# The Occupation Automation-Exposure Score, Read Before You Retrain

*You will score your occupation's automation exposure from public task data and reach one defensible call: hold, reskill in place, or retrain.*

- Canonical URL: https://www.refolk.ai/candidates/guides/occupation-automation-exposure-score
- Pillar: Reading the market
- Format: Framework
- Published: 2026-08-31
- Last reviewed: 2026-08-31
- Reading time: 16 min

You are deciding whether your occupation is exposed enough to automation that you should retrain into a different one, or whether you can stay and adapt in place. This guide is for a working person - entry-level or mid-career - who wants to convert their own job's task-level exposure into one call, using only public data. It gives you scored dimensions tied to real task datasets and a stated decision boundary, so a scary number becomes a defensible move.

Most pages that rank for "is my job at risk of automation" hand you a single risk percentage and no action rule. That number is built by assigning whole occupations to a probability bin, which hides the one thing you can actually change: your own task mix. This document does the opposite. It decomposes your job into tasks, scores each, aggregates to one figure, and tells you what the figure means.

## What "automation exposure" actually measures, and what it does not

Automation exposure is the time-weighted share of your tasks where a tool cuts the work time by at least half without losing quality. It is a measure of technical feasibility, not a forecast of job loss.

This distinction is the whole game. The Eloundou et al. "GPTs are GPTs" study defines exposure at the task level: a task is exposed if access to an LLM tool reduces the time to complete it by at least 50% without reducing quality. Their scores measure the technical feasibility of an LLM circa early 2023, decomposed into discrete tasks. Feasibility is not the same as displacement, and treating them as one is the most common way readers get this wrong.

Three public sources give you the raw material, each measuring a different thing:

| Source | What it scores | What each field measures |
|---|---|---|
| O*NET | The task taxonomy | Over 19,000 occupation-specific task statements linked to more than 2,000 detailed work activities |
| AIOE (Felten, Raj, Seamans) | Occupation-AI relatedness | Crowd-sourced task-AI application relatedness, aggregated to a score per 6-digit SOC code |
| Eloundou et al. | LLM exposure per activity | Whether an LLM cuts task time by at least 50% without quality loss, scored per detailed work activity |

The Bureau of Labor Statistics includes a technological-change variable inside its projections, but it does not release that field publicly, so a standalone "Degree of Automation" score is not established publicly. Build your work on O*NET, AIOE, and Eloundou instead.

**19% - Share of workers who may see at least half their tasks impacted by LLMs**

Eloundou et al. estimate about 80% of the workforce could have at least 10% of tasks affected, but only around 19% cross the 50% line.

## The dimensions that move the call

Score four dimensions, in order of how much they change your decision: substitution versus augmentation, codifiability, bottleneck protection, and the entry-level hiring trend. Exposure level alone is the weakest of these, which is why a raw percentage misleads.

The task-based literature splits work on two axes. Autor, Levy and Murnane partitioned tasks along routine versus non-routine and cognitive versus manual, and argued that automation substitutes for the routine while complementing the non-routine. Machines and software have taken over codifiable and repetitive tasks. That gives you the first two dimensions.

Frey and Osborne named three bottlenecks that protect work: occupations requiring complex perception or manipulation, occupations requiring creativity, and occupations requiring social intelligence. Computers are significantly challenged in these three areas. That is dimension three. LLMs shifted the frontier toward standardized text work, so the most exposed jobs are heavily reliant on the standardized processing of textual information.

The dimension that actually decides hold versus retrain is substitution, not exposure level. Identical exposure scores diverge on outcomes: declines are concentrated where AI substitutes for human tasks, while where it complements workers, employment is flat or rising. So when you score a task as exposed, immediately ask a second question - does the tool do the task instead of you, or does it do it faster with you? That second answer carries more weight than the first.

#### The four scoring dimensions, ordered by decision weight

1. **Exposure level** - Time-weighted share of tasks an LLM can cut by 50% or more
2. **Codifiability** - Whether the knowledge is written and rule-based or tacit and situational
3. **Bottleneck protection** - Share of tasks shielded by perception, creativity, or social intelligence
4. **Substitution vs augmentation** - Whether the tool replaces you on the task or speeds you up

*Exposure level is the outermost, weakest layer; substitution at the core is what moves the hold-or-retrain call.*

Codifiability is the specific trigger inside "white collar." Declines hit occupations that involve codified knowledge, while tacit-knowledge occupations see faster employment growth for experienced workers. This is why paralegal-type text-standardization roles score high and tacit trades do not. When you tag a task, ask whether the knowledge it uses is written down and rule-following, or whether it lives in judgement built from experience.

