Introducing Experience Levels in the AI Resilience Report

Product update · September 18, 2026 · 11 min read

An occupation is not one thing

Since the AI Resilience project launched, the question underneath every career page has been “how resilient is this occupation?” As the body of research into AI impacts on work has grown, we are able to ask additional questions. Among the most pressing are questions about experience: Does AI impact entry-level workers differently than mid- or senior-level workers? A first-year software developer and a staff engineer with fifteen years experience, for example, share a job title and a Bureau of Labor Statistics code, but the tasks comprising their day-to-day work are very different. So, how might these different daily task compositions affect automation pressure?

This release is our first attempt at answering that question.

The research informing our approach

Accounting clerks and inventory clerks had large, roughly equal shares of their work automated by computers between 1980 and 2018. A model that only counts how much of a job gets automated would expect them to end up in the same place. But, as MIT researchers David Autor and Neil Thompson show, accounting clerks saw wages rise and employment fall, while inventory clerks saw wages fall and employment rise. The same degree of exposure led to opposite outcomes.

Autor and Thompson’s explanation is that tasks within an occupation are not interchangeable. They differ in the expertise they require: the specialized knowledge, training, and judgment it takes to do them well. Expertise is hierarchical. A worker who can do the hardest task in a job can do all the easier ones too, but not the other way around. So the most expert task in an occupation sets its barrier to entry.

Automation reshapes that barrier depending on which tasks it removes:

• Remove the routine tasks (basic bookkeeping for the accounting clerks) and the work that remains is more specialized. The barrier to entry rises. Fewer people qualify, wages go up, and there are fewer jobs.
• Remove the expert tasks (the verification work that set experienced inventory clerks apart) and the work that remains is more generic. The barrier drops. More people can do the job, wages fall, and employment rises. The premium on experience erodes.

Autor and Thompson tested this across 303 occupations and four decades of U.S. data, and the expertise direction predicted wage and employment changes independently of how many tasks were added or removed. A 2026 follow-up by Teeselink and Carey (AI, Automation, and Expertise) asked whether the same pattern holds for generative AI, using hundreds of millions of job postings across 39 countries after ChatGPT’s release. It does. Where AI exposure raised expertise requirements, wages rose; where it lowered them, they did not.

That gave us a framework that is both well-evidenced and, conveniently, built from data our model already has: a scored list of every task in every occupation.

Our approach to including experience in the report

We scored every task for expertise. Every O*NET task in the report already carried an automation likelihood, our model’s estimate of how likely AI is to fully take a task over in the near term. Each task now carries a second score, from 0 to 1, for the expertise it requires, judged in the context of its own occupation. A language model did the scoring against five fixed anchors: 0.0 to 0.2 is routine work most adults could do without preparation (filing documents, greeting customers); 0.2 to 0.4 takes weeks to months of training (standard office software, basic data entry); 0.4 to 0.6 takes years of training and some professional judgment (drafting reports, troubleshooting equipment); 0.6 to 0.8 takes significant experience, professional judgment, or credentials (analyzing complex data, managing projects); and 0.8 to 1.0 is elite work that defines the barrier to entry (performing surgery, designing a structural system).

We split each occupation into three experience bands. One of the core challenges we needed to address is that O*NET does not say which tasks a first-year worker does and which ones take a decade to earn. The expertise scores give us a way to infer it. We sort each occupation’s tasks by their expertise score and cut the list into thirds, weighting core tasks twice as heavily as supplemental ones: the bottom third defines Entry-Level work, the middle third Mid-Level, and the top third Senior. So a task’s band is set by its score, but the cut points are relative to the occupation rather than fixed score ranges. A task scored 0.5 is senior work in one occupation and entry-level work in another. That follows Autor and Thompson, who treat expertise as occupation-relative: entry-level surgery is still surgery. A worker at a given band does that band’s tasks plus everything below it, since expertise is hierarchical.

We drew the task map. Every career page now includes a scatterplot that displays individual tasks by resilience and expertise. A dashed line at 50% marks the automation threshold. Pick a band and a shaded window highlights the tasks that define that level.

We did not include experience in the topline resilience score. More on this below.

Results

Across the 938 occupations with scored tasks, 459 are more favorable to senior-level workers, 73 are more favorable to entry- and mid-level workers, and 406 are neutral. In other words: in about half of occupations, AI is poised to erode (or already eroding) junior-level work, while eroding senior-level work in just 8% of occupations.

Software developers are a clean example of an occupation in which AI is eroding junior-level work. About 56% of the tasks that define entry-level work there (e.g., writing status reports, storing and retrieving data for analysis) are already past the automation threshold. At senior level, none of the tasks are past the automation threshold. The expertise shift ranks in the top 10% of all careers.

Information security analysts are the second case. Three of the four tasks that define senior work (regulating access to data files, encrypting transmissions, running risk assessments) are past the automation threshold, while entry-level work like training users and reviewing security violations mostly holds up. The occupation is also growing, with BLS projecting employment up 21% from 2025 to 2035, which is what the framework predicts when the barrier to entry falls.

Why experience level does not (yet) factor into occupational resilience scoring

The experience view sits alongside an occupation’s resilience score rather than inside it, the same arrangement we use for location data. Selecting Entry, Mid, or Senior changes the task map and the outlook text, but does not change the topline resilience score.

We are keeping it out because the experience layer currently rests on a single scoring run and a framework with one AI-era study behind it. The four limitations below are the reasons in detail. They are also the list of what would have to improve before we consider factoring experience into the score, and we’ll return to the question as the expertise scores accumulate refresh history and as more evidence on generative AI comes in.

Bands are thirds of a task list, not years of experience. The year ranges on the picker (“0 to 2 years,” “8+ years”) are there to orient users and don’t correspond to anyone’s actual tenure. Because we split each occupation’s scored tasks into thirds, band size depends on how many tasks the occupation has. An occupation with twelve scored tasks has about four per band, which makes the bands coarse and means a single task’s expertise score can shift where the cut falls. Occupations with no scored tasks (59 of 997 detailed occupations) don’t show the feature.

Expertise scores are model estimates. A language model assigned them by reading O*NET task descriptions against fixed anchors, the same way our automation scores are produced. Two things follow. Scores can shift between scoring runs, so an occupation’s label may change at a quarterly refresh without any change in the underlying work. And a score like 0.62 is an ordinal judgment, useful for ranking tasks within an occupation, not a measured quantity. Teeselink and Carey found LLM expertise rankings stable across prompt variations and correlated with worker education and pay. That supports using the scores to rank tasks. It does not support reading precision into any particular value.

The 50% threshold is a choice. A task “survives automation” if our model puts its automation likelihood below 50%. We use that cut because it is where the model’s estimate flips from more likely automated to more likely not, and because it is the same automation line already drawn on every task map. It also mirrors Autor and Thompson’s own design, which treats a task as either removed by automation or kept, with no partial credit. It is still a single cut on a continuous score. Move it to 40% or 60% and the expertise shift moves with it. The occupations most likely to flip are the ones whose tasks cluster near the threshold.

The framework was validated on earlier automation, not on generative AI. Autor and Thompson’s evidence covers computerization from 1980 to 2018. Teeselink and Carey’s replication on post-ChatGPT job postings is the only AI-era test so far, and it covers a few years of postings rather than decades of wage and employment outcomes. We are applying a pattern established under one technology to a newer one on the assumption that it holds. If generative AI reaches expert tasks in ways earlier automation did not, the framework would point the wrong way for those occupations.

If you’re using the experience view in advising, in the classroom, or for your own decisions, tell us what’s helping and what’s confusing. Reach us at air@careervillage.org