How We Measure Your AI Risk

The methodology behind your personalized AI displacement assessment.

This document describes how Expreso generates your AI risk report - the data sources we draw on, the models we use, where our analysis is strong, and where uncertainty remains. We believe that if a system is going to tell you something important about your career, you should be able to see exactly how it reaches its conclusions.

The Question That Matters

Here is the question most AI risk assessments try to answer: how exposed is your job to AI?

It's the wrong question.

Your job title is a label. What actually determines your relationship with AI is the specific work you do every day - the tasks, the decisions, the relationships, the tools, the judgment calls. Two people with the title "HR Director" can have completely different AI exposure depending on whether they spend their time on payroll compliance (increasingly automatable) or organizational design (deeply human). Two software engineers can have opposite risk profiles depending on whether they write boilerplate CRUD applications or architect complex distributed systems.

The major AI exposure studies - from Oxford, OpenAI, the IMF, and others - assign a single risk score to everyone who shares a job title. These studies are useful for understanding broad economic trends, but they cannot tell any individual person where they actually stand. They treat all HR Directors the same. All software engineers the same. All marketing managers the same.

We don't. Our system asks: what do you actually do?

How We Learn About You

The assessment begins with a conversation. An AI-guided interview asks you about your work - not in the generic way a survey would, but adaptively. If you mention managing a team, it asks how many people, what decisions you make about their work, whether you handle hiring and firing. If you mention client relationships, it probes the depth: are these transactional or strategic? Could someone else step in tomorrow, or are you the reason the client stays?

The interview isn't following a fixed script. It's building a picture of your professional identity across three dimensions:

If you connect your LinkedIn, email, and calendar, the system supplements your self-report with observable context: your job history and tenure, professional network signals, and job search activity (or lack thereof). Connected accounts provide supplementary context, though the interview remains the primary evidence source. People tend to overstate impressive-sounding work and understate the routine tasks that actually consume their time - connected data helps us calibrate.

What We Measure Your Work Against

Once we understand what you do, we need to know how AI intersects with those specific tasks. For this, we use two foundational datasets.

The O*NET Task Ontology

The U.S. Department of Labor maintains a database called O*NET that describes nearly every occupation in the American economy in terms of the specific tasks workers perform. It contains over 17,000 distinct task descriptions across roughly 1,000 occupations, each rated for importance by surveyed workers actually doing those jobs.

O*NET gives us a common language. When you tell us you "manage payroll," we translate that into the specific O*NET tasks it maps to - "Process payroll information," "Review time sheets and work charts," "Ensure compliance with federal, state, and local payroll regulations." This translation matters because it connects your personal description of your work to a standardized taxonomy that we can score.

The Anthropic Economic Index

Anthropic, the AI company behind Claude, publishes an Economic Index based on analysis of millions of real AI conversations. Unlike academic forecasts that ask "could AI theoretically do this task?" the Anthropic data answers a different question: are people actually using AI for this task right now?

That distinction is critical. Many tasks that AI could theoretically handle are not yet being adopted in practice - because of regulatory barriers, trust requirements, workflow integration costs, or simply because nobody has built the right tool yet. The Anthropic data captures what is actually happening, not what might happen someday.

For each O*NET task, the Anthropic dataset tells us what fraction of observed AI conversations involved that task. We transform this raw usage signal into a pressure score that reflects where each task sits relative to all other tasks in terms of real-world AI adoption.

How We Calculate Your Score

Your assessment produces a personalized displacement risk score. Here's how we build it, step by step.

Step 1: Build your task portfolio

From the interview and connected evidence, we extract the specific tasks that make up your role and assign each a share of your working time. A marketing director might have 25% of their time in campaign strategy, 20% in budget management, 15% in team leadership, 15% in client presentations, and 25% in performance analytics. These shares matter: a role that's 80% analytics and 20% relationship management has a very different AI profile than one that's 20% analytics and 80% relationships. When you confirm your time allocation during the interview, we use those proportions directly and label them as "confirmed." When we cannot map your confirmed allocation to specific tasks, we distribute time equally and label it as "estimated" - so you always know which shares are grounded in your input and which are approximations.

Step 2: Score each task for AI pressure

Each task in your portfolio is scored for AI pressure using a tiered evidence system. We don't treat all evidence equally - we're explicit about how strong our signal is for each task:

Your AI pressure score is the weighted average of these individual task scores, where the weights are how much of your time each task consumes.

