SCTY
Back

December 10, 2025

AI at Work: The 2025 Evidence

The debate about whether AI changes work is over. Seven major 2025 studies—from the ILO, WEF, Anthropic, OpenAI, and academic researchers—now converge on the same core findings:

  1. Productivity gains are real and large. Controlled experiments show +15-30% in cognitive service work. Not vibes. Measured output.

  2. AI compresses performance. The biggest gains go to lower-skill workers. Top performers see modest improvements or sometimes get worse when they over-rely on the model.

  3. Automation and augmentation are both happening—but in different jobs. Automation-prone roles (clerical, structured cognitive) are shrinking. Augmentation-prone roles (mixed judgment, social, technical) are growing and demanding more skills.

  4. Adoption is deeply unequal. By geography (Israel 7× vs Nigeria 0.2× per capita), by gender (women overrepresented in high-exposure clerical roles), and by firm (frontier companies send 7× more messages per seat than median).

  5. The frontier moved past chat. Agents, API automation, embedded workflows—that's where the action is now. The gap isn't "uses AI vs doesn't" anymore. It's "integrated deep vs dabbling."

The question is no longer whether AI affects work. It's who controls the gradient between augmentation and displacement—by occupation, by firm, by country.

The Eight Findings

1. AI is already moving real productivity

The Brynjolfsson RCT: 5,172 customer support agents, staggered access to a GPT-3-based assistant. Result: +15% productivity on average. Lower-skill agents: +30%. Top agents: small gains, sometimes quality declines from over-reliance.

OpenAI enterprise surveys align: 75% of workers report improved speed or quality. Typical user saves 40-60 minutes per active day.

This is industrial-revolution-scale at the micro level.

2. Automation risk is real but concentrated

ILO's 2025 task exposure analysis: only 3.3% of global employment sits in the highest automation-risk bucket. Another ~20% faces transformation, not outright redundancy.

The clear loser: clerical and secretarial work. Data entry, bank tellers, cashiers, ticket clerks, postal workers, admin assistants, legal secretaries, bookkeeping.

3. Augmentation shows up in job postings

Chen et al. analyzed the near-universe of US job postings against ChatGPT's November 2022 launch:

  • Automation-prone occupations: postings down 17%
  • Augmentation-prone occupations: postings up 22%, with higher skill requirements

Firms are learning to put AI inside workflows, not just replace headcount.

4. Inequality is structural

Anthropic's usage index: per-capita Claude usage in Israel is 7× population share. Singapore 4.6×. US 3.6×. India 0.27×. Nigeria 0.2×.

High-exposure roles skew female and clerical. Productivity gains help lower-skill workers within roles—but nothing guarantees they capture the surplus.

5. Skills are shifting toward cognitive + social + AI-literate

WEF's fastest-growing skills: AI/big data, cybersecurity, tech literacy—but also creative thinking, resilience, curiosity, leadership, talent management.

Manual/physical skills decline overall. Pure clerical numeracy and writing become AI-assist territory. High-end manual work persists and specializes.

6. The frontier is agents, not chat

Perplexity: agentic queries in the hundreds of millions. "Productivity & Workflow" + "Learning & Research" = 57% of usage.

OpenAI: ~20% of enterprise messages now go through custom GPTs—repeatable workflows, not ad-hoc prompting.

Anthropic: API usage is 77% automation patterns. The shift from augmentation to automation is already measurable.

7. Most organizations under-use what they bought

Within enterprises, many active users have never touched data analysis, reasoning, or search tools. Frontier firms send 7× more messages per seat than median.

The constraint isn't model capability. It's workflow design, data access, evaluation, culture.

8. The research questions changed

No longer "does AI affect work?" Now:

  • How persistent are learning gains as models and orgs keep changing?
  • How fast does augmentation become automation in specific task clusters?
  • How do wage structures adapt when bottom-quartile workers jump in performance?
  • How do low-adoption countries plug in before standards harden without them?

The Supporting Data

Productivity Evidence

The Brynjolfsson RCT is the cleanest causal study: 5,172 customer support agents with staggered access to a GPT-3-based assistant. Average productivity up 15%. Lower-skill agents up 30%. During AI outages, productivity stayed above pre-AI baseline—workers were learning from the model, not just using it.

Side effects: customers became more polite, escalations dropped, new-worker attrition fell. AI improved English fluency and made low-skill agents' language converge toward high-skill style.

Job Postings Analysis

Chen et al. scored US occupations on automation vs augmentation potential using O*NET tasks + GPT-4o, then tracked job postings from 2019 through Q2 2024 against ChatGPT's November 2022 launch.

The data: automation-prone jobs saw postings drop 17% and skill requirements fall 20-25%. Augmentation-prone jobs saw postings rise 22% with skill requirements up 15%. AI is simultaneously destroying and creating jobs—along different occupational bundles.

WEF Projections to 2030

Based on 1,000+ employers covering 14M workers:

  • Jobs created: 170M (14%)
  • Jobs destroyed: 92M (8%)
  • Net: +78M (7%)

The only clearly net-negative driver: robots and autonomous systems (−4.8M). AI and information-processing tech is net positive but with high churn both ways.

Growing: Big Data, AI/ML specialists, software devs, cybersecurity, renewable energy engineers, nurses, teachers, project managers

Declining: Data entry, bank tellers, cashiers, postal workers, admin assistants, legal secretaries, bookkeeping

Skills Shift

WEF estimates 39% of worker core skills will change by 2030. The fastest-growing: AI/big data, cybersecurity, tech literacy, creative thinking, resilience, curiosity, leadership. The only clear decline: manual dexterity and endurance.

The discriminators between growing and declining jobs: resilience, resource management, quality control, programming, tech literacy, analytical thinking.

Geographic Distribution

Anthropic's AI Usage Index: Israel at 7× per-capita share, Singapore 4.6×, US 3.6×. India at 0.27×, Nigeria 0.2×. Usage rises 0.7% for each 1% increase in GDP per working-age capita.

Perplexity: 28% of agent adopters and 30% of agentic queries come from digital-tech occupations. The frontier is narrow and rich.

Organizational Reality

WEF: biggest barrier to transformation is skills gaps (63%), then culture/resistance (46%), then regulation (39%). 85% plan to prioritize upskilling; 41% plan to reduce staff where skills become obsolete.

OpenAI's enterprise data: frontier firms send 7× more messages per seat than median. The constraint isn't model capability—it's workflow design, data access, and culture.

Sources

  1. ILO – Generative AI and Jobs: A 2025 Update (May 2025)
  2. World Economic Forum – Future of Jobs Report 2025 (January 2025)
  3. Brynjolfsson, Li, Raymond – Generative AI at Work (2024)
  4. Chen, Srinivasan, Zakerinia – Displacement or Complementarity? The Labor Market Impact of Generative AI (July 2025)
  5. Salari et al. – Impacts of Generative AI on the Future of the Labour Market: A Systematic Review (March 2025)
  6. OpenAI – The State of Enterprise AI (December 2025)
  7. Anthropic – The Anthropic Economic Index Report (September 2025)
  8. Perplexity – The Adoption and Usage of AI Agents: Early Evidence (December 2025)