Digests

This weekly digest tracks what is NEW or CHANGED in AI-economics research. For the cumulative state of evidence on any topic, see the /syntheses pages. A single study rarely overturns a body of evidence.

The Delta

• Strengthened: system-level design and verification, not just model choice, drive measurable productivity outcomes in deployment-scale tests.
• Better measured: a platform-scale RCT shows higher recommender quality reduces consumption concentration and broadens demand to the middle tail.
• Newly observed: tension between demand diffusion on consumer platforms and reviews warning labor-market gains concentrate among AI-skilled workers.

What Moved & What Held

Coming in, the standing view is that AI can raise productivity when paired with complementary investments (skills, process, governance), that platform algorithms shape demand patterns, and that distributional gains are uneven without those complements.

This week adds two large, clean experiments and a careful agent-harness study showing that verification steps, harness rules, and surfacing choices shift engagement, conversion, and task success without changing the base model, and it better measures a decline in superstar concentration from improved recommendations in a very large RCT; at the same time, several reviews reiterate that labor-market gains likely concentrate without policy and skill complements. Still holds this week: the macro picture of uneven gains and the importance of institutional and managerial complements; the news is sharper measurement of where system design moves realized outcomes.

Top Papers

• Tension · established Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix — Guy Aridor, Winston Chou, Nathan Kallus, Antoine Scheid, Allen Tren, Kevin Zielincki
  • In a randomized controlled trial on 8.5M users, better recommendations raise engagement and reduce title concentration (HHI down 5.7%), shifting viewing toward a broader middle tail with little change in the extreme long tail; this pushes against the common view that recommender improvements concentrate demand.
  • So what: Risk of misdiagnosing market power if concentration metrics are interpreted without testing recommender effects; monitor producer-side concentration separately.
  • Full numbers
• Extends · established TRACE: Agentic Catalog Enrichment with Multi-source Evidence Grounding — Rohan Kumar, Steven Xu, Kyle MacDonald, Matthew Long, Bernice Chow, Mac VanRenterghem, Sudeep Das
  • An agentic pipeline with a verify-before-write judge achieves a 98.4% human-confirmed PASS rate on attribute extraction and, in an online A/B test, surfacing enriched attributes lifts checkout conversion by about 0.48%, showing system design plus verification moves commercial metrics.
  • So what: Risk of eroding trust and conversion if noisy attributes are surfaced without verification and governance gates.
  • Full numbers
• New · suggestive Same Model, Different Harness: Different Coding-Agent Results — Sydney Lewis
  • Holding the model fixed, a shortened-history, interventionist harness increases complete solutions on a 169-task SWE-bench Verified cohort (43 to 72 tasks) and raises partial-repair rates on FeatureBench under context pressure, with effects largest at mid-size token windows.
  • So what: Risk of misallocating spend to bigger models while leaving cheaper harness-level gains untapped under realistic context pressure.
  • Full numbers

