A conservative, sector-level model finds AI could add around $1.06 trillion a year to US GDP by 2036 (3.6% of 2024 GDP), mostly driven by services and finance; even under bearish assumptions direct productivity gains approach $800–940 billion and infrastructure payback occurs before 2036.

Main Finding

A bottom-up, sectoral model estimates that AI could raise US GDP by roughly $1,057 billion per year (3.6% of 2024 GDP) by 2036 in the paper’s base case. Alternative scenarios produce a range from $796bn (bear) to $1,368bn (bull), with an agentic / full-workflow-transformation scenario up to $2,521bn. Even under conservative assumptions the model finds direct AI-driven productivity alone of ≈$940bn/year by 2036, and cumulative net GDP gains exceed cumulative AI infrastructure investment before 2036 (base-case payback in 2033).

Key Points

• Modeling approach: bottom-up, sectoral across 21 NAICS industries, restricting productivity gains to the labor-generated portion of gross value added (reduces the naive addressable base by ≈72% versus applying aggregate multipliers to total output).
• Core per-sector formula multiplies six factors: base GDP, labor share, AI coverage, productivity gain percentage, adjusted adoption rate, and a skill-weighted capture rate.
• Annual overlays: demand expansion (adds positive effect), a robotics “unlock” for physical sectors starting in 2030, and an electricity drag subtracted yearly.
• AI coverage: taken from Massenkoff & McCrory (2026) (LLM task coverage across 22 SOC groups), mapped to NAICS via employment-weighted averages using BLS OES 2023.
• Productivity anchors: sector percentages grounded in published evidence (e.g., GitHub Copilot RCT, JPMorgan disclosures, Cognizant internal research).
• Frictions/adoption: regulatory and labor frictions scored using actual compliance regimes (Basel III, FDA AI guidance, HIPAA) plus BLS union density; applied as a haircut to adoption with an S-curve ramp and a permanent friction floor.
• Scenarios: four variants change capture rates and friction assumptions (bear, base, bull, agentic).
• Sectoral concentration: Professional & Technical Services, Information, and Finance & Insurance account for ~86% of the base-case direct contribution.
• Distributional result: lower-skill workers show larger individual productivity gains in some firm-level studies, but the model’s skill-weighted capture mechanism means higher individual gains do not automatically produce proportionate GDP capture.
• Conservatism: paper documents 14 deliberate conservative assumptions (e.g., frozen base GDP, no AI-on-AI compounding, permanent friction floor, conservative capture rates), so estimates likely understate upside.
• Sensitivity: aggregate results most sensitive to the high-skill capture rate and the pace at which friction decays.
• Transparency: full model (11 analytical tabs) is publicly available for replication and sensitivity testing.

Data & Methods

• Coverage mapping:
  • Source: Massenkoff & McCrory (2026) LLM task-coverage scores by Standard Occupational Classification (22 SOC groups).
  • Mapping: SOC → NAICS via employment-weighted averages using BLS Occupational Employment and Wage Statistics (2023).
• Productivity gains: sector-specific percentages anchored to empirical and disclosed sources:
  • Kalliamvakou et al. (2023) Copilot randomized controlled trial (developer productivity).
  • Firm disclosures (e.g., JPMorgan CEO insights).
  • Industry research (Cognizant New Work New World 2026).
• Friction/adoption modeling:
  • Regulatory frameworks (Basel III, FDA AI guidance, HIPAA) and BLS union density inform sector friction scores.
  • Friction applied via S-curve adoption ramp with permanent floor; scenarios vary friction decay and capture rates.
• Model mechanics:
  • Six-factor multiplicative sector formula applied annually to 2036.
  • Additional yearly layers: demand expansion, robotics unlock starting 2030 (for physical sectors), electricity drag subtraction.
  • Four scenarios produced by adjusting capture/friction parameters (bear/base/bull/agentic).
• Validation and robustness:
  • Sensitivity analysis across key parameters (notably high-skill capture and friction decay).
  • Conservatively biased parameter choices and explicit documentation of 14 conservative assumptions.
  • Public release of model for independent replication.

Implications for AI Economics

• Methodological: A sectoral, labor-share-constrained bottom-up approach yields materially smaller “addressable” output than naive top-down multipliers, offering a more granular and defensible estimate of direct productivity effects.
• Policy:
  • Regulatory and labor frictions materially affect adoption and near-term GDP capture; policy choices that lower unnecessary compliance friction or clarify AI regulation can accelerate realized gains.
  • Investments in worker upskilling and mechanisms that increase high-skill capture of productivity gains are high-leverage levers for aggregate outcomes.
• Investment and returns:
  • Model implies AI infrastructure investment is paid back before 2036 under all scenarios (base-case payback 2033), supporting continued private and public capital deployment.
  • Sectoral concentration suggests concentrated returns for finance, information, and professional services—implications for sector-specific investment strategies and capital allocation.
• Distributional:
  • The finding that lower-skill workers may experience higher individual productivity increases but not proportionate GDP capture highlights potential wage and employment shifts and underscores the need for policies addressing distributional consequences.
• Research:
  • Key uncertainties (high-skill capture, friction decay, AI-on-AI effects, robotics timing/scale) should be prioritized in empirical work; relaxing conservative assumptions could substantially raise estimated gains.
  • The publicly available model provides a platform for independent sensitivity testing, extensions (e.g., macro feedbacks, dynamic GDP growth), and more detailed sectoral or regional analyses.

Limitations worth noting: intentionally conservative design (frozen base GDP, no AI-on-AI compounding), mapping and productivity-anchor uncertainties, and simplified robotics/electricity treatments — all mean the results are directional and contingent on assumptions.