Agentic AI Job Exposure: 93% of Analyzed Roles Hit Moderate Risk by 2030

The standard framework for measuring AI’s impact on jobs has a flaw: it assumes automation picks off individual tasks while the occupation survives around the remaining work. A new research paper argues that agentic AI systems — which execute entire multi-step workflows autonomously — break that assumption at the foundation.

The paper, Agentic AI and Occupational Displacement: Multi-Regional Task Exposure Analysis by Ravish Gupta and Saket Kumar, introduces the Agentic Task Exposure (ATE) score, a framework built specifically for systems that receive goal specifications rather than task assignments and deliver completed outputs without human handoffs. Applying it across 236 occupations in five US tech regions, the authors find that 93.2% of analyzed roles cross the moderate-risk threshold in Tier 1 regions by 2030.

Context

Prior AI exposure frameworks — including those built on GPT-4 and Frey and Osborne’s foundational computerizability study — share a common structure: they assess how much of an occupation’s task content is automatable, then implicitly assume the occupation persists as a “coordination shell” around whatever work remains. This reasoning has historically justified optimism about white-collar employment resilience. Even if 80% of a paralegal’s tasks become LLM-automatable, the coordination, judgment, and client-facing work supposedly keeps the role intact.

Agentic AI systems challenge that assumption directly. They do not automate tasks in isolation — they receive research goals, retrieve case law, draft memoranda, cross-check citations, revise for style, and deliver complete work products. The coordination bottleneck that protected knowledge worker roles is now inside the agent’s planning module.

The Detail

The ATE score formula weights three components for each task in each occupation: an AI capability score derived from benchmark performance data, a workflow coverage factor that penalizes tasks requiring human coordination or exception handling that agents cannot yet handle autonomously, and a regional adoption velocity function following a logistic S-curve calibrated to each metro’s AI concentration.

The workflow coverage factor is the methodological innovation. Prior exposure indices implicitly set this to 1.0, treating every task as independently completable. The ATE score explicitly discounts tasks involving real-time interpersonal dynamics (negotiation, counseling, de-escalation), regulatory fiduciary accountability (diagnosis, certification, sworn testimony), physical presence, and high-frequency exception handling. Even with these penalties applied, the resulting exposure scores are higher than task-level analyses would suggest — because agentic systems handle the workflow coordination that prior frameworks credited as a human advantage.

Regional adoption velocity is calibrated from fastest to slowest: San Francisco Bay Area (Tier 1, growth rate k=0.85, inflection 2024), Seattle and Austin (Tier 2, k=0.62, inflection 2025), New York (Tier 3, k=0.48, inflection 2025.75). The model treats Seattle’s 2026–2027 workforce effects as a leading indicator of national patterns that emerge 2028–2030.

The highest-exposure occupations in the study are credit analysts, judges, and sustainability specialists — reaching ATE scores of 0.43–0.47 in Tier 1 regions. These roles share a structural characteristic: high-volume, document-intensive analytical work where the workflow coverage factor stays high even after applying penalty discounts. Credit analysts receive zero keyword-based workflow penalties across their 11 O*NET tasks; their ATE scores represent genuine exposure, not artifact.

Remote work introduces a convergence effect that matters for geographic interpretation. Remote workers in San Francisco firms are employed across all regional tiers, transmitting higher-tier adoption pressure into slower-adopting metros. For New York (Tier 3), the remote-adjusted model increases Financial ATE by 9.2% — meaning the adoption frontier arrives earlier than regional economic data alone would predict. This connects to the broader labor market dynamics that prior analysis of AI’s impact on developer workflows has begun to quantify.

The paper also identifies 17 occupational categories benefiting from reinstatement effects — new roles concentrated in human-AI collaboration, AI governance, and domain-specific AI operations. These are not speculative; they reflect documented hiring patterns in the organizations deploying agentic systems at scale.

Source: arXiv

Content created with AI assistance and reviewed for accuracy.