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.

Better measured: large-scale field randomization finds timers speed work by 2.8% but raise refunds by 0.77%. Separate studies find AI can help meet joint quality-and-time targets, while perceived usefulness can decline with repeated use.

Strengthened: LLM outreach raises cash recovery in a debt-collection experiment. Separate market studies link LLM availability to faster information processing and more synchronized trading, while an AI-exposure factor captures residual common risk.

Better measured: data-center expansion is linked to more establishments and property taxes alongside higher electricity prices and water withdrawals, tightening the siting externalities picture.

Extendsestablished

When Algorithms Time Workers: Speed-Quality Tradeoffs in a Randomized Field Experiment

Youngsoo Kim, Dmitry Mitrofanov

randomized field experiment, high evidence

A multi-year platform randomized controlled trial (random assignment to an informational timer) across tens of millions of grocery-picking tasks finds 2.8% faster picking on average and a 0.77% increase in refunds, with larger speed gains on large baskets, early in shifts, and on weekends, and limited response among permanent employees. This extends the standing view by quantifying a live, large-scale speed–quality tradeoff in an operational setting rather than lab or short-run pilots.

So what: if this holds, throughput nudges can externalize costs into refunds and consumer harm, turning small per-task degradations into material back-end expense at scale.

Full numbers

Newestablished

Large Language Models, Debt Renegotiation, and Contract Enforcement

Fangzhou Lu

field RCT, high evidence

In a field randomized design comparing human versus large language model (LLM)-based outreach across telephone and messaging, LLMs increase cash recovery; holding content fixed, human voice delivery outperforms synthetic voice, and autonomous LLM messaging beats human-written messages, with larger gains when cases require integrating hardship and procedural information. This brings causal evidence that general-purpose LLMs change bargaining outcomes and information elicitation in debt renegotiation.

So what: if this holds, credit outcomes and borrower disclosures shift with who generates and delivers messages, raising fairness and oversight risk in high-stakes collections.

Full numbers

Extendssuggestive

The Impact of Data Centers on Local Economies

Diana Van Patten, Fernando Alvarez, David Argente, Joyce Chow

instrumental variables, medium evidence

Using a US county-level long-difference shift–share instrumental-variable design (1995–2023), a 10% rise in annualized data-center revenue is associated with 2.3% more establishments and 14.7% higher property-tax receipts, alongside 0.5% higher retail electricity prices. The water-withdrawal estimate, measured through 2020, is 3.1%. This extends infrastructure externality estimates from case studies toward an economy-wide, quasi-experimental baseline.

So what: if this generalizes, siting wins on jobs and taxes travel with utility-price and water-stress risk that can outlast incentive windows.

Full numbers

Tensionestablished Less is more: Monologue versus Dialogue in AI-enabled Telemarketing Xincheng Ma, Brian Rongqing Han, Dongwon Lee, Tianshu Sun, Junbum Kwon: A 90k-customer randomized controlled trial finds dialogue robocalls did not outperform monologues for coupon conversion and some openers reduced take-up, suggesting format changes can backfire in low-stakes outreach.

Newsuggestive Artificial Intelligence Exposure and Financial Stability: Cross-Industry Common Risk and Connectedness in U.S. Equity Markets Vamsidhar Ambatipudi: An AI-exposure factor appears to be about 82% unspanned by standard factors and leaves residual co-movement after adjustments, pointing to a distinct common AI risk theme.

Extendssuggestive Restart Economics for AI Agents: Cutoffs, Retries, and Cost per Resolved Task James E. Murphy: Replay tests on 500 SWE-bench Verified tasks suggest retries raise resolution rates when verification is automated, but manual review can erase gains, making stopping rules contingent on fallback costs.

Extendssuggestive Large Language Models, Price Discovery, and Synchronized Trading Jiawei Wang: After ChatGPT, earnings information appears to be incorporated faster for hard-to-read firms even as non-retail order flow becomes more synchronized, consistent with shared decoder effects.

Extendssuggestive Large Language Models and Information Frictions in Corporate Bonds Moazzam Khoja: LLM-based call readings add information beyond dictionaries and help predict rating migrations and short-term bond returns when analysts disagree.

