How the Job Resilience Score works

What the model measures

The model starts with three occupation baselines: AI exposure, human moat and augmentation upside. It then adjusts them using the respondent's task mix, career level, digital workflow, repeatability, formal accountability, relationships, unusual situations and workplace AI adoption.

AI exposure measures how much work current generative AI can reach. Automation pressure adds workflow and organisational incentives. Human advantage measures protection from trust, accountability, physical context and judgment. Augmentation upside estimates how much AI may improve the work while a person remains responsible.

Calculation sequence

All displayed scores are rounded and clamped to 0–100. Career-level adjustments recognise that junior work is often more routine while senior and management work usually carries more accountability. These are model assumptions, not universal facts.

Occupation data

The lightweight tool uses a compiled index of 4,861 occupation titles across 13 countries. For each normalised occupation it stores only the fields needed for search and scoring: title, category, represented markets, exposure baseline, human-moat baseline and augmentation baseline. Full salary, migration and career records are deliberately kept outside this tool.

Where multiple country records share a normalised occupation slug, the tool averages valid baseline values. A US record is preferred as the display label when available, followed by Australia and then the first available market record. The current client dataset was generated from the project's occupation snapshot dated 2026-07-16.

Baseline provenance and fallbacks

The compiled source records carry an exposure-method label. Depending on the source occupation and available crosswalk, that label identifies either an Eloundou-style occupational AI exposure mapping or an ILO generative-AI occupational exposure mapping. Crosswalks may be direct through SOC or ISCO codes, grouped at a broader code level, or produced through a documented title-to-occupation mapping step.

The compact tool does not claim that these research measures directly predict layoffs. They are used only to establish a relative starting point. During index generation, an explicit automation_exposure value is preferred; the occupation's AI-risk rating is the fallback. Valid values are averaged across records that share a normalised slug. Human-moat values fall back to 5/10 and augmentation-upside values to 6/10 when those fields are absent. These fallbacks prevent a missing field from becoming a false zero.

A method label describes lineage, not certainty. Country records may use different occupational classifications, granularity and publication dates. The normalised slug improves search coverage but can merge specialties whose real task profiles differ. That is one reason the personal task questions carry substantial weight.

How to read the score bands

Exposure-oriented scores use four descriptive bands: 75–100 High, 50–74 Moderate, 30–49 Limited and 0–29 Low. The occupation landing pages use a slightly more detailed presentation for baseline exposure: 75–100 High, 60–74 Elevated, 40–59 Moderate and 0–39 Lower. These labels organise the interface; they are not statistical confidence intervals.

The Job Resilience Score is interpreted as 76–100 strong human defenses, 61–75 likely to change more than disappear, 46–60 meaningful exposure and 0–45 substantial automation pressure. A higher resilience score does not mean no change. It can also reflect high augmentation potential, where the person remains responsible but the tools and workflow change quickly.

Small differences should not drive major decisions. A score of 62 is not meaningfully safer than 60 without additional evidence. Look first at which component moved, then at the tasks and context that produced it.

Worked interpretation example

Consider an occupation with a relatively high digital-work baseline. If a respondent reports that routine information work is central, inputs are mostly digital, rules repeat often and AI is already widely adopted, both task automation and context risk rise. Automation pressure can therefore exceed the occupation baseline.

The same title can produce a different result when the respondent owns regulated outcomes, handles unusual cases, maintains long-term relationships or performs on-site work. Human advantage rises and automation pressure falls even though the occupation baseline has not changed. Seniority adjustments also move the scores because junior task mixes are often more execution-heavy while senior roles more often carry judgment and accountability.

This example demonstrates what the tool is designed to do: reveal which assumptions drive the output. It does not prove that either worker will keep or lose a job.

Calibration and interpretation

The weights are editorial model parameters designed to keep the result interpretable and sensitive to actual work. They have not been validated as a causal forecast of layoffs. A 70 exposure score does not mean a 70% probability of replacement, and differences of a few points should not be treated as statistically significant.

Use the scores to compare task patterns, identify where human review matters and plan experiments. Do not use them as the sole basis for employment, education, financial or legal decisions.

Known limitations

• Occupation averages hide differences between employers, countries and specialties.
• AI capability and adoption can change faster than the data snapshot.
• The model does not estimate labour demand, wages, regulation or employer strategy.
• Self-reported task estimates may be incomplete or biased.
• Physical robotics exposure is not modelled in the same depth as generative AI exposure.

Privacy and reproducibility

The calculation runs in the browser. Answers are not submitted to the site. A shared result URL contains only the selected occupation slug, the five displayed scores and the model version; it excludes task answers, country, career level and goal. See the privacy notice for details.

The formulas above correspond to Model 1.0. Material scoring changes will increment the model version and be recorded here.

Update policy and model history

Occupation-data refreshes update the snapshot date. Changes to coefficients, task families, context questions, fallback values or score interpretation require a model-version change. Copy edits, accessibility improvements and corrections that do not change a result may be released without incrementing the model.

Model 1.0, July 2026: initial public scoring model with six task families, six context factors, career-level adjustments, four component scores and a weighted Job Resilience Score. The public occupation index contains 4,861 normalised titles from 13 represented countries.

When results from different model versions are compared, the version should be reported with the score. Shared-result links reject unknown model versions instead of silently interpreting them with newer logic.