Back to home

Our Methodology

How we assess AI's impact on jobs

1. The Dual Indicator: AI alone vs Augmentation

Unlike simplistic approaches that give a single automation percentage, we distinguish two fundamental dimensions of AI's impact on each task.

AI alone

Measures to what extent AI can perform this task independently. An 80% index means AI produces an equivalent result in 80% of cases, without human intervention.

Ex: Data entry, email sorting, standard report generation

Augmentation Potential

Measures the potential for AI augmentation. A high index means AI can significantly improve human productivity or work quality.

Ex: AI-assisted diagnosis, augmented writing, data analysis

2. The AI Positioning Quadrant

The two indicators are combined in a 4-zone quadrant that visualizes a job's position relative to AI.

Augmented
In Transformation
Low Exposure
High Automation
AI alone →
← Augmentation Potential
40%
40%

Why a 40% threshold instead of 50%?

  • Realistic distribution: most jobs have AI alone indices < 50%. With 50%, the priority transformation zone would be underrepresented.
  • Early warning: 40% already represents a significant level. Waiting for 50% would be too late to prepare.
  • Academic consistency: reference studies (Frey & Osborne, McKinsey) use ~35-40% as a significant threshold.

Low Exposure

Low AI alone, low augmentation. Jobs minimally impacted by current AI (e.g., plumber, nurse).

Augmented

Low AI alone, high augmentation. AI amplifies human capabilities (e.g., doctor, lawyer, developer).

In Transformation

High AI alone, high augmentation. AI transforms key tasks in this job — this is the priority transformation zone (e.g., accountant, data entry operator, designer).

High Automation

High AI alone, low augmentation. AI can perform these tasks without providing complementary value to the worker (e.g., telemarketer, industrial sorter).

3. The Transformation Index (V5)

Our V5 method distinguishes two types of AI assistance and integrates human verification cost for more realistic indices.

V5 Formula

Effective_assistance = (E1 + E2 × tool_adoption) × (1 - verification_cost)Index = Elimination + (1 - Elimination) × Effective_assistance

E1 measures standalone LLM help (writing, analysis, translation). E2 measures LLM combined with external tools (code agents, RPA, CRM). Verification cost (workslop) reduces effective assistance since humans must still verify.

JobAI aloneAugmentationTransformation Index
Data Entry Operator85%10%87%
Web Developer20%60%68%
Nurse5%15%19%

4. Task Weighting

Not all tasks in a job carry the same weight. We use two criteria to weight their impact on the overall index.

Importance (1-5)

From marginal (1) to critical (5) for the job. A critical task weighs more in the final index.

1
2
3
4
5

Frequency (1-5)

From rare/annual (1) to continuous/daily (5). A frequent task impacts daily work more.

1
2
3
4
5

Weight calculation

Weight = Importance × Frequency

5. Reliability: Multi-run Scoring

LLMs can give variable responses. To ensure index reliability, we use a multi-run approach.

Validation Process

  1. 1Each task is evaluated 3 times independently by the LLM
  2. 2We detect outliers (gap > 15 points between runs)
  3. 3Outliers are re-evaluated with a 4th run
  4. 4Final index is the median of valid runs

We use a large language model in a frozen version to ensure the reproducibility of the indices over time.

6. Job Categories and Corrections

We apply corrections based on job category to fix systematic LLM biases.

Routine Cognitive

Accountant, secretary, call center agent

Skilled Cognitive

Engineer, lawyer, doctor

Field Manual

Plumber, electrician, driver

Care & Relational

Nurse, caregiver, educator

Creative

Designer, artist, writer

Commercial

Salesperson, sales rep, real estate agent

Mixed Office

Manager, project manager, HR

7. Sources and References

Our analysis is based on the most comprehensive job reference frameworks and leading scientific studies.

Job Reference Frameworks

  • ROME 4.0

    French job classification (France Travail). 532 job sheets, 11,000 titles, detailed tasks.

  • O*NET 30.1

    US Department of Labor database. 1,000+ occupations with skills, tasks and work context.

  • ESCO v1.2

    European multilingual classification. Harmonization of jobs and skills at EU level.

Scientific References

  • Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023).

    GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models.

    OpenAI Research. arXiv:2303.10130

    openai.com/research/gpts-are-gpts →
  • Frey, C. B., & Osborne, M. A. (2017).

    The future of employment: How susceptible are jobs to computerisation?

    Technological Forecasting and Social Change, 114, 254-280.

    doi.org/10.1016/j.techfore.2016.08.019 →
  • Briggs, J., & Kodnani, D. (2023).

    The Potentially Large Effects of Artificial Intelligence on Economic Growth.

    Goldman Sachs Global Economics Research.

    goldmansachs.com →
  • International Labour Organization (2024).

    Generative AI and Jobs: A global analysis of potential effects on job quantity and quality.

    ILO Working Paper 96.

    ilo.org →
  • Nedelkoska, L., & Quintini, G. (2018).

    Automation, skills use and training.

    OECD Social, Employment and Migration Working Papers, No. 202.

    doi.org/10.1787/2e2f4eea-en →
  • Microsoft Research (2025).

    New Future of Work Report 2025.

    Microsoft Research Technical Report.

    microsoft.com/research →

8. Calibration and Ground Truth

Our indices are calibrated against 42 reference jobs covering 7 categories, with targets derived from 7 leading academic studies.

Calibration Process

  1. 142 reference jobs manually calibrated with targets based on scientific literature
  2. 2Multi-source formula: 15% Frey & Osborne (adjusted), 40% Eloundou/OpenAI, 30% OECD, 15% Goldman Sachs
  3. 3Automated scoring compared to targets, goal RMSE < 10 points
  4. 4Category corrections applied to reduce systematic LLM biases

Technical Exposure

Technical index: what AI CAN theoretically do

Real Impact

Realistic index: what IS actually happening, after country/sector adoption coefficient and regulatory friction

Ready to discover your profile?

Test your job's AI exposure for free

Take the test