## What counts as "high exposure": the thresholds

High exposure begins where more than half your time-weighted tasks are exposed. That is the 50%-of-tasks line, and it is the decision boundary this guide is built on. Three published models set thresholds differently, and knowing which is which stops you from importing the wrong number.

| Model | Unit scored | High-exposure threshold | Headline share |
|---|---|---|---|
| Eloundou (LLM, direct) | O*NET activities, 50% time-saving | 50% or more of tasks | ~19% of workers |
| Eloundou (LLM plus software) | O*NET activities | 50% or more of tasks | ~46% of jobs |
| Frey & Osborne | Whole occupation | Over 70% probability | 47% of employment |

The load-bearing number is the Eloundou 50%-of-tasks line. It splits the workforce cleanly: around 19% of workers may see at least 50% of their tasks impacted by LLMs directly. Add complementary software and the share of jobs with over half their tasks affected rises to just over 46%. The Frey and Osborne 47% figure is a different animal - a 70% probability applied to whole occupations - and it is the one that overstates risk, because it does not allow for changes in the tasks within an occupation.

> **Rule:** Use the 50%-of-tasks line, not a risk percentage
>
> Your decision boundary is the time-weighted share of your own tasks that pass the 50% time-saving test. If that share is at or above 50%, you are a retrain candidate. If it is below, reskill in place is defensible.

There is convergent evidence the task method is sound. Mean task-exposure scores correlate with the AIOE index at a Spearman rho of 0.84, and an independent GPT-4o pass estimated roughly 6,000 of about 19,200 O*NET tasks - close to one-third - could be done without human intervention. Different methods, same direction.

## The leading employment signal you should not skip

The earliest measurable sign that an occupation is contracting shows up in entry-level hiring, not in aggregate employment. Watch the youngest cohort in your occupation, because adjustment runs through hiring rather than layoffs.

The Stanford Digital Economy Lab and ADP "Canaries in the Coal Mine" series uses monthly payroll records for millions of workers. Early-career workers aged 22 to 25 in AI-exposed occupations experienced a 16% relative employment decline while employment for experienced workers stayed stable. The gap has since widened to about 19% below where it would be if it had kept pace with less-exposed peers. In levels, 22-25 employment in the two most exposed quintiles fell about 11% from late 2022 to mid 2026, while the three least-exposed quintiles grew about 10%.

Two features make this a clean signal. First, it operates primarily through reduced hiring of young workers rather than increased separations. Second, declines concentrate in occupations where AI substitutes for human tasks; where it complements, employment is flat or rising. That is the same substitution dimension, showing up in payroll data.

> A mid-career worker should read the entry-level trend in their own occupation as their personal early warning.

One honest caveat on timing. A February reanalysis found that with the broadest controls, declines in AI-exposed occupations become statistically significant only in 2024; earlier declines are likely driven by non-AI factors such as interest rates and the tech-sector cycle. So when you read the trend, confirm it persists after those controls before you treat it as an AI signal.

## Score your own occupation, step by step

Here is the procedure. It takes roughly two to three hours end to end. No single published tool outputs one defensible number with a hold, reskill, or retrain rule, but every component below is public, and doing it by hand is what makes the number yours rather than the occupation average.

#### From job title to one exposure figure and a call

1. **Find your O*NET code** - Enter your job title into O*NET OnLine and read the 6-digit O*NET-SOC code from the page header. Skim the task list to confirm it fits the work you actually do.
2. **Pull tasks and DWAs** - Open the Details or Custom report for your code and export it to a CSV of task statements and their linked detailed work activities.
3. **Weight tasks by time** - Assign each task a rough share of your working week so the shares sum to 100%. This is the step public scores skip, and where within-occupation honesty lives.
4. **Tag routine vs bottleneck** - Mark each task routine-codifiable, or protected by perception and manipulation, creative intelligence, or social intelligence.
5. **Apply the 50% test** - For each task, ask whether an LLM tool cuts the time by at least 50% without reducing quality. Flag each exposed or not exposed.
6. **Aggregate to one figure** - Sum the time-weighted share of tasks flagged exposed to get a single percentage you can place against the 50% threshold.
7. **Cross-check an index** - Look up your SOC code in the AIOE dataset and the Eloundou occupation score, and confirm your DIY figure points the same direction.
8. **Read the entry-level signal** - Check the Canaries Dashboard trend for exposed occupations in your age band to see whether hiring is already contracting.

A note on ordering. Some people run the employment-signal step first, driven by a headline. Resist that. The task decomposition should precede the employment signal, so your own number - not the news - drives the call. The signal confirms or complicates the number; it does not replace it.

When you tag exposed tasks, add a second column beside the exposed flag: automate or augment. This is the single most valuable annotation in the whole exercise, because it is the dimension that predicts whether high exposure means fewer jobs or faster work.

