How We Answer the
Questions That Matter
AI is not coming for your job. It's coming for the parts of your job that were never the best use of you in the first place.
Understanding which tasks are changing is more actionable than fearing a headline. When you can see exactly which tasks are at risk and which are uniquely yours, the conversation shifts from “will I be replaced?” to “where do I double down?”
Contents
1. Why This Matters
A new headline every week tells you AI will take your job. The next week, another says it won't. Both are probably wrong about your role, because neither one looked at it.
They're using blunt instruments: single frameworks that treat entire occupations as a unit. “Accountants are 78% exposed.” “Nurses are safe.” But you're not a statistic. Your role is a collection of individual tasks, and some of those tasks are deeply human while others are already being handled by AI. The only way to know where you stand is to look at each one.
No single AI exposure model is a reliable predictor of actual unemployment risk. However, an ensemble approach combining multiple models accounts for an additional 18% of variation in unemployment risk beyond baseline controls.
Frank, Ahn & Moro, PNAS Nexus (2025)
A landmark study from Harvard Business School (Dell'Acqua et al., 2023) tested 758 BCG consultants and found the “jagged technological frontier”: for tasks inside AI's capability zone, consultants using AI produced 40% higher quality work. But for tasks outside that zone, those who relied on AI performed 19 percentage points worse.
You need to know which of your tasks are inside that frontier and which are outside. Getting it wrong costs you either way: in missed productivity or in misplaced trust.
22%
of jobs structurally disrupted by 2030 (WEF)
39%
of core skills will change (WEF)
63%
of employers cite skills gap as #1 barrier
2. The Three Questions
Every role is different, and so is the answer. We analyse the tasks that make up yours and answer the three questions you're really asking:
Can AI already do parts of my job?
We measure AI feasibility for every task in your role. Not the occupation, not the industry, but your actual work.
What should I double down on?
We profile the human capabilities each task demands, the things AI can't replicate today, so you know where to invest in yourself.
How much time do I have to adapt?
We project how AI capabilities advance in your domain, and how long before your organisation is likely to deploy them.
The result is your personal roadmap: a task-by-task picture of what makes you irreplaceable and exactly where to focus next.
See where your role stands.
Upload a job description or CV and get your task-by-task breakdown.
Try the Task Breakdown3. Can AI Already Do Parts of My Job?
For every task in your role, we measure how much of it AI can handle today. This isn't a guess. It's a score from 0 to 1, built from four independent research frameworks that each look at the question from a different angle.
How exposed is this task to AI?
Eloundou et al. (2024). “GPTs are GPTs.” Science, 384(6702).
Published in Science, this study classified every occupational task in the US economy into three exposure levels:
AI alone can cut this task's time by 50%+. Drafting emails, translating documents, generating code.
AI plus tools can cut the time by 50%, but a human still guides and approves. Financial modelling, code architecture.
AI cannot currently reduce this task's time by 50%. Physical presence, complex judgment, trust relationships.
80%
of US workers have some task exposure
19%
have 50%+ of tasks exposed
#1
surprise: higher income = more exposed
Is this task actually suitable for AI?
Brynjolfsson, Mitchell & Rock (2018). AEA P&P, 108.
Exposure doesn't mean suitability. An 8-factor rubric separates tasks AI can technically do from tasks a business would actually hand over:
Favours AI (+1 each)
- • Well-defined output
- • Abundant training data
- • Repetitive or rule-based
- • Errors are tolerable
Favours Human (-1 each)
- • Requires empathy or trust
- • Requires creative judgment
- • Requires physical presence
- • Requires zero-error accountability
Is this routine work or expert judgment?
Autor, Levy & Murnane (2003). QJE, 118(4). Updated 2025/2026.
AMPLIFY
Coding, analysis, research
OWN
Surgery, skilled trades, therapy
OFFLOAD
Data entry, bookkeeping
OFFLOAD
Assembly, inspection, filing
Before generative AI, the pattern was simple: routine tasks at risk, expert tasks safe. AI has broken that pattern. For the first time, nonroutine cognitive work (coding, research, analysis) is exposed. But nonroutine interpersonal and physical tasks remain resilient.
