Updating Our Scoring Model to Better Measure AI Resilience

From the beginning, we've defined AI resilience as the degree to which an occupation continues to offer sustained economic opportunity, employer demand, and meaningful human contribution as AI transforms tasks. When we launched the AI Resilience Report in March, our model could measure two of those three dimensions well: through AI exposure datasets, we can begin to understand how AI might change the way humans engage in the tasks that make up an occupation, and through long-term employment projections, we can understand how demand for occupations may change over time.

The last dimension, sustained economic opportunity, which we're defining as the degree to which an occupation can continue to provide real economic opportunity over time, was more challenging, and we didn't have great data for it. A job could show up as “Stable” (low AI exposure, steady or growing demand) and yet still leave the person doing it in a precarious position economically. Or, a job could show up as “Changing Fast” because of high AI exposure, and yet still be highly resilient. Our early model, heavily weighted for AI exposure, couldn't present a holistic picture of occupational resilience.

Now, with two additional datasets, we've taken a step closer to a richer, more actionable model of AI resilience.

A new dimension: sustained economic opportunity

Our updated model adds a dimension we are calling sustained economic opportunity. It draws on two recent pieces of research that get at something our original model couldn't capture.

The first is Lukas Althoff and Hugo Reichardt's wage bill projections, which measure the total dollars flowing to workers in an occupation. It tells you whether a career is gaining or losing ground as a share of the overall economy, not just what any individual worker earns.

The second is Sam Manning and Tomas Aguirre's Adaptive Capacity Index, which measures how well workers in a given occupation can adapt to economic shifts, whether through transferable skills, adjacent career pathways, or the flexibility of the role itself.

Together, these datasets help us assess whether an occupation can continue to provide real economic opportunity over time. This composite is our best current proxy for that broader concept, not a complete measure of it.

How the pieces fit together

The AI Resilience Report now evaluates occupations across three dimensions:

AI and automation exposureaccounts for 40% of the overall score. This is the core of the report and it draws on exposure estimates from Anthropic, Microsoft, the WRTMJ framework, and our own internal model. A higher exposure score means more of the job's tasks can be performed or significantly accelerated by AI. As before, we ‘flip’ the exposure score, subtracting it from 100, to derive resilience to AI and automation, which we're calling ‘meaningful human contribution’.

Long-term employer demandaccounts for 30% of the score. This is what we previously called the “labor market outlook”: a composite of BLS projected growth rate and annual openings over the next decade.

Sustained economic opportunity accounts for the remaining 30%. This is the new dimension, combining the Adaptive Capacity Index and the Wage Bill data described above.

AI exposure remains the dominant signal, largely because we have more data feeding into it, but the additional dimensions provide fundamentally important context. A job with low AI exposure and strong demand looks different if the workers in it have limited economic mobility and declining wage share than if they have robust adaptive capacity and a growing economic footprint. Additionally, we expect the weighting to change over time as research around AI's labor market impact matures and more data feeds into our model.

Why the labels changed

The shift in methodology called for a shift in how we label and communicate results. Our updated model uses six categories along a resilience spectrum:

• Highly Resilient
• Resilient
• Mostly Resilient
• Somewhat Resilient
• Not Very Resilient
• Vulnerable

Every category now answers the same question: how resilient is this occupation? But the label alone can't tell you why. Our original three categories, Changing Fast, Evolving, and Stable, were interpretive. They described how fast things were changing, which worked when the model was built primarily around one signal. With three dimensions that interact in different ways, that kind of interpretation would be misleading. Two occupations can score identically and land in the same category for completely different reasons. The new labels are more observational. They tell you where an occupation falls on the resilience spectrum, but you have to look at the underlying dimensions to understand what's driving that placement.

What moves, and why it matters

When you add a new dimension to a model, some occupations shift. But how much things move, and how broadly, matters.

Across all occupations in the report, the average resilience score shifted by −1.69 percentage points under the updated model. The median shift was slightly larger at −2.31 percentage points. What this tells us is that a lot of occupations that looked solid on exposure and demand alone have softer economic fundamentals once you account for adaptive capacity and wage bill data. The new dimension meaningfully adds to our understanding of occupational resilience without dramatically reshuffling the model.

About two-thirds of occupations shifted less than 10 percentage points in either direction, and roughly a third shifted less than 5. Where it gets interesting is the remaining third: occupations that shifted more than 10 percentage points. Some of these are occupations that appeared safe on exposure and demand alone but turn out to sit on shaky economic ground. Others looked threatened by AI but have strong enough economic fundamentals to absorb the disruption.

Most occupations shifted less than 10 percentage points from their prior score.

Case studies: Three occupations

Each of the case studies below shows how the updated model changes the picture of resilience, and where it still has gaps.

Carpet Installers dropped 25.2 percentage points under the new model, landing at Mostly Resilient, although just barely, with a score of 50.5% against a 50% cutoff. Previously, they would have scored as Resilient. This is an occupation with low AI exposure: the work is physical and hard to automate, and that alone used to be enough to look solid. The other two dimensions tell a different story, however. Long-term employer demand is very low: the ten-year growth rate projection sits at −10%, and there are only 1,100 annual openings. While wage bill data presents a picture of moderate economic opportunity, adaptive capacity is low, meaning workers in this field don't have great options if conditions shift.