**Per-task scoring row**

```
Task statement | Time share % | Routine or bottleneck | 50% test (exposed/not) | Automate or augment
Draft standard contract clauses | 20 | routine-codifiable | exposed | augment
Interview client to establish facts | 15 | social intelligence | not exposed | -
Cross-check filings against a database | 25 | routine-codifiable | exposed | automate
Advise client on a judgement call | 10 | social + creative | not exposed | -
```

*One row per task. Time share must sum to 100% across all rows. The automate-or-augment column is what turns exposure into a decision.*

## Turning the score into hold, reskill, or retrain

The call is a two-variable judgement: your exposed task share against the 50% line, crossed with whether your exposed tasks are mostly substituted or mostly augmented. A high share of augmented tasks is not a retrain signal; a moderate share of substituted tasks in a contracting occupation is.

#### The hold, reskill, or retrain map

Horizontal axis runs from Mostly augmented to Mostly substituted. Vertical axis runs from Exposed share below 50% to Exposed share at or above 50%.

| Quadrant | What it means |
| --- | --- |
| High exposure, augmented | Reskill in place: add the tool, re-weight toward bottleneck tasks |
| High exposure, substituted | Retrain candidate: confirm the entry-level trend, then plan a move |
| Low exposure, augmented | Hold: adopt the tool opportunistically, re-score yearly |
| Low exposure, substituted | Hold with watch: small substituted share, monitor the trend |

*Exposure share sets the vertical position; substitution versus augmentation sets the horizontal, and the horizontal usually wins.*

Read the map like this. If your exposed share is below 50%, hold is defensible regardless of the substitution split - you are adapting, not retreating. If your share is at or above 50% and the exposed tasks are mostly augmentation, reskill in place: the tool makes you faster, so learn it and re-weight your week toward the tasks the bottlenecks protect. Only the high-share, mostly-substituted quadrant is a genuine retrain candidate, and even there you confirm the entry-level hiring trend before committing.

There is real room to lower your own score legitimately. Off-the-shelf models overstate risk by freezing task mix; a reader who re-weights toward bottleneck-protected tasks can defensibly reduce their exposure. That re-weighting is the "adapt in place" move made concrete.

The reskill-in-place path is less crowded than it looks. In Refolk's index the "Prompt Engineering" pool is only 12.2% the size of the "Machine Learning" pool - roughly one-eighth. The adaptation skill is far from saturated, so a worker who adds it stands out rather than joining a crowd.

| Skill listed | US profiles | Share vs Machine Learning |
|---|---|---|
| Machine Learning | 481,968 | 100% |
| Prompt Engineering | 58,696 | 12.2% |

If you land in the retrain quadrant and want to see what the move actually looks like, study people who have already made it. Refolk searches its index of professional profiles by the exact pivot you are considering, so you can read real paths from a text-heavy role into a lower-exposure one rather than guessing.

Ask me this: `People who reskilled from a text-heavy role into a perception/manipulation or social-intelligence occupation (nursing, skilled trades, therapy)` - [run the search](https://www.refolk.ai/start?q=People%20who%20reskilled%20from%20a%20text-heavy%20role%20into%20a%20perception%2Fmanipulation%20or%20social-intelligence%20occupation%20(nursing%2C%20skilled%20trades%2C%20therapy)).

*Returns profiles that made the retrain move this guide describes, so you can see the starting roles, timelines, and destination occupations.*

Once you have a target occupation, [Refolk](/candidates) can write the resume for it from your own history and score how well you actually fit each posting, which is the friction that stops most in-progress pivots from ever getting applications out the door.

## How this goes wrong: the failure modes

The score fails in predictable ways, and every one of them inflates or deflates your call in a direction you can check. Read this section as carefully as the procedure; a standard that overclaims is worse than no standard.

- **Scoring the occupation, not your job.** Two workers with the same SOC do different work, and O*NET-level scores hide it. Research shows considerable variation in tasks among workers within the same occupation. Check by re-running the score on your own time-weighted task list, not the occupation average.
- **Treating exposure as displacement.** Exposure means time-saving, not job loss; the Eloundou score is technical feasibility circa early 2023. Check by separating "task exposed" from "occupation shrinking" and never collapsing the two.
- **Substitute versus complement confusion.** A high score can mean AI helps you, not replaces you, because declines hit codified-knowledge work while tacit work grows. Check by classifying each exposed task as automate or augment.
- **Reading a raw percentage as a verdict.** A single fear number assigns whole occupations to a probability bin and ignores within-occupation task change. Check by using the task-share method, not the occupation label.
- **Ignoring the entry-level signal for mid-career comfort.** Aggregate stability can mask a hiring freeze at the bottom; young workers in exposed occupations are falling behind while totals look calm. Check by reading the age-banded Canaries trend, not the headline.
- **Timing false positive.** Early declines may not be AI. With the broadest controls, declines become significant only in 2024, and earlier ones likely reflect non-AI factors. Check by confirming the trend persists after interest-rate and tech-sector controls.
- **Manual or social work treated as safe forever.** Bottlenecks erode. Kiva Systems solved warehouse navigation by putting barcode stickers on the floor, sidestepping a perception problem. Check by re-scoring annually rather than once.