Also Notable

• Extends · descriptive Board Oversight Mechanisms in Financial Distress Decision-Making: A PRISMA-Based Systematic Review — Andrei Popescu, Elena Ionescu — Board oversight in crises works via interacting structural, procedural, and behavioral mechanisms, suggesting single-lever reforms underperform.
• New · descriptive Data-driven comparison of airline passenger flight supply using optimal transport theory — Qian Liu, Paul Rochet, Chantal Roucolle — An interpretable optimal-transport distance tracks multidimensional carrier differences and shows heterogeneous post-pandemic adjustments among 32 European airlines (2016–2025).
• New · descriptive LLMs Can Design Near-Optimal OR Algorithms — Jackie Baek — Frontier LLMs generate solutions and reusable algorithms that match or beat state-of-the-art across multiple OR benchmarks in this sample.
• Extends · suggestive Fintech credit and corporate cash holdings around the world — Manoja Behera, Jitendra Mahakud — Greater fintech credit availability is associated with lower corporate cash buffers, especially for more constrained firms and in deeper financial systems.
• Tension · suggestive How Is AI Transforming the Task Characteristics and the Experience of Vulnerable and Minority Employees in the Hospitality Sector? — Deepak Bangwal, Shobha Maindola, Rupesh Kumar, Pankaj Chamola, Sarbjit Singh Oberoi — In this mixed-method sample, task–technology alignment with mechanical/analytical/intuitive AI is associated with improved worker experience for vulnerable hotel employees, tempering broad polarization narratives.
• Extends · descriptive ARTIFICIAL INTELLIGENCE IN DIGITAL BANKING: APPLICATIONS AND IMPLICATIONS FOR LABOR TRANSFORMATION — Nguyen Thi Hang, Huynh Thi Huong Thao — Banking use cases cluster around ML, chatbots, and RPA, and the literature points to rising demand for hybrid finance–AI roles and skill polarization.
• Confirms · framework Technological Polarization and Unequal Growth in the Era of Generative Artificial Intelligence — Kyra Mahindru — A PRISMA-style review synthesizes evidence that generative and agentic AI amplify wage polarization absent policy, reinforcing distributional risk concerns.
• Extends · suggestive Forecasting Fashion Sales With Social Media Signals: Insights From Amazon, Twitter, and Google Trends — Olena Rudna, Alex Rudniy, Arim Park — Combining Twitter attribute counts with Google Trends in vector error-correction models improves out-of-sample demand forecasts across 77 apparel attribute series over univariate baselines.
• Extends · suggestive Explainable Machine Learning-Driven Supplier Risk Prediction Using ERP Procurement Data and Power BI Decision Dashboards: Evidence from Manufacturing Supply Chains — Gautam Kumar, Raviteja Narra, Srilatha Batchu, Manoj Kumar — On a 100-supplier ERP panel, XGBoost outperforms a traditional scorecard (ROC-AUC ~0.93) and SHAP explanations support dashboarded operational insights.
• Extends · descriptive Enterprise Architecture for Digital Transformation in a Fragile State: Evidence from a Mixed-Methods Case Study of Afghanistan — Bilal Himmat, Amir Kror Shahidzay, Mohammad Rashid Sapai — Formal enterprise architecture adoption is low (about 16%) yet correlates with stronger standardization, innovation enablement, and risk reduction in this context.
• Extends · suggestive Can MD&A tone convey information? Evidence from mismatched firms under unified fiscal years — Mu Xing, Hong-Mei Zhang, Dong Chen — Chinese firms with misaligned business and reporting cycles adjust MD&A tone in ways associated with better subsequent performance, lower stock synchronicity, and higher valuations, especially with strong internal controls.
• Extends · descriptive Precision psychiatry in clinical practice: What is clinically actionable, what is promising, and what remains experimental? — Julio Torales, Marcelo O’Higgins, Iván Barrios, Victor-Guillermo Sequera, Gladys Mercedes Estigarribia Sanabria, João Mauricio Castaldelli-Maia, Antonio Ventriglio — Few precision-psychiatry approaches are routine-actionable today; most biomarker and AI methods remain promising but experimental.
• Extends · suggestive Determinants of researchers’ copyright licensing behavior in open access — Byoung-Goon An — Funder mandates and international co-authorship correlate with permissive CC licenses, while high-prestige venues and medical fields lean restrictive.
• Extends · descriptive Institutional Drivers and Governance Mechanisms of Debt Accumulation in Local Authorities: A Critical Review of the Literature with Reference to Zambia — Mulenga Kelvin Mutale, Norman Kachamba, Nsama Musawa, Daniel Chisanga — The literature points to institutions and governance over raw resources as drivers of subnational debt variation and flags a lack of council-level causal studies in Zambia.
• Extends · descriptive Children in the Digital Age: Health, Behavioral, Learning Impacts, and Management Strategies — Rishikesh Upadhyay — Synthesis argues digital tools can personalize learning but also create health and attention risks when unmanaged, with equity gaps shaping outcomes.
• Extends · descriptive Supply Chain Resilience Enablers: A Review on the Last-15-Years Research — Modestus Pinto, Yosephine Suharyanti, Slamet Wigati — Research emphasis has shifted from visibility to agility, analytics, and sustainability as dominant resilience enablers, varying by sector.
• Extends · framework The UK–Israel FTA in a politically dynamic era: economic rationale, moral diplomacy, and trade policy — Erez Cohen, Daniel Schiffman — Process tracing links shared role conceptions to deep digital-services talks through 2024 and a May 2025 suspension following UK policy reorientation.
• Extends · framework EXPRESS: A Review of Upper Echelons Theory in Hospitality: General Managers as an Overlooked Echelon — Mahsa Javdanmehr, Stephen X. Zhang, Kim Huynh, Rob LAW — Proposes a hospitality-tailored UET where GMs’ blended strategic–operational roles and tenure dynamics shape managerial influence and AI adoption heterogeneity.
• Tension · suggestive A Machine Learning Framework for Price Estimation in Air Force Acquisition — Kefallinos, Paola, O'brien, Cuyler — XGBoost explains much of log-price variation (test R2 ~0.71) but median absolute percent error around 51% suggests operational accuracy remains a hurdle.
• Confirms · descriptive Impact of Artificial Intelligence on the Global Economy — Niharika Sharma Mahajan — Review of international studies reiterates sizable productivity potential from AI with uneven realized gains tied to infrastructure, skills, and institutions.
• Extends · descriptive Artificial Intelligence and Labour Market Transformation: A Comparative Analysis of Occupation Exposure to AI in Selected Economic Sectors — Suhan Deepak Chandwani — Descriptive comparisons show routine tasks are more exposed while demand rises for hybrid analytic, creative, and interpersonal skills; net employment effects vary by context.
• Extends · framework Synergy Paradigm: Reimagining Innovation through Interdisciplinary Collaboration — Iskilu Abayomi Akintola — Conceptual chapter argues AI-driven task reallocation widens skill premia without reskilling and supportive institutions; no new estimates provided.

