⚡ electrical engineering

Agentic AI and Occupational Displacement: A Multi-Regional Task Exposure Analysis of Emerging Labor Market Disruption

This paper introduces the Agentic Task Exposure (ATE) framework to demonstrate that autonomous AI agents, which execute entire workflows rather than discrete tasks, pose a significantly higher displacement risk to information-intensive occupations across major US regions by 2030, with 93.2% of analyzed roles exceeding moderate-risk thresholds while simultaneously creating new opportunities in AI governance and human-AI collaboration.

Original authors: Ravish Gupta, Saket Kumar

Published 2026-04-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ravish Gupta, Saket Kumar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the labor market as a giant, complex kitchen. For decades, economists have been watching the chefs (workers) and the new gadgets (robots/AI) trying to figure out who gets replaced.

Previously, the story was simple: The "Chopping" Robot.
Imagine a robot that can only chop onions. It does that one job perfectly. The human chef still has to sauté the steak, season the sauce, and plate the dish. The human is still needed; they just lost the onion-chopping part. This is how we've understood automation for years: AI takes a few small tasks, but the job itself survives.

The New Story: The "Autonomous Sous-Chef"
This paper argues that we are no longer just getting onion-choppers. We are getting Agentic AI—think of them as fully autonomous sous-chefs.

You don't tell an Agentic AI, "Chop the onions." You tell it, "Make the sauce."
The AI then:

Reads the recipe.
Checks if you have the ingredients.
Chops the onions, simmers the pot, and tastes the sauce.
Realizes it's too salty, adds water, and tastes again.
Serves the finished dish to the customer.

It does the whole workflow from start to finish without asking for help in the middle. If one human can now supervise 50 of these AI sous-chefs, the restaurant doesn't need 50 human chefs to do the remaining 20% of the work. It just fires the 50 chefs and keeps the one supervisor.

What Did the Authors Do?

The authors created a new "Thermometer" called the Agentic Task Exposure (ATE) Score.

Old Thermometers: Measured how many individual tasks (like typing or calculating) a job had.
The New ATE Thermometer: Measures how much of the entire workflow (the whole sauce-making process) an AI can handle on its own.

They tested this thermometer on 236 different jobs (like accountants, lawyers, and salespeople) across five major US tech hubs: San Francisco, Seattle, Austin, Boston, and New York.

The Big Findings (The "Menu" of Change)

1. The "White-Collar" Shock
For a long time, people thought AI would only replace factory workers or data entry clerks. This paper says: No, the office jobs are next.
Because Agentic AI is great at reasoning, reading contracts, and analyzing financial data, jobs like Credit Analysts, Judges, and Legal Specialists are now at high risk.

Analogy: It's like a robot that can read a 100-page legal contract, find the dangerous clauses, and rewrite them in seconds. The human lawyer used to do this for days. Now, the AI does it in minutes, and the human just checks the final result.

2. The "Tech Hub" Time Machine
The paper found that the risk isn't the same everywhere. It depends on where you live and how fast your local companies adopt new tech.

San Francisco (Tier 1): This is the "Ground Zero." By 2027, nearly 92% of financial jobs and 100% of legal jobs there will be at "moderate risk" of being replaced by AI.
Seattle, Austin, Boston (Tier 2): These cities are about 2–3 years behind. They will see the same risks in 2030 that San Francisco sees in 2027.
New York (Tier 3): They are a bit further behind, but the wave is coming.
Analogy: Think of it like a storm. San Francisco is getting hit by the rain right now. Seattle is getting the first drops. New York is still sunny, but the storm is moving their way.

3. The "Reinstatement" Effect (New Jobs)
It's not all bad news. When you fire the 50 chefs to hire 1 AI manager, you don't just fire everyone. You create new roles.
The paper identifies 17 new types of jobs that will appear. These aren't "coding" jobs; they are "manager" jobs.

The AI Whisperer: Someone who knows how to tell the AI exactly what to do.
The AI Auditor: Someone who checks the AI's work to make sure it didn't hallucinate (make things up).
The Ethics Guard: Someone who makes sure the AI isn't breaking laws or being unfair.
Analogy: If a self-driving car replaces the driver, you don't just lose the driver. You gain a "Fleet Safety Manager" who monitors the cars and a "Route Optimizer" who plans the trips.

What Should We Do? (The Chef's Advice)

1. Stop Training for the Past
Don't try to retrain a fired accountant to become a software engineer. That's like trying to teach a fish to ride a bicycle.
Instead, train them to become an "AI-Augmented Accountant." Teach them how to use the AI tool to do the boring math, so they can focus on the hard stuff: spotting errors, talking to clients, and making judgment calls.