Confirmssuggestive Association between adoption of Artificial Intelligence and Financial Performance of the Companies: Evidence from Global Companies Wei Chen: Firm-level AI adoption is associated with modest gains in return on assets, return on equity, and Tobin's Q that accumulate over two years. These are financial-performance measures, not direct estimates of productivity.

Extendsdescriptive Who Captures the Gains? The Global Distribution of AI Value Chain Jobs Yan Liu: AI value-chain postings increased after 2024 and concentrate in rich countries and top-wage occupations, flagging distributional skew in new AI jobs.

Extendssuggestive How Does AI in Managers' Hands Shape Employee Effort and Decision Performance? Yinglin Ruan, Tian Lu, Peijian Song, Yingjie Zhang, Jing Long: In this study, restricting AI recommendations to managers is associated with higher frontline effort and savings, while sharing recommendations with employees is associated with weaker gains, highlighting governance design effects.

Extendsestablished Quality Within a Time Allowance: AI-Assisted Legal Work in a Randomized Experiment Zhiwei Zhuang: A randomized controlled trial with 153 law students finds higher rates of meeting a joint quality-and-time benchmark with AI assistance, though conservative bounds admit zero in worst case.

Extendsestablished Generative AI in Higher Education Staff Work: Perceived versus Measured Effects Manika Garg, Samuel Kernan Freire, Sunčica Bruck, Aslam Tanjung, Theo Bakker: Staff complete tasks faster with higher automated quality ratings, yet perceived usefulness declines over four weeks, revealing a performance–perception gap.

Extendsestablished Personalized Recommendations Without Inducing Congestion: Mitigating Disparities in the NYC High School Match Erica Chiang, Kenny Peng, Rebecca Lichtenstein, Brielle McDaniel, Kristen O'Neil, Deja Thomas, Lianna Wright, Jon Kleinberg, Eva Tardos, Nikhil Garg: An equilibrium-aware recommender appears to increase ranking of recommended nearby high-performing programs without appearing to induce rejections, suggesting congestion-aware design can avoid crowding harms.

Tensionframework Common AI Information and Capital Allocation Haoning Sun: Model and patterns suggest shared AI errors can contaminate prices and produce a J-shaped accuracy–productivity link, cautioning on common-signal exposure.

Extendsestablished Structured Pre-Generation Elicitation Versus Single-Shot Prompting In AI-Assisted Enterprise Decision-Making: A Randomised Online Experiment William Scott-Jackson: A metacognitive scaffold raises rated deliverable quality by ~32% and reduces off-task outputs at a small time cost.

Extendssuggestive Evaluating AI pilot zones’ impact on enterprise digital transformation: a quasi-natural experiment Li Mingyue, Sun Xianmin: Staggered difference-in-differences evidence links China's AI pilot zones to higher firm digital transformation via talent and finance channels, with regional heterogeneity.

Newestablished Do Tool Calls Execute as Intended? Measuring and Repairing Intent-Execution Correspondence in LLM Agents Boyang Yang, Zhenhao Li, Ziyao Yang, Kanghui Jia, Xin Yin, Mingmou Liu, Haoye Tian: Production harnesses can mutate tool calls silently; an intent–execution correspondence (IEC) protocol plus IntAct repairs recover ~79% of changed-call failures.

Extendsestablished When Recruiters Meet Algorithms: Advice Taking and Bias in AI‐Assisted Recruitment Alain Lacroux, Christelle Martin Lacroux: The study finds recruiters behaviorally rely more on algorithmic advice even with lower stated trust, and inaccurate advice is associated with larger decision losses than the gains from accurate advice.

Newdescriptive Evidentiary Integrity in Automated Regulatory Auditing: A Deterministic Tool-Mediated Architecture for Cross-Jurisdictional Privacy Compliance Verification Septian Adi Nugraha, Muhamad Erza Aminanto: In tests, a gated, deterministic audit workflow yields 100% quote provenance and removes hallucinated citations while matching human inter-rater agreement ranges.