What Moved

• System design over model specs: Two deployment-scale experiments and a careful harness comparison moved confidence that verification, context management, and surfacing decisions materially shift outcomes, beyond base-model capability; this tightens the earlier, more general claim that “complements matter” by quantifying effects on conversion, engagement, and task success. The harness paper’s paired tests also suggest diminishing marginal returns from giant context windows relative to better orchestration in constrained regimes, an editorial inference bridging benchmark and field evidence.
• Concentration dynamics on platforms: The Netflix RCT better measures a reduction in consumption concentration and a shift toward the middle tail when recommendations improve, which runs counter to simple “algorithms amplify superstars” priors and sets up a domain-specific contrast with labor-market polarization findings. This points to layered markets where AI can diffuse consumer demand yet still concentrate producer returns or wages.

Contested & Watch

• Do better recommenders reduce or increase concentration?
  • Finding: In a platform-scale RCT on 8.5M users, improved recommendations lower title concentration (HHI down 5.7%) and broaden the middle tail.
  • Watch: Additional platform-level randomized experiments that report concentration metrics and supply responses over longer horizons.
• Where are the high-return margins: system engineering or base-model upgrades?
  • Finding: Holding the model fixed, harness and intervention rules substantially lift coding-agent repair and solution rates under context pressure.
  • Standing evidence: Capability benchmarks show strong base-model performance in some domains (e.g., OR tasks) even with light prompting.
  • Watch: Head-to-head A/Bs comparing model swaps versus harness changes on the same tasks and budgets, with cost-adjusted outcome deltas.
• Does AI widen wage polarization economy-wide even as some local deployments aid vulnerable workers?
  • Finding: Reviews synthesize that generative and agentic AI amplify wage polarization without policy; a hospitality study finds aligned AI can improve vulnerable workers’ experience in this sample.
  • Standing evidence: Multiple reviews and sector syntheses lean toward uneven gains and rising skill premia, with limited causal estimates across labor markets.
  • Watch: Sector-specific causal studies that jointly estimate productivity and distributional effects, including wage and task composition changes.
• Can procurement-ready ML hit accuracy thresholds for price estimation?
  • Finding: In Air Force acquisition data, XGBoost attains test R2 around 0.71 but median absolute percent error near 51%, limiting operational usefulness.
  • Standing evidence: Many applications report strong explanatory power of ML on logs but face scale-error constraints in high-stakes settings.
  • Watch: Prospective evaluations with procurement-ready loss functions, human-in-the-loop calibration, and error-cost reporting.

Methods Spotlight

• Optimal-transport distances for multi-dimensional firm comparisons — Data-driven comparison of airline passenger flight supply using optimal transport theory: interpretable OT metrics let analysts compare seasonality, geography, and route-length structure in one coherent framework.
• Verify-before-write agentic pipelines — TRACE: Agentic Catalog Enrichment with Multi-source Evidence Grounding: tying multimodal evidence to an automated judge yields high-accuracy extractions and measurable conversion gains in an online A/B.
• Paired harness experiments for agent evaluation — Same Model, Different Harness: Different Coding-Agent Results: holding model, tasks, and budgets fixed isolates orchestration effects and helps separate system design from model capability.