The AI-Exposure Read, Scored to Hold, Specialize, or Pivot

This guide is for anyone deciding where to aim a job search while headlines insist their occupation is being automated away. The job is specific: score the role you currently search for, decide whether it is shrinking or merely being reshaped, and convert that into one of three actions - hold your target, specialize within it, or pivot to an adjacent role. You will do this with five named public datasets and your own judgement, not a single doom score lifted off a ranked list.

Most pages on this topic hand you a leaderboard of doomed occupations and a headline risk number. That tells you nothing you can act on. This document gives you a repeatable read you apply to the one role in front of you, and it keeps the two things that get conflated apart: capability exposure (what a model can do) and displacement (what employers actually do).

Why exposure is not the same as displacement

Capability exposure measures how much of your job a model could do. Displacement measures how many people actually lose the work. They move on different clocks, and confusing them is the single most expensive mistake in this whole exercise.

The exposure research is blunt about its own limits. GPTs-are-GPTs scores each O*NET occupation on how much its task content overlaps LLM capability, and the authors describe it as a proxy for potential economic impact "without distinguishing between labor-augmenting or labor-displacing effects." In that paper, about 80% of the US workforce could have at least 10% of tasks affected and around 19% at least 50%. Those are capability ceilings, not pink slips.

Realized data tells a quieter story. Studies using the Current Population Survey show at most small changes in hiring in AI-exposed jobs. Anthropic's first Economic Index, built from roughly a million real conversations, found a slight lean toward augmentation: 57% of tasks augmented, 43% automated. So the capability is broad, but the labor-market outcome so far is narrow and uneven.

The practical rule: capability exposure sets the ceiling on what could change. Your BLS projection and your local postings tell you what is changing. Score both, and weight the realized signals higher for any decision you make this year.

The five signals and what each one proves

The read combines five signals, each from a named public source, each measuring a different thing. One signal on its own lies; the five together are hard to fool.

A note on why knowledge workers cannot skip this. Earlier automation waves hit clerical and manual work. This one runs up the wage ladder. Because LLMs do writing, coding, and analysis, occupations with higher wages generally present higher exposure, contrary to earlier machine-learning evaluations. If you are an analyst, a writer, or a developer, you are squarely in scope.

What drives a high score inside the capability datasets is worth knowing so you can read your own tasks. In the LLM rubric, roles reliant on science and critical-thinking skills correlate negatively with exposure, while programming and writing skills correlate positively. Frey and Osborne identify three bottlenecks that protect roles: perception and manipulation, creativity, and social intelligence. Routine information processing scores high; judgment, physical presence, and human-trust content score low.

What the realized numbers say, by role type

Realized employment data is the spine of this read because it lags capability and splits by cohort. The headline aggregate is calm; the detail is not.

BLS projects the economy adding 6.7 million jobs from 2023 to 2033, total employment up 4.0% to 174.6 million. Growth clusters in care and advisory work; decline clusters in routine office, sales, and processing roles. The AI case studies name specific declines.

Read that table as a pattern, not a list. The shrinking roles are routine information processing. The growing roles carry client trust (advisors) or physical presence (aides) - precisely the bottlenecks that protect work. Home health and personal care aides alone are projected to add 820,500 jobs by 2033.

The employer-intent layer agrees in direction. WEF's survey of 1,000 companies across 22 industries and 55 economies projects 170 million jobs created and 92 million displaced by 2030, a net gain of 78 million, with 86% of businesses expecting transformation and nearly 40% of job skills expected to change.

Now the part the aggregates hide. Stanford's realized-outcome work found that early-career workers aged 22 to 25 in the most AI-exposed occupations experienced a 13% relative decline in employment even after controlling for firm-level shocks, a figure a later revision pushed to 16% and then a widening gap of 19%. In the same data, employment among 22-to-25-year-olds fell 3.8% while it rose 2% for workers aged 35 to 40. The field looked stable in aggregate and was anything but at the entry rung.

Displacement is showing up as a missing bottom rung, not as mass layoffs across the field.

Specialize-within-role is a demand-backed strategy

When a role is being reshaped rather than erased, the right move is usually to specialize upward toward the work machines cannot yet codify. This is not motivational advice; it is visible in posting data.

AI absorbs the easily automated work first - retrieval, summarization, scheduling, formatting, routine drafting. Senior staff hold expertise that is harder to codify, so they stay. That shows up in demand: senior-level postings surged by almost 15% between consecutive years, anchored by experienced workers in tech, engineering, and professional services, while entry-level postings declined 7.5% year-over-year. In one striking cut, software development senior-level positions accounted for 69.3% of postings.

So "specialize" has a concrete meaning. It means moving your target toward judgment, client trust, and senior or oversight tasks, and away from the routine core a model does cheaply. For a customer service representative, that is the team lead who rolls out and supervises conversational AI rather than the agent it displaces. For a credit analyst, it is complex-claims or advisory work rather than standard scoring.

The crowding is real. In my index of professional profiles, there are 420,327 entry-level US software engineer records and 437,839 US customer service representative records. These are the exact tiers where hiring is tightening. A Senior filter on the same software titles returned no matched records in that cut, which I treat as not resolvable in this snapshot rather than as zero supply - the point stands that the visible supply sits at the exposed bottom.

When you have scored your role as reshaped and decided to specialize, the next problem is finding the people who already made that move so you can see the titles and skills they carry. That is a targeted people search, and it is what Refolk is for.