Step 3: Assess your environment

Beyond the tasks themselves, thirteen features of your work environment and personal position affect how quickly AI disruption would actually reach you. These include how digital your work is, how repeatable your processes are, whether your employer is actively adopting AI, whether your work requires physical presence, the liability stakes of your decisions, the depth of your professional relationships, your decision-making authority, your team management responsibilities, how much revenue you directly own, the specificity of your domain expertise, how differentiated your skill combination is, the strength of your professional network, and your current job search traction.

Each of these is estimated from your interview responses by an AI model. They reflect our best assessment of your situation, but they are not direct measurements. Because of this inherent uncertainty, we present these factors as qualitative assessments (Strong, Moderate, Low) rather than precise numbers.

Step 4: Combine into a displacement risk

The final score brings these layers together through a principled multiplicative model with four inputs and a calibration layer. The same inputs always produce the same outputs, with no hidden randomness.

How Confident We Are (And Aren't)

Every score we produce comes with a confidence level. This is one of the most important parts of our methodology, and one that most AI risk tools skip entirely.

Confidence reflects how much of your task portfolio is supported by strong evidence versus weaker estimates. If 80% of your tasks have direct or family-level Anthropic data, our confidence is high. If 60% of your tasks fall to occupation-level priors or unknown estimates, our confidence drops - and we tell you that.

In practice, our Anthropic coverage is strongest for digital, knowledge-work tasks - software development, data analysis, technical writing, business operations - and weakest for physical, regulatory, and highly specialized work. If your role is primarily computer-based, our evidence is relatively strong. If your role involves physical presence, complex regulatory judgment, or highly specialized domain work, our signal is thinner and we'll tell you so.

An additional source of uncertainty: because your protection factors (deployability, human moat, personal buffer) are estimated by an AI model from your interview, some variation between assessments is expected. Running the same interview twice may produce slightly different estimates for these qualitative factors. The task exposure component - which is grounded in the Anthropic Economic Index - is stable and reproducible.

What This Assessment Is Not

Transparency requires saying what we don't claim as clearly as what we do.

The Research Behind This

This methodology builds on a substantial body of academic research on AI and labor markets.

The foundational insight - that occupations should be analyzed as bundles of tasks rather than monolithic jobs - comes from Autor, Levy, and Murnane's influential 2003 study. Frey and Osborne (2017) brought this framework to public attention with their widely cited estimate that 47% of U.S. jobs face high automation risk, though subsequent work by Arntz, Gregory, and Zierahn (2016) showed that task-level analysis yields substantially lower estimates.

The most directly relevant prior work is Eloundou et al. (2024), published in Science, which used both human annotators and GPT-4 to assess the LLM exposure of every O*NET task. Their finding - that roughly 80% of workers could see at least 10% of their tasks affected by LLMs - established the methodology of task-level AI exposure scoring. This was validated by Tomlinson et al. (2025) using Microsoft Copilot usage data.

The Anthropic Economic Index (Handa et al., 2025) introduced the empirical advance our system relies on: measuring actual AI usage rather than theoretical exposure. Massenkoff and McCrory (2026) extended this work by examining actual labor market impacts, validating a core assumption of our model: task exposure alone overstates risk.

More recently, Gans and Goldfarb (2026) have argued that linear task aggregation may overstate risk for roles where tasks are complementary rather than independent. In their "O-ring" framework, automating nine of ten tasks doesn't eliminate the job; it concentrates value on the tenth. Our human moat and personal buffer scores partially capture this dynamic, but we acknowledge the limitation.

One More Thing

The goal of this assessment is not to scare you. It's to give you clarity.

AI is changing the landscape of professional work. Some tasks that were valuable last year are becoming commoditized. Some skills that seemed niche are becoming critical. Some roles are being compressed; others are expanding. The people who navigate this well will be the ones who understand, specifically and honestly, where they stand.

That's what the interview is designed to give you: not a generic anxiety about AI, but a specific, evidence-grounded map of your professional exposure - where you're vulnerable, where you're defended, and what you might do about it. Your report includes a personalized action plan that identifies the highest-leverage moves you can make based on your specific exposure and protection profile.