year. No job has quality evidence yet, so every job in the current release shows a band.
Known limits
- It is a model, not a forecast you can bank on. The range is wide because the future is uncertain.
- The ranges are our judgment. The pace, lag and slow-progress ranges come from published research, but we have not yet tested them against what happened to jobs in the past.
- Blockers are not in it yet. Licensing, regulation, liability and client preference are shown on each page but do not yet move the year.
- It assumes no shocks. A breakthrough, a regulation or a recession could move any job.
- Replacement is not the only outcome. Many jobs will shrink or change long before they could be replaced, and some will grow because AI makes the work cheaper and demand rises.
- We check it. We compare our scores with official data and publish the result on the methodology page. A test of the timelines themselves is still to come.
Sources
- METR, research on AI task time horizons. metr.org
- US Census Bureau, Business Trends and Outlook Survey. census.gov
- Anthropic Economic Index, observed AI use and robot exposure data, CC BY.
- Quality evidence as listed on each job page.