Still needs a human is our headline score for each job, from 0 to 100. Higher is safer. It answers the question “Will AI replace this job?” in a single number and a single word.
It is always shown with the three answers it is built from, never on its own:
What goes into it
The score combines two things:
- How much of the job AI can do (coverage, 0–100).
- How well AI does it compared with a qualified person (quality parity, 0–100, where 50 is parity).
A job where AI can do a lot, and do it well, scores low. A job where AI can do little, or does it badly, scores high.
Quality term = min(100, 1.43 × max(0, Quality parity − 30))
Score = 100 − (0.6 × Coverage + 0.4 × Quality term)
The quality term only starts counting once AI’s quality reaches 30 on the parity scale, because below that, AI’s work is not good enough to replace anyone. From 30 upward it is stretched onto 0–100, so quality at the top of the parity scale counts in full.
So 100 means AI can do nothing in the job, and 0 means AI can do all of it, better than a person. Every score in between is reachable.
Adoption is not in the headline score. How fast businesses take up AI affects when a job could be replaced, so it lives in the replacement year. The headline score is about the work itself, and the replacement range is shown beside it, not folded into it.
The verdict words
| Score | Will AI replace it? |
|---|---|
| 80–100 | Nah. |
| 60–79 | A little. |
| 40–59 | Partly. |
| 20–39 | Mostly. |
| 0–19 | Largely. |
The verdict is read from the score as shown, rounded to a whole number, so a score of 79.6 shows as 80, and the answer to “Will AI replace it?” is “Nah.”
“Nah.” is the answer for the top band only. It means the job still rests on people, by a wide margin, on today’s evidence.
The bands are fixed numbers. We do not grade on a curve, so a band can be empty if no job’s evidence puts it there.
Why no job scores below about 50 yet
In the current release (2026-Q4, 923 jobs), scores run from 51.4 to 88.6, with a median of 75.4:
| Will AI replace it? | Jobs |
|---|---|
| Nah. (80–100) | 325 |
| A little. (60–79) | 562 |
| Partly. (40–59) | 36 |
| Mostly. (20–39) | 0 |
| Largely. (0–19) | 0 |
When the bottom two bands are empty, as in the October 2026 release, it is because of the evidence, not the formula. Three things hold scores up:
- No job has quality evidence yet. Every job takes the neutral parity value of 50 (below), which takes the same 11.4 points off every score. So today every job scores 88.6 − 0.6 × coverage, and none can score above 88.6.
- Observed use counts for a lot. How much people already use AI chat assistants for a task makes up 40% of each task’s coverage score by design, and more while no benchmark results are in (see the coverage method). Usage is low even on tasks AI could do end to end, so it holds coverage down. The highest coverage of any job is 62.
- Even the top rating keeps a person in the loop. A task AI can do end to end scores 90, not 100, because review remains.
A job would reach the second-lowest band (Mostly.) only if AI could handle more than about 82% of its task time at parity, or if evidence showed AI’s work beating people’s. The bottom band (Largely.) needs that evidence: even at the top of the quality scale, a job needs coverage above about 68 to get there. No job meets either test today. The jobs most exposed on our data, such as telemarketers and customer service representatives, are in the middle band: AI could partly do them (Partly.).
The neutral value also decides where the line between the top two bands (A little. and Nah.) falls: at a coverage of about 15. Jobs close to that line can move across it when quality evidence for their family arrives.
The figures on the page
The figures on each job page fill with amber to the job’s Still needs a human score, and the rest is grey. A job scoring 72 is 72% amber. The page’s background tint follows the same number.
On the page, the task list keeps its own split: “AI does this now”, “AI helps” and “Still needs a human”. That split is about task time. The headline score also weighs quality and keeps room for review, so the two will not always match. A job can have almost no task time marked “Still needs a human” and still score above 50.
Jobs with no quality evidence
Most jobs have no direct test of AI against people. For those, the quality term uses an imputed value: the job family’s average once any job in the family has usable evidence, and a neutral default of 50 (parity) until then. The page still shows “Not yet measured” for quality, and the dataset flags every imputed value.
Our evidence register has no entries yet, so in this release every job uses the neutral default. It lowers every score by the same amount, so it shifts where the bands fall but does not change the order of jobs.
Why we publish one number at all
A single number is easy to misread, and we thought hard about leaving it out. We publish it because people asking “will AI replace my job?” want a straight answer first. We keep it honest by:
- always showing the three answers beside it,
- always showing the evidence grade and the replacement range,
- publishing the formula and every change to it, and
- keeping the bands fixed.
Known limits
- It blends two different things (how much and how well) into one number, which hides detail. The job page shows the detail.
- Today it is coverage on a fixed scale. Quality is imputed for every job, so ranking jobs by this score is the same as ranking them by coverage. Quality evidence is what will separate jobs with similar coverage.
- It measures software AI. Robot data only caps what AI is credited with on hands-on tasks, so robots never lower a job’s score, and factory automation does not show in it.
- It says nothing about pay or demand. A job can score low and still be growing.
Our own score
The footer of every page shows our own job’s score: web developers, SOC 15-1254. It currently reads Still needs a human: 55/100 ↑ safer. We are scored by the same rules as everyone else.