2. Schools Need to Change
Universities shouldn't just teach students how to build a financial model. They should teach students how to spot a fake financial model built by an AI. The skill of the future isn't "doing the work"; it's "checking the work."

3. Governments Need to Act Faster
Because the AI wave hits San Francisco first, then Seattle, then New York, the government can't wait for a national crisis. They need to have retraining programs ready before the layoffs happen, tailored to specific cities.

The Bottom Line

The paper concludes that Agentic AI is different because it doesn't just take a "task"; it takes the "job."

The Risk: High for office jobs that involve thinking, analyzing, and writing.
The Timeline: It's happening fast in tech hubs (2025–2027) and will spread everywhere by 2030.
The Hope: We won't lose all jobs, but the nature of the jobs will change. We will move from being "doers" to being "directors" of AI.

In short: The robot isn't just chopping onions anymore; it's cooking the whole meal. We need to learn how to be the head chef who knows how to taste the food, not just the one who chops the vegetables.

1. Problem Statement

Existing labor market frameworks (e.g., Autor–Levy–Murnane, Acemoglu–Restrepo) model automation as the substitution of discrete, individual tasks while assuming the occupation itself remains a "coordination shell" for human labor. This paper argues that Agentic AI fundamentally breaks this assumption. Unlike narrow AI (which performs single tasks like translation or image classification), Agentic AI systems execute end-to-end workflows involving multi-step reasoning, tool invocation, autonomous decision-making, and self-correction without intermediate human handoffs.

The core problem is that current exposure metrics fail to capture the risk of workflow-level displacement, where an AI agent replaces the entire occupational function rather than just subtasks. This leads to an underestimation of displacement risk for information-intensive professional roles (e.g., lawyers, analysts) that were previously considered "automation-proof" due to their need for coordination.

2. Methodology: The Agentic Task Exposure (ATE) Framework

The authors introduce the Agentic Task Exposure (ATE) score, a composite metric designed to measure the probability of end-to-end occupational displacement by Agentic AI. The score is computed algorithmically (not via regression) using O*NET task data and calibrated adoption parameters.

2.1 The ATE Formula

For an occupation $o$ in region $r$ at time $\tau$ , the ATE score is calculated as:
$ATE_o(r, \tau) = \sum_{t \in T_o} w_{o,t} \cdot CAP(t) \cdot COV(t, o) \cdot V(r, \tau)$

Where:

$w_{o,t}$ (Task Weight): Importance-weighted relevance of task $t$ within occupation $o$ (from O*NET).
$CAP(t)$ (AI Capability): A score (0–1) representing the current demonstrated ability of Agentic AI to perform task $t$ at professional standards, derived from benchmarks (e.g., MMLU, SWE-bench, AgentBench) and enterprise deployment reports.
$COV(t, o)$ (Workflow Coverage Factor): The critical innovation. This factor (0–1) measures whether an AI agent can complete the task within the standard workflow context without human intervention for coordination, context-setting, or exception handling.
- Penalties: $COV$ is reduced multiplicatively if a task triggers specific barriers:
  - P1 (Interpersonal): Negotiation, counseling, emotional engagement (-25%).
  - P2 (Regulatory/Fiduciary): Legal liability, diagnosis, certification (-30%).
  - P3 (Physical Presence): Sensorimotor coordination, on-site work (-40%).
  - P4 (Exception Handling): Crisis response, novel judgments (-20%).
$V(r, \tau)$ (Adoption Velocity): A logistic (S-curve) function modeling the rate of Agentic AI deployment in region $r$ by year $\tau$ . Parameters ( $k$ , $\tau_0$ , $L$ ) are calibrated based on regional AI ecosystem density (patents, VC, enterprise adoption).

2.2 Data Sources & Scope

Occupations: 236 Standard Occupational Classification (SOC) codes across six information-intensive groups: Financial, Legal, Healthcare, Healthcare Support, Sales, and Administrative/Clerical.
Regions: Five major US tech hubs categorized by adoption tier:
- Tier 1: San Francisco Bay Area (Leading indicator).
- Tier 2: Seattle, Austin, Boston.
- Tier 3: New York.
Time Horizon: Projections for 2025–2030.
Remote Work Adjustment: The model incorporates a remote work feasibility index ( $R_o$ ) to adjust adoption velocity based on the employer's location rather than just the worker's residence, acknowledging that remote workers in Tier 3 regions may be employed by Tier 1 firms.

