AI reshapes jobs and markets: it automates routine tasks while creating complementary roles, boosts productivity for firms that invest in complementary assets, and concentrates market power — producing net employment gains that are unevenly distributed and policy-sensitive.

Main Finding

AI is reshaping the digital economy along three linked dimensions: (1) labor — a dual effect of displacing routine tasks while creating AI‑complementary roles that produces uneven employment outcomes; (2) productivity — sizable firm‑level gains that depend on complementary investments and exhibit a J‑curve delay before aggregate effects appear; and (3) market structure — stronger winner‑takes‑most dynamics driven by data feedback loops, raising concentration, entry barriers, and novel competition risks (e.g., killer acquisitions, algorithmic collusion). Existing policy frameworks for labor, competition, and data governance are poorly aligned with these dynamics.

Key Points

• Employment

  • Task‑based analyses revise early high displacement estimates downward (Frey & Osborne ~47% vs. Arntz et al. ~9%) by recognizing within‑job task heterogeneity.
  • Acemoglu & Restrepo’s task model distinguishes automation (substitution) from augmentation (new task creation); current evidence suggests automation often outpaces new task creation, pressuring middle‑skill jobs and raising inequality.
  • Parallel evidence finds net job creation in AI‑related roles (WEF projections), but distributional frictions and measurement limits (survey vs. econometric methods) matter.
  • Labor‑market polarization continues to be important, with some indications AI is beginning to affect previously non‑routine cognitive occupations.

• Productivity

  • Firm‑level evidence is robust: AI‑intensive firms show materially higher TFP (reported ~20–30% above sector averages in some European studies) and higher sales per worker in US data (~5% in some estimates).
  • Productivity gains are contingent on complementary intangible investments (skills, reorganization, data infrastructure) and often show a lag (the productivity J‑curve).
  • Sectoral concentration of gains: manufacturing, finance, healthcare, logistics/retail show particularly strong AI impacts.

• Market Structure & Competition

  • Data acts as a cumulative, increasing‑returns input: larger datasets → better models → more users → more data (self‑reinforcing advantages).
  • Evidence of rising market concentration and “superstar” firms capturing disproportionate shares of profits and market value.
  • Strategic behaviors of incumbents (e.g., killer acquisitions) and the potential for algorithmic collusion pose challenges for price‑based antitrust frameworks, particularly in zero‑price markets.

• Policy & Governance

  • Current labor, antitrust, and data rules are often misaligned with AI‑driven realities (e.g., passive unemployment insurance, price‑centric antitrust, incentives for data hoarding).
  • Tension between rapid capability development (concentration aids frontier R&D) and the need for competitive plurality and equitable distribution of gains.

Data & Methods

• Review design: Systematic Literature Review (SLR) following PRISMA.
• Search scope: Web of Science, Scopus, SSRN, Google Scholar; keywords combining “artificial intelligence”, “machine learning”, “digital economy”, “employment”, “productivity”, “market structure”, etc.
• Time window: January 2015 – December 2025 (includes emergence of LLMs/generative AI).
• Screening: initial n = 1,847 records → de-duplicated (432 removed) → 1,415 screened → 214 full texts assessed → 78 studies retained.
• Inclusion criteria: peer‑reviewed articles and major institutional reports (IMF/OECD/WEF/WB), English language, empirical/theoretical/systematic reviews with generalizability.
• Coding and synthesis: studies coded by thematic domain (employment, productivity, market structure), methodology (quantitative, qualitative, simulations), geographic scope, and level of analysis (firm, sector, country).
• Evidence types in the corpus: firm‑level panel regressions, TFP analyses, matched employer‑employee data, sectoral case studies, simulations of algorithmic pricing, policy/legal analyses, and institutional reports.
• Noted limitations: English‑only selection; exclusion of grey literature and non‑peer opinion pieces; heterogeneity in methods and counterfactuals across studies.

Implications for AI Economics

• Policy design

  • Labor policy: scale up active labor market policies (retraining, reskilling, job search assistance), rethink unemployment insurance for faster transitions, and target support for middle‑skill workers at risk of displacement.
  • Competition policy: move beyond price‑only consumer welfare metrics to account for innovation harms, privacy erosion, and reduced plurality; strengthen merger review (including low‑turnover "killer" deals) and consider ex‑ante obligations for gatekeepers.
  • Data governance: incentivize competitive data access/sharing (while protecting privacy), reassess trade secrecy rules that encourage data hoarding, and align intellectual property rules with pro‑competitive innovation.
  • Regulatory trade‑offs: explicitly manage the innovation vs. distribution trade‑off — preserve incentives for frontier R&D while mitigating entrenchment and unequal gains.

• Research priorities

  • Longitudinal microdata tracking workers and firms through AI adoption to measure displacement, reallocation, and wage effects over time.
  • Comparative institutional studies on how regulation and governance shape distributional outcomes across countries.
  • Focused empirical work on generative models and LLMs to capture economic dynamics distinct from earlier AI waves (e.g., effects on cognitive tasks, content markets).
  • Analyses of global value chains and cross‑border spillovers of AI adoption and data flows.
  • New empirical approaches to detect algorithmic collusion, killer acquisitions, and non‑price forms of market harm in zero‑price markets.

• Measurement & methodology recommendations

  • Disaggregate occupations into tasks when estimating automation risk.
  • Allow sufficient post‑adoption windows to capture the J‑curve in productivity studies.
  • Combine employer surveys with matched administrative and panel data to improve projections of job creation/destruction.

Limitations of this review: language and publication‑type restrictions, heterogeneity in methods across studies, and the rapidly evolving character of AI (post‑2025 developments may alter dynamics).