Carpet installation isn't being disrupted by AI, but the economic ground underneath it is softer than an exposure-only model could see.

There is, however, something the model isn't yet capturing here: many carpet installers are sole proprietors or work for very small businesses. The adaptive capacity measure doesn't account for the transferable skills that come with running a business, like managing clients, pricing jobs, and handling logistics. It's possible, then, that the economic picture for carpet installers is somewhat stronger than the data suggests.

Software Developers gained 28.9 percentage points, the largest positive shift on the list. This is probably the occupation people most associate with AI job disruption, and the exposure scores back that up: software development ranks among the most exposed in the report. But developers land at Mostly Resilient under the updated model, because the economic picture is strong across the board. Software developers rank fifth out of all occupations for long-term employer demand. Adaptive capacity is also high: the skills are broadly transferable and the adjacent career pathways are wide. The wage bill is large and growing. These signals, taken together, put AI exposure in a very different context.

One thing worth noting, though, is that the demand and wage bill data reflect the market as it exists today. It is possible that AI exposure functions more like a leading indicator that forecasts a shift in economic fundamentals. The next round of BLS employment projections, due in May, may tell us more.

Exercise Trainers and Group Fitness Instructors barely moved, shifting by just 0.38 percentage points. Low AI exposure, very high employer demand, and a high wage bill that reflects the massive and expanding global fitness industry. Adaptive capacity is low-to-moderate, given the fairly specialized skillset. By the numbers, this occupation scores Resilient under both the old and new models.

But exercise trainers earn a median salary of about $46,000, which is below the national median individual wage. This tells us that the updated model, while improved, doesn't fullyresolve a tension mentioned in the introduction: An occupation can be resilient in the sense that it will continue to exist and grow, while still not providing the kind of economic security the word “resilient” might imply. The wage bill is high because the fitness industry is huge, not because individual trainers are well-compensated. And while adaptive capacity offers an indirect measure of economic stability, there are more direct measures, like individual compensation, benefits, and employment stability, where data exists and could be integrated. It's possible, then, that the model needs additional job quality data for sustained economic opportunity to be fully assessed.

A note about occupations designated as Vulnerable

Thirty-five occupations in our model score as Vulnerable. We made a deliberate decision to include this category, and we want to be straightforward about what it means and why we chose it.

Occupations labeled Vulnerable are ones where the data points to a serious risk of major erosion or displacement. We don't use the word lightly, but we think it would be dishonest not to use it at all. If the purpose of this report is to help people make informed decisions about their careers, we owe them a clear signal when an occupation faces that level of risk. Softening the language wouldn't change the underlying reality, but it would make the report less useful to the people who need it most.

What's revealing about these 35 occupations is that they don't tell a single story. They aren't all “jobs AI will replace.” They arrive at vulnerability through different paths, which is exactly what a three-dimensional model should surface.

One large cluster of vulnerable occupations is administrative and clerical roles where AI is the primary threat. Telephone operators, data entry keyers, typists, proofreaders, credit authorizers, procurement clerks: These are occupations with high AI exposure, and core tasks, such as processing information, routing communications, checking records, are increasingly automated. It's important to note that many of these roles were already contracting before the current wave of AI, and the combination of high exposure, weakening demand, and limited adaptive capacity puts them squarely in the Vulnerable category.

Another large cluster is specialized manufacturing and industrial roles. Foundry mold and coremakers, lathe and turning machine operators, pourers and casters, tool grinders: These occupations may not score especially high on AI exposure, given that the work is physical and skilled, but long-term employer demand is declining, the roles offer limited transferability to adjacent careers, and the wage bill data shows a shrinking economic footprint. These jobs aren't being displaced by AI so much as they're being squeezed by broader economic forces, with no countervailing strengths to offset the pressure.

This mix demonstrates that vulnerability isn't one thing. A multi-dimensional model helps us start to understand where the risk factors are and how they interact.

What comes next

We expect the methodology to continue evolving as new research and data become available. One area we're watching is entrepreneurship and self-employment. Our current model evaluates occupations as employee roles, but a significant share of workers in some occupations are business owners or gig workers, and that may change the economic picture in ways the model doesn't yet reflect. We're also thinking about job quality: whether measures like individual compensation, benefits, and employment stability can sharpen our picture of sustained economic opportunity. Another open question is whether AI exposure functions as a leading indicator, forecasting shifts in economic fundamentals that demand and wage bill data haven't yet registered. And we're interested in geographic and industry variation: it is possible that occupational resilience varies to some extent based on these factors. For example, a marketing manager in sporting goods in Portland may face a different resilience picture than a marketing manager in consumer tech in Seattle.

Research into AI's impact on the labor market is advancing rapidly, and we are working to keep pace with it to ensure that students and job seekers have the most accurate, up-to-date information available.

If you have questions about the updated methodology, please reach out to us.