3. Key Contributions

Theoretical Extension: Extends the Acemoglu–Restrepo framework to account for end-to-end workflow coverage, moving beyond task-level substitution to model the collapse of the "coordination shell."
New Metric (ATE): Introduces a reproducible, algorithmic scoring methodology that explicitly penalizes tasks requiring human judgment, regulatory accountability, or physical presence, distinguishing it from prior "exposure" scores that assumed all tasks were independently automatable.
Multi-Regional Empirical Analysis: Provides the first granular, multi-regional projection of Agentic AI displacement, revealing significant temporal lags (2–3 years) between Tier 1 and Tier 2/3 regions.
Reinstatement Identification: Identifies 17 emerging occupational categories where Agentic AI creates net new demand for human labor, specifically in AI operations, governance, and human-AI collaboration design.

4. Key Results

4.1 Displacement Risk Projections

High Exposure in Tier 1: By 2027, 70.8% of occupations in the San Francisco Bay Area cross the moderate-risk threshold (ATE $\ge$ 0.35). By 2030, 93.2% of all 236 analyzed occupations in Tier 1 cross this threshold.
Top-Risk Occupations: The highest ATE scores (0.43–0.47 by 2030) are found in Credit Analysts, Judges, Sustainability Specialists, and Regulatory Affairs Specialists. These roles have high cognitive AI capability scores and well-bounded digital workflows, despite previous assumptions that they were safe due to "judgment" requirements.
Regional Lag: Tier 2 regions (Seattle, Austin, Boston) show zero occupations crossing the moderate-risk threshold in 2027 but converge to Tier 1 levels by 2030. Tier 3 (New York) lags significantly, with only ~28% of Financial/Legal roles crossing the threshold by 2030.
No "High Risk" (ATE $\ge$ 0.65): Even by 2030, no occupation in the study reaches the "High Risk" threshold. The framework predicts widespread moderate risk rather than catastrophic total displacement, suggesting a gradual workforce recomposition.

4.2 Reinstatement Effects

The study identifies 17 emerging roles driven by the need for human oversight, governance, and collaboration with agents. These are not primarily technical coding roles but domain-expert roles:

AI Operations & Oversight: AI Incident Manager, Agentic System Monitor.
Human-AI Collaboration Design: AI Workflow Designer, Change Management Specialist.
Domain-Specific Direction: Senior AI-Augmented Paralegal, Clinical AI Supervisor.
Governance & Ethics: AI Compliance Officer, Algorithmic Accountability Manager.
Evidence: Early 2026 labor market data shows rapid growth in titles like "AI Reviewer" and "AI Workflow Designer," with wage premiums for AI-skilled workers doubling to 56%.

4.3 Validation

Convergent Validity: ATE scores show strong correlation with the AI Occupational Exposure (AIOE) index ( $\rho = 0.84$ ) and GPT exposure scores ( $\rho = 0.72$ ).
Divergence: ATE scores are lower than raw GPT exposure for roles with high physical/interpersonal barriers (e.g., Travel Agents, Postal Clerks), confirming the $COV$ penalty mechanism works.
Empirical Outcomes: Industry-level analysis (Q1-Q3 2025) shows a 4.2 percentage point divergence in employment growth between high-ATE industries (declining) and low-ATE industries (growing) in Tier 1 regions, supporting the model's predictive validity.

5. Significance and Implications

Policy Timing: The 2–3 year adoption lag between regions implies that workforce transition policies must be regionally staged. Tier 1 regions require immediate intervention, while Tier 2 regions have a short planning window (2028–2029).
Reskilling Strategy: The paper argues against retraining displaced workers as software engineers. Instead, the "reinstatement" pathway involves upskilling domain experts to become AI directors (e.g., a paralegal learning to audit AI legal research). The scarce input is domain knowledge; AI tool proficiency is a learnable complement.
Education Reform: Higher education must shift from teaching workflow execution (e.g., building a DCF model) to AI-augmented judgment (e.g., auditing an AI-generated model for hallucinations or circular logic).
Employer Responsibility: Current WARN Act thresholds (mass layoffs) may not capture the gradual displacement caused by Agentic AI substitution. The authors advocate for voluntary disclosure of AI deployment plans to allow for proactive transition support.

Conclusion

The paper concludes that Agentic AI represents a structural shift in automation risk, moving from task-level substitution to workflow-level displacement. While the risk is broad and moderate rather than catastrophic, it is highly concentrated in information-intensive professional roles and follows a distinct regional adoption curve. The framework provides a reproducible tool for policymakers and educators to anticipate these shifts and design targeted interventions for workforce transition and curriculum reform.