> **Watch out:** A single risk percentage is not a decision
>
> The widely cited 47% figure assigns entire occupations to a greater-than-70% probability bin and ignores that you can re-weight your own tasks. Arntz and colleagues showed this overstates the share susceptible. Never let one occupation-level number stand in for your own task decomposition.

> **Tip:** Score the automate-or-augment column twice
>
> The first pass tends to be optimistic on your own tasks and pessimistic on others'. Re-run the automate-or-augment tag a day later with the question flipped - "what would a manager cut?" - and reconcile the two.

## Keeping the score current

An exposure score is a snapshot, not a settlement. Re-run it on a schedule, because the frontier moves and the bottlenecks erode. The mechanism to watch is not this year's model release but whether the tasks you rely on for protection are still protected.

Set a yearly cadence. Each pass, redo the time-weighting first, because your own task mix shifts as you adapt, and that shift is the point. Then re-apply the 50% test to the tasks you moved toward - the bottleneck-protected ones - and confirm they still fail the test. If a perception or social task that used to be safe now passes, that is your signal that a bottleneck is eroding in your specific work, well before any index catches up.

Between full re-scores, keep one leading indicator on a shorter loop: the entry-level hiring trend for your occupation. It moves before aggregate employment and before your own experience does. If the 22-25 cohort in your occupation is contracting after controls, treat it as your early warning even if your personal role feels stable.

#### Before you call the score done

- [ ] You scored your own time-weighted task list, not the occupation average
- [ ] Every task carries a routine-or-bottleneck tag and an exposed-or-not flag
- [ ] Each exposed task is labelled automate or augment
- [ ] Your aggregate exposed share is placed against the 50% line
- [ ] Your DIY figure agrees in direction with the AIOE and Eloundou scores, or the discrepancy is flagged
- [ ] You read the age-banded entry-level trend for your occupation, after controls
- [ ] You landed on one call: hold, reskill in place, or retrain
- [ ] You set a calendar reminder to re-score in twelve months

Do the decomposition first, let the number lead, and let the employment signal confirm or complicate it. That order is what separates a defensible move from a fear-driven one - and it is why, when someone asks how you decided to hold, reskill, or retrain, you will have a task list and a threshold to point at rather than a headline.

## Frequently asked questions

### Is my job at risk of automation, or just exposed to it?

Exposure and risk are different things. The Eloundou score measures technical feasibility circa early 2023 - whether an LLM saves at least 50% of the time on a task without losing quality. That is not job loss. To move from exposure to actual risk, classify each exposed task as automate or augment, and check the entry-level hiring trend for your occupation. High exposure where AI complements you often means faster growth, not displacement.

### Should I retrain because of AI if my score is above 50%?

A weighted task share above 50% puts you in the retrain-candidate band, but it is a trigger to investigate, not a verdict. First separate the tasks AI substitutes from the ones it augments. If most of your exposed share is augmentation, reskill in place by adding the tool to your workflow. Retrain only when a large substituted share coincides with a contracting entry-level hiring trend in your occupation.

### How exposed is my occupation to automation compared to others?

Look up your 6-digit SOC code in the public AIOE dataset, which scores occupations on a scale running from -2.67 for Dancers to 1.58 for Genetic Counselors. Then cross-check the Eloundou occupation score. Neither replaces your own time-weighted task decomposition, because within-occupation task variation is real and two workers with the same code often do different work.

### Why not just use a published automation risk percentage?

Because whole-occupation scores freeze your task mix. The widely cited 47% figure assigns entire occupations to a single high-risk probability bin and ignores that you can re-weight your week toward bottleneck-protected tasks. Arntz and colleagues showed this overstates the share susceptible. Score your own weighted task list instead, and the number becomes a move rather than a fear.

### How do I future proof my career if I decide to adapt in place?

Re-weight your working week toward tasks protected by the three bottlenecks - complex perception and manipulation, creative intelligence, and social intelligence - and add the adaptation skill. In Refolk's index the Prompt Engineering pool is only 12.2% the size of the Machine Learning pool, so the skill is far from saturated and a worker who adds it stands out. Re-score annually, because bottlenecks erode.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/candidates/guides/occupation-automation-exposure-score*