Run the posting check yourself

The one signal you can refresh on any given day is the Indeed Job Postings Index. It is public, daily, and free on GitHub and FRED, with per-sector series, so you never have to trust a stale number.

The index is the percentage change in seasonally adjusted postings since 1 February 2020. A reading of 101 means postings are 1% higher than that baseline. Read three things: your sector's level, its year-over-year change, and the entry-versus-senior split for your title.

The spread is the lesson. Civil engineering at 154.0 and media and communications at 64.1 are the same economy on the same day. A national headline - the overall JPI reached 103.5, its highest since late March and about 3% above the pre-pandemic level - tells you nothing about a cold sector. Always read your specific sector and geography.

Score your role in eight steps

This is the procedure. Budget about two hours the first time. Each step produces one artifact you carry into the next.

On weighting, the sources genuinely disagree, so say so. Capability papers imply exposure leads. Stanford and Indeed show realized demand lags and splits by cohort. For a decision you are making this year, weight the realized signals - your BLS projection and your local postings - above the capability score. The capability score tells you where the pressure will eventually land; the realized signals tell you whether it has landed yet.

Separate a shrinking title from a reshaped one

This is the judgement the whole read exists to produce. A shrinking title and a reshaped one can show the same aggregate headcount and the same exposure score, and they demand opposite actions.

A shrinking title loses headcount across all cohorts, and postings fall at every seniority. There is no safe rung to climb to. A reshaped title holds or grows at senior level while the entry rung erodes, and its task mix skews toward augmentation. The ladder is being pulled up, not knocked down.

The test is the cohort and seniority split. Highly AI-exposed positions contracted only 0.2% year-over-year in one cut, but within that, 22-to-25-year-olds fell 3.8% while 35-to-40-year-olds rose 2%. That divergence is the fingerprint of a reshaped field. If you see it in your role, specializing upward is viable. If headcount is falling for everyone, it is not.

A worked read: customer service representative. Capability exposure is high. BLS projects -5.0% through 2033. Employer intent is declining. The task mix is automation-heavy at the agent level but augmentation-heavy at the lead level. Local postings show entry collapsing. That is a reshaped title: pivot out of the agent role or specialize upward into the lead who deploys and supervises conversational AI.

How this read goes wrong

The read fails in predictable ways, and every failure is a false positive - a conclusion that feels certain and is not. This is the most important section to internalize.

• Reading exposure as destiny. High capability exposure is potential, not outcome; the authors refuse to call it displacement. The false positive is panicking over a GPTs score while your BLS projection and sector JPI are both positive. Check: never act on step 1 without steps 2 and 5.
• Aggregate numbers hiding the cohort split. Exposed positions contracted 0.2% in aggregate while 22-to-25-year-olds fell 3.8% and 35-to-40-year-olds rose 2%. The false positive is "my field is stable." Check: always read your own age and seniority cohort separately.
• Mistaking a hiring freeze for a safe field. The entry-level decline is driven not by layoffs but by a collapse in hiring. Incumbents feel secure while new entrants cannot get in. Check: compare the new-postings trend, not just headcount.
• Treating augmentation as permanent. The augment share flips toward automation as models improve - directive conversations rose from 27% in late 2024 to 39% by August 2025. The false positive is "AI only assists my role." Check: re-score your automation share every six months.
• Task-count exposure overstating risk. Anthropic notes data entry keyers and radiologists show higher effective exposure than raw task counts suggest, while teachers and software developers appear less affected once reliability and duration are considered. Check: weight by task success and time, not task count.
• Confusing a national series with your local market. JPI sectors diverge from 154.0 to 64.1. The false positive is applying the national index to a cold sector. Check: use your specific sector and geography.
• Over-trusting Frey-Osborne-style scores. Critics argue the hand-labeling ignores social-perceptiveness activities, inflating risk. Check: cross-read capability scores against O*NET bottleneck variables before concluding.

Convert the read into a search decision

The output of this whole read is one of three actions, and each maps to a different search. Do not leave the exercise with a feeling; leave it with a target.

For a pivot, aim at an adjacent title that shares your transferable skills but scores lower on exposure - an insurance claims professional moving into fraud investigation or complex claims, for instance. For a specialize-upward move, aim at the senior or oversight version of your own role and rebuild your materials around judgment, trust, and supervision rather than the routine core. For a hold, keep your materials current and re-run the read every six months, because the augment share drifts.

Once you have the decision, the materials have to match the new target precisely, which is slow to do by hand across a batch of postings. Refolk writes a resume from your own history and tailors it to each posting, which is exactly the repeated work a pivot or a specialize-upward move generates.

Role (O*NET-SOC): __________
Capability exposure rank: __________ (driving tasks: __________)
BLS 2023-33 projection: ____%  (grow / shrink)
WEF intent: grow / decline / absent
Automation share of my core tasks: ____%  (dated: __________)
Sector JPI level / YoY / entry-vs-senior split: ____ / ____ / ____
Shrinking or reshaped: __________
Decision: hold / specialize / pivot
Re-score on (date six months out): __________

Keep the read current

Before you act, verify you actually ran the full read and did not stop at the scary number. Run this checklist against the role in front of you.

The read has a half-life. The augment share moves - directive automation climbed from 27% to 39% in under a year - so a "specialize" verdict can become a "pivot" verdict without any headline event. Re-score every six months using the one-line record, and refresh the Indeed JPI whenever you are about to commit to a batch of applications, since that is the one signal that updates daily and for free. The occupation you score today is not the occupation you will search for next year, and the point of this framework is that you will know which one you are looking at.