Automation Exposure

Bottom Line

AI exposure is broad and uneven across tasks. So far, firms are reconfiguring work more than eliminating jobs Mertens (2026); Wang (2026); Zhang (2026). Outcomes vary with adoption barriers, technical limits, liability, and governance; work redesign and skills investment shape results Gao (2026); Chopra. Exposure indices disagree on who is most at risk and often gauge potential, not realized outcomes; early evidence shows unequal access to higher-paying, AI-complementary roles Moreira Tomei (2026); Jetha (2026); Mishra (2026).

What This Means in Practice

• Map exposure by task, not job. Score feasibility and deployment risk to pick safe, high-benefit use cases Papagiannaki (2026); Gao (2026); Vaid (2026).
• Start with narrow, low-risk automations plus audit logs. Add measurable human oversight to reduce errors and cycle time without more incidents Adams (2026); Sabouri (2026); Lipsanen (2026).
• Plan for partial automation. Keep humans on high-complexity judgment; move routine sub-tasks to AI with clear accountability Li (2026); Choudhuri (2026).
• Build equity into upskilling. Use apprenticeships and on-the-job training to shift workers into AI-complementary roles, and track who gets access to higher-exposure, higher-pay paths Jacobs (2026); Mishra (2026).
• Use policy levers to favor labor-complementary uses and curb demand-reducing over-automation. Consider targeted tax incentives, safety standards, and accounting for machine labor Korinek (2026); Hemenway Falk (2026); DOUCOURE (2026).

What the Research Finds

Measuring exposure: capability, learnability, and risk give different maps

• Task-level scoring outperforms static job lists and shows large within-occupation differences Papagiannaki (2026); Vaid (2026).
• Adding liability and safety risk to technical feasibility re-ranks exposure: high-stakes care and unstructured physical work look less exposed; symbolic cognitive work looks more exposed Gao (2026).
• Learnability can diverge from capability overlap: a feedback-driven learning index rates some operators and conductors as highly learnable despite low exposure scores; creative and interpersonal roles score lower Moreira Tomei (2026).
• Benchmarks show limits and steady progress: agents complete a minority of live web tasks today (best about 33%), with rising performance; some audits argue earlier tests understated capabilities Zhang (2026); Mertens (2026); Zanoli (2026).
• Task structure matters: breaking work into micro-actions yields stable clusters. Planning and design tasks look more substitutable, tool-mediated physical tasks less so, but the highest-risk cluster can flip as models improve Gao (2026).

Who is exposed: sectors and groups

• Canada: lower-precarity jobs have higher large language model (LLM) exposure; unstable jobs have lower exposure Jetha (2026).
• India: graduates from Scheduled Castes and Tribes are less represented in generative-AI roles in the same district, and AI-exposed roles pay up to 20% more, pointing to widening caste wage gaps Mishra (2026).
• Sweden: women are concentrated in occupations predicted to be more affected by generative AI, implying gendered risk from pre-AI sorting Gardberg.
• Routine and mid-skill roles are often flagged as higher exposure across contexts, including service and administrative work in Albania and routine admin tasks in Nigerian firms Arslan; Bimbari; Peter.
• China (policy-driven automation diffusion): wider skill wage gaps were associated with slower wage growth in high-exposure occupations; vocational education and on-the-job training mitigated the effect Wei (2026).

How organizations respond: task redesign and bounded automation

• As AI diffuses, firms reallocate hiring across jobs and redesign tasks within jobs; senior roles shift earlier via reallocation and junior roles via both margins Wang (2026).
• Automating low-risk code review was associated with shorter time-to-close and lower revert and production-incident rates than human-only review Adams (2026).
• Exposing and controlling agent actions in spreadsheet work increased error detection and shared ownership versus opaque agents Sabouri (2026).
• Teams prefer bounded delegation with audit trails, uncertainty flags, and least-privilege permissions (minimum access an agent needs); these controls reduced drift and shifted effort toward higher-level design Choudhuri (2026); Lipsanen (2026).
• In live trading, reliability depended more on system-level controls than the base model; control-harness fixes were associated with more capital actually deployed by agents Barton (2026).

Exposure isn't destiny: adaptation, partial automation, and policy steering

• Large U.S. retraining programs rarely move participants into less automation-exposed occupations; employer-led programs such as apprenticeships do better on occupation switches Jacobs (2026).
• Theory: partial automation is often cost-optimal because accuracy costs rise quickly; more substitution on low-complexity than high-complexity tasks Li (2026).
• Competition can over-automate relative to the social optimum; corrective automation taxes can address this in theory Hemenway Falk (2026).
• A model-based macro-financial stress test suggests rapid AI adoption could depress labor income and demand if substitution outpaces creation of paid demand or redistribution Chen (2026).
• Steering technical progress toward labor-complementary uses raises worker welfare when redistribution is costly; as labor is devalued, steering power wanes and policy shifts toward redistribution and non-monetary well-being Korinek (2026).

New since 2024-25: how the balance of evidence shifted

• 2026 studies strengthen the case for dynamic, task-level exposure: firms are redesigning jobs and hiring in response to generative AI diffusion, and performance gains are broad-based Wang (2026); Mertens (2026).
• New measures sharpen disagreements: learnability and risk-aware scoring materially re-rank exposure and reveal stable structure in what is substitutable Moreira Tomei (2026); Gao (2026); Gao (2026).
• Demographic gaps in access to AI-exposed, higher-paying roles are now documented across countries, elevating equity concerns Jetha (2026); Mishra (2026).

What We Still Don't Know

• Long-run causal impacts on employment and wages from LLM and agent adoption remain largely unmeasured; method reviews cite reliance on static task exposure and limited identification Papagiannaki (2026).
• Measurement disagreement persists: capability overlap, learnability, and risk-aware indices often re-rank occupations, and rankings can flip over time, so there is no consensus exposure map for planning Moreira Tomei (2026); Gao (2026); Gao (2026).
• The pace at which capability gains turn into deployed automation is unclear; benchmarks show low current completion on live tasks alongside steady improvement, making timing hard to forecast Zhang (2026); Mertens (2026).
• Evidence in lower- and middle-income settings is thin and mostly descriptive, limiting generalization beyond high-income contexts Dawoud (2025); Ligot; Peter.
• Organization-wide controls that separate capability from authority at scale have not been tested in field randomized controlled trials; studies using natural experiments show promise but generalizability is open Adams (2026); Barton (2026).