Extendsdescriptive Artificial Intelligence as an Industrial System in 2026: Demand, Infrastructure Capital, Supply Constraints, and Regulatory Divergence, with Implications for B2B Buyers, Suppliers, and Investors Liza Phukan: Filings show concentrated capital expenditure (capex), trillion-scale backlogs, and binding high-bandwidth memory (HBM), foundry, and grid bottlenecks alongside regulatory divergence.

Newsuggestive Do Markets Dream of Artificial Intelligence? Macroeconomic Effects and International Spillovers of AI News Shocks Maria Sole Pagliari: AI news shocks are followed by an investment-led US expansion with producer prices rising before consumer prices and distinct euro-area spillovers.

Extendsframework Competing Multi-Product Firms Using Single-Product Pricing Algorithms Karsten Hansen, Kohei Hayashida, Kanishka Misra, Mallesh Pai: Simulations suggest cross-product linkages (constant markups, joint-pricing agents) can sustain supra-competitive prices even when per-product learners do not.

Newsuggestive Tariffs, Export Controls, and the Reshoring of AI Compute Yogender .: Policy carve-outs and export controls are associated with larger capex shifts toward exempted jurisdictions, especially Taiwan, in 2026 guidance.

Extendsdescriptive Compute, Power, and Sovereignty: The Political Economy of AI Critical Infrastructure Zola Valashiya: Contracts reallocate demand risk across developers, utilities, and governments, trading cheaper financing against exposure for ratepayers and taxpayers.

Tensionframework AI Displacement without Immiseration: A Ricardian Perspective on Jagged AI Growth Ahmad Lashkaripour: A calibrated Ricardian model implies jagged AI progress can displace tasks while lifting average real wages, with outcomes sensitive to frontier "jaggedness".

Extendssuggestive Agentic Strategists Arnab Choudhury, Bruce Kogut, Patryk Perkowski: In these experiments, personalization and domain grounding are each associated with higher profits for LLM strategists, but combining them appears to have negative complementarity at the aggregate profitability level.

Operational speed–quality frontier: a platform-scale randomized controlled trial measures faster picking and higher refunds from a simple timer. Separate AI-assistance studies report gains on joint quality-and-time benchmarks and automated quality scores; in the higher-education staff pilot, perceived usefulness declined over four weeks. The settings and outcome measures differ, so these results do not establish a single general speed–quality effect.

Who delivers and who sees AI advice: debt-collection experiments show human delivery of LLM-generated content outperforms synthetic voice, and manager-only recommendations are associated with higher frontline effort and savings than shared recommendations, sharpening the governance design lever. This is an editorial juxtaposition across domains, but it points to delivery-channel and disclosure as first-order mediators of realized gains.

Common AI information in finance: studies report faster incorporation of hard-to-read disclosures, more synchronized order flow, and a residual AI-exposure factor. LLM readings of earnings calls also contain information not already reflected in bond spreads. These findings motivate tests of the model-based concern that shared AI errors contaminate prices; they do not by themselves establish that error channel.

Compute infrastructure externalities: an instrumental-variables design links local gains in establishments and property taxes to data-center growth while quantifying higher electricity prices and water withdrawals, moving beyond anecdote to long-difference measurement. This lands alongside descriptive evidence of capital concentration and supply bottlenecks that constrain where and how fast these effects propagate.

Throughput nudges versus downstream quality

Dialogue formats in outreach versus LLM messaging in collections

Common AI signals: efficiency gains versus synchronized risk

Capability coverage versus realized exposure constraints

Agent retries and verification: savings versus null deployment value

Large-scale randomized field intervention at operational scale (timers on tens of millions of tasks), When Algorithms Time Workers: Speed-Quality Tradeoffs in a Randomized Field Experiment: brings external validity and heterogeneity mapping to speed–quality tradeoffs that are often studied in labs.

Intent–Execution Correspondence protocol with IntAct repairs, Do Tool Calls Execute as Intended?: isolates the first divergent hop in agent tool calls across real harnesses and recovers most changed-call failures, a prerequisite for trustworthy agent deployment.

Task-to-wage-bill capability mapping, How Much Work Can AI Agents Do?: ties external evaluation success curves to O*NET (Occupational Information Network) task times and wages, offering an updating, economically weighted capability benchmark rather than coarse occupation labels.

Browse the full paper archive →

Website · LinkedIn