The 2025 update: expert vs. inexpert
When automation eliminates inexpert (supporting) tasks, workers shift to higher-value work and wages rise. When it eliminates expert (core) tasks, wages decline. Which kind of tasks make up your role matters enormously.
Can AI agents chain tasks together?
Shao et al. (2025). Stanford WORKBank. Dominski & Lee (2025). OAIES.
When tasks form unbroken chains where output flows directly to the next step without human judgment, reliability compounds downward. Even at 90% accuracy per step, a 10-step chain delivers only 35% end-to-end reliability. We factor this into the risk assessment.
90% × 90% × 90% … (10 steps) = 35% end-to-end
This is why high-stakes multi-step workflows still need human oversight.
How does this compare to real-world occupational data?
Felten, Raj & Seamans (2021). Strategic Management Journal, 42(12). O*NET v29.1.
Every task is cross-referenced against the U.S. Department of Labor's O*NET database, the largest occupational dataset in the world. This grounds our scores in empirical reality, not just AI model judgment.
1,000+
occupations mapped
19,000+
task statements classified
55,000+
alternate titles indexed
4. What Should I Double Down On?
Even when AI can do a task, that doesn't mean the human is redundant. The question is: what human capabilities does this task demand that AI cannot yet replicate? That's what tells you where to invest in yourself.
The difference between replacement and amplification
Cazzaniga et al. (2024). IMF SDN/2024/001. Jaumotte et al. (2026). IMF SDN/2026/001.
AMPLIFY
AI makes you better
OFFLOAD
AI replaces you
OWN
Human domain
OWN
Fully human territory
About 40% of global jobs are exposed to AI. That's the number that makes headlines. The part most articles leave out: for a large share of those workers, AI isn't a threat. It's a force multiplier. The key variable isn't exposure. It's complementarity: does AI replace you, or make you more valuable?
The five capabilities AI can't replicate today
Loaiza, I. & Rigobon, R. (2024). The EPOCH of AI. SSRN 5028371. MIT Sloan.
We profile every task across five dimensions of human capability, derived from Loaiza & Rigobon's EPOCH framework. They approach from AI's statistical limitations: scenarios where universal approximation functions fail (biased data, extrapolation beyond training, moral dilemmas) and identify the human capabilities that complement those gaps.
Empathy
Social cognition, interpersonal connection, emotional attunement. The ability to understand and respond to someone's emotional state in a way that builds genuine trust.
Presence
Embodied cognition. Not just being physically there, but integrating sensory-motor experience into the work itself: surgery, skilled trades, physical therapy, and any work where the body is the instrument.
Opinion
Moral reasoning, causal judgment, tacit expertise. Making decisions under genuine uncertainty where there is no objectively correct answer, only better and worse judgment calls.
Creativity
Novel ideation, unconventional framing, aesthetic judgment. Not pattern matching from training data, but genuinely new ideas that emerge from lived experience and individual perspective.
Hope
Meaning-making, inspiration, affective cognition. The capacity to provide motivation, purpose, and emotional grounding: leadership, mentoring, chaplaincy, and community building.
Your strongest EPOCH dimension defines your resilience floor. A task requiring deep empathy is irreplaceable even if its creativity demand is low. The dominant capability is what matters.
Implementation note: Loaiza & Rigobon provide the five EPOCH dimensions and occupation-level scores. Our aggregation into a single Human Capability Demand scalar (75% dominant dimension, 25% averaged remainder) is our own engineering choice, designed to reflect that a task's resilience floor is set by its strongest human capability demand, not an average across all five.
Risk signals: when AI can do it but shouldn't be left alone
Some tasks are technically automatable but the stakes are too high. We assess five risk signals that can push a task into human oversight regardless of AI capability:
5. How Much Time Do I Have to Adapt?
Your score today isn't your score in two years. AI capabilities are advancing at different speeds in different domains. We project how your score evolves at 6, 12, and 24 months, then factor in the reality that most organisations move much slower than the technology.
How fast is AI improving in your domain?
Sharpe & Tyndall (2025). Measuring AI Agent Performance on Time-Limited Tasks. METR.
Based on METR's empirical evaluations of frontier AI models, we use domain-specific doubling rates, measuring how quickly AI capability doubles in each area:
90d
Digital
Software, data, ML
150d
Reasoning
Strategy, judgment
300d
Interpersonal
Empathy, trust
600d
Physical
Sensorimotor tasks
Digital tasks see AI capability double every 3 months. Interpersonal tasks, where empathy, trust, and motivation matter, take nearly a year. This difference is enormous for your career planning.
When will your organisation actually deploy it?
McKinsey (2024). OECD (2024). Acemoglu (2025).
Just because AI can do something doesn't mean your employer will use it tomorrow. The gap between technical capability and organisational deployment varies dramatically:
Regulatory requirements add further delay. A therapist role in healthcare with HIPAA compliance: METR projects capability in 18 months, add 30 months sector adoption, add 12 months HIPAA. that's 5 years from “AI can do therapy” to “your employer deploys it.” That difference changes your entire strategy.
6. How the Engine Works
Behind the three questions is a classification engine that separates measurement from judgment. An AI model measures each task's features. A rule-based decision tree on our server makes the classification. The AI never assigns labels. It only provides the data the rules operate on.
Step 1: We break your role into tasks
Your job description or CV is decomposed into 8–15 high-level work activities. Each task is matched against 19,000+ pre-classified task statements in our library for consistency.
Step 2: Each task is measured on three dimensions
How much of this task can AI handle today?
How much does this task rely on empathy, presence, judgment, creativity, or meaning-making?
What are the consequences if AI gets this wrong? Is it regulated? Is the decision reversible?
Step 3: The decision tree assigns a label
These rules fire in order. The first match wins:
AI feasibility is low. This task is yours.
AI has moderate capability, but the human skills this task demands are too high to hand over.
AI is capable, but failure is catastrophic, regulated, or irreversible. You must oversee.
AI is highly capable, the task needs little human skill, and the stakes are low.
Everything else. You drive quality and judgment; AI accelerates execution.
7. Your Score: The Human Contribution Index
Your per-task labels are aggregated into a single number from 0 to 100: the Human Contribution Index (HCI). It represents how much of your role's value comes from irreplaceable human contribution.
HCI Weight Distribution
OWN
Full human contribution
GUARD
AI assists, human governs
AMPLIFY
AI multiplies effectiveness
Formula
HCI = (OWN×1.00 + GUARD×0.70 + AMPLIFY×0.45 + OFFLOAD×0.00) / n × 100
Why those weights?
OWN = 1.00. The human is irreplaceable. Full contribution.
GUARD = 0.70. AI handles execution, but the human governs. The role shifts from doing to overseeing, still essential, but augmented.
“If AI makes me more productive, why does my score drop?”
Because resilience is not productivity; it's irreplaceability. If AI makes you 3× faster, it also makes your competitor 3× faster. The barrier to entry collapses and the wage premium erodes. Amplification creates value for the firm but fragility for the worker.
AMPLIFY = 0.45. Three independent studies converge: ~15% AI contribution (Brynjolfsson, QJE 2025), ~27% labour cost savings (Acemoglu 2025), and E2 weighted at 0.5 × E1 (Eloundou). All point to 55–85% human retention. The 0.45 weight centres this range.
OFFLOAD = 0.00. When AI fully substitutes for human labour, the human contribution is zero by definition (Acemoglu & Restrepo, JEP 2019).
Transparency note: These weights are author-determined based on synthesis of the evidence above, not directly measured from a single study. No competing framework offers empirically validated contribution weights: Eloundou uses binary exposure tiers, Felten uses continuous scores, WORKBank uses categorical zones. We chose explicit quantification over black-box classification. As the task library grows, we retrospectively test whether these weights produce scores consistent with observed outcomes.
What your score means
Transitional
0–24
Adaptive
25–49
Strategic
50–74
Essential
75–100
Automate & Reinvest
Upskill Aggressively
Direct & Audit
Retain & Reward
Transitional
Automate & Reinvest. Most of this role's tasks can be handled by AI. Focus on efficiency now, and build toward AMPLIFY or OWN capabilities.
Adaptive
Upskill Aggressively. This role is being reshaped, not eliminated. Shift from doing the work to directing and auditing the AI.
Strategic
Direct & Audit. Significant human judgment combined with AI-augmented execution. You design the strategy; AI handles the throughput.
Essential
Retain & Reward. These roles hold institutional value: trust, judgment, strategy, relationships. Their value increases as AI generates more noise.
8. Worked Example
Here's how it looks in practice. Ten tasks from a Marketing Manager job description, each measured for AI feasibility, human capability demand, and risk, then classified:
Worked Example
How a Marketing Manager Scores 63% HCI
HCI Calculation
4 OWN × 1.00 = 4.00
2 GUARD × 0.70 = 1.40
2 AMPLIFY × 0.45 = 0.90
2 OFFLOAD × 0.00 = 0.00
Sum = 6.30 / 10 × 100
63%
HCI
9. How We Keep It Honest
Every role is broken into 8–15 tasks
This ensures a CEO and a Data Entry Clerk are measured on the same scale. It prevents score inflation through micro-task decomposition and makes results comparable across roles. Senior roles are evaluated at the strategic level. “Database management” for a Head of Engineering means architecture governance, not data entry.
Tasks are matched before the AI sees them
Before our AI model scores anything, each task is matched against a library of pre-classified tasks. Close matches (≥90% similarity) are locked to established scores. The AI can't override them. This means identical tasks always get identical scores, regardless of when you run the analysis.
≥90% match, locked to established classification.
82–89% match, suggested as reference, not enforced.
No close match. Full AI scoring, then added to the library for future consistency.
Everything is grounded in O*NET
Every analysis is cross-referenced against the U.S. Department of Labor's database of approximately 1,000 occupations and 19,000+ task statements. This anchors our scores to occupational research, not just AI judgment.
The system gets more accurate over time
Every classification is recorded. When the same task appears in a future analysis, the established score is inherited. The library grows with every analysis, building a compounding empirical foundation.
Scores are calibrated against known benchmarks
We maintain occupational benchmarks spanning the full 0–100 spectrum, derived from the consensus of published research (Eloundou, Autor, IMF, ILO). These serve as the answer key that keeps scoring consistent.
10. What We Don't Know
- Scores are a snapshot. They reflect the current technology cycle (~6 months). We recommend re-auditing every 6–12 months. Our projections extend this window but they're not guarantees.
- Capability isn't deployment. A low score means a task can be automated, not that it will be tomorrow. Our adoption model estimates the lag, but can't predict individual organisations.
- GUARD is transitional. Tasks protected by risk signals today may shift to OFFLOAD as AI reliability improves and regulations evolve.
- Doubling rates are estimates. They come from published METR evaluations but may not hold if AI progress accelerates or plateaus.
- O*NET descriptions can lag. Task descriptions from the Department of Labor may not fully reflect how jobs are performed in 2026.
- Some agentic research is early-stage. Stanford WORKBank and OAIES are at the preprint stage. Core weights are anchored to peer-reviewed research.
- Weight sensitivity. A single GUARD↔AMPLIFY reclassification shifts the HCI by ~2.5 points in a 10-task role. Under alternative weight assumptions (GUARD=0.65–0.75, AMPLIFY=0.40–0.50), overall scores shift by ±3–5 points. Your score is a structured estimate, not a precise measurement.
- Job descriptions aren't reality. JDs include aspirational language. We discount this, but input quality affects output quality.
11. The Research
Everything above is grounded in peer-reviewed journals, institutional reports, and working papers from leading research institutions. Here's the full list.
PR = peer-reviewed | IR = institutional report | PP = preprint / working paper
AI Feasibility Frameworks
- Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs. Science, 384(6702). PR
- Eloundou, T., et al. (2025). Extending “GPTs are GPTs” to Firms. AEA Papers & Proceedings, 115. PR
- Brynjolfsson, E., Mitchell, T., & Rock, D. (2018). What Can Machines Learn? AEA Papers & Proceedings, 108. PR
- Autor, D.H., Levy, F., & Murnane, R.J. (2003). The Skill Content of Recent Technological Change. QJE, 118(4). PR
- Autor, D.H. & Thompson, N. (2025). Expertise. NBER WP No. 33941. PP
- Autor, D.H. & Kausik, B.N. (2026). Resolving the Automation Paradox. arXiv:2601.06343. PP
- Felten, E.W., Raj, M., & Seamans, R. (2021). Occupational AI Exposure. Strategic Management Journal, 42(12). PR
- Shao, Z., et al. (2025). WORKBank and the Human Agency Scale. Stanford. arXiv:2506.06576. PP
- Dominski, J. & Lee, Y.S. (2025). The OAIES. arXiv:2507.08244. PP
Human Capabilities & Complementarity
- Loaiza, I. & Rigobon, R. (2024). The EPOCH of AI: Human-Machine Complementarities at Work. SSRN 5028371. MIT Sloan. PP
- Cazzaniga, M., Jaumotte, F., Li, L., et al. (2024). Gen-AI: Artificial Intelligence and the Future of Work. IMF SDN 2024/001. IR
- Jaumotte, F., Kim, J., Koll, D., et al. (2026). Bridging Skill Gaps for the Future. IMF SDN 2026/001. IR
- Barsalou, L.W. (2008). Grounded Cognition. Annual Review of Psychology, 59, 617–645. PR
Time Horizon & Adoption
- Sharpe, T. & Tyndall, J. (2025). Measuring AI Agent Performance on Time-Limited Tasks. METR. IR
- METR. (2025). Frontier Model Evaluations: Domain Doubling Rates. evaluations.metr.org. IR
- McKinsey Global Institute. (2024). The State of AI: 2024. IR
- OECD. (2024). AI Adoption and Impact in Enterprises. OECD Digital Economy Papers. IR
- Acemoglu, D. (2025). The Simple Macroeconomics of AI. Economic Policy, 40(121). PR
Scoring & Validation
- Frank, M.R., Ahn, S.J., & Moro, E. (2025). AI Exposure Predicts Unemployment Risk. PNAS Nexus, 4(4). PR
- Brynjolfsson, E., Li, D., & Raymond, L.R. (2025). Generative AI at Work. QJE, 140(2), 889–938. PR
- Dell'Acqua, F., et al. (2023). Navigating the Jagged Technological Frontier. HBS WP 24-013. PP
- Acemoglu, D. & Restrepo, P. (2022). Tasks, Automation, and Wage Inequality. Econometrica, 90(5). PR
- Acemoglu, D. & Restrepo, P. (2019). Automation and New Tasks. JEP, 33(2). PR
- Noy, S. & Zhang, W. (2023). Experimental Evidence on the Productivity Effects of Generative AI. Science, 381(6654). PR
- Arntz, M., Gregory, T., & Zierahn, U. (2016). The Risk of Automation for Jobs in OECD Countries. OECD Social, Employment and Migration WP No. 189. IR
- Svanberg, M., et al. (2024). Beyond AI Exposure. MIT Working Paper. PP
- Gmyrek, P., et al. (2025). Global Index of Occupational Exposure. ILO WP No. 140. IR
- Anthropic. (2026). Economic Index: January 2026 Report. IR
- Gruen, C., et al. (2025). AI and Jobs: A Review. arXiv:2509.15265. PP
- Manning, S., et al. (2026). Adaptive Capacity for AI Displacement. Brookings Institution. IR
The Resilience Matrix is a product of applied computational labour economics.
Try the Task Breakdowntheresiliencematrix.com
© 2026 The Resilience Matrix. All rights reserved. Proprietary methodology. Unauthorised reproduction prohibited.