The Commonplace · Sep 28
|
The Commonplace
Weekly Research Digest · September 28, 2026
|
From Alex
Over on the New Future of Work Reader, we recently published The Half-Finished Workplace, a researcher take from Nabiha Jiwani, JP Jaramillo, Jaime Teevan and Alexia Cambon. A cohort of NYU Stern MBA students spent a week with an unfinished Microsoft AI prototype, and the piece asks who should be in the room when the future of work gets built.
One more I'd add to this week's reading, though it didn't make the automated selection: “How Does AI Change Labor Demand?”, new cross-country work posted on SSRN.
Correction: the mid-September issue listed the wrong authors for “Sustainability accounting and ESG reporting quality: a systematic literature review and future directions.” The authors are Wasim K. Al-Shattarat, Basiem K. Al-Shattarat and Ruba S. Hamed, and the web version is now corrected.
The Delta
What Moved & What Held
Coming in, the standing view was that AI often boosts near-term output in real firms, with stronger gains for lower-skill workers and mixed effects for experts, benefits depend on information and organizational complements, and headline machine learning (ML) metrics or vendor claims can diverge from decision value.
This week adds three pieces: first, agentic delegation cleared a high bar in a live commercial setting (airline pricing RCT, +5.8% revenue per available seat kilometer) and complements prior assistance results; second, a three-month legal-services RCT documents lasting, unaided quality gains for seniors but not juniors, refining expectations about learning; third, an RCT in public administration suggests algorithmic assignment can erode human matching and lower throughput without improving quality, a caution for automating discretion.
Still holds this week: gains are context-dependent, complements and governance shape who benefits, and measurement choices can misstate effects.
Top Papers
Key: each paper is tagged Relation (New, Confirms, Extends, Tension, Challenges) and evidence status (established, suggestive, framework, descriptive). Study design (RCT, quasi-experiment) is shown separately in parentheses. full key
Extendsestablished
Agentic AI as a Lens on Human Decision Making: Field Experimental Evidence from Dynamic Pricing
Zheng Kang, Fan Li, Xiaokun Wu, Pei-Yu Chen
Field RCT, high evidence
In a large field randomized controlled trial (RCT) at a major Chinese airline, delegating route-level pricing to an autonomous agent raised revenue per available seat kilometer by 5.8% relative to experienced human setters, with larger gains under higher demand and strategic uncertainty; the AI adopted a different intertemporal pricing path than humans.
So what: if this holds, overreliance on human routines under uncertainty is an earnings risk in high-frequency commercial decisions.
Confirmsestablished
Revanth Ayinala
RCT, high evidence
A large-scale RCT with 5,172 support agents finds generative AI raises average productivity by ~34–35%, with gains concentrated among novices and a small quality decline for highly experienced agents, compressing the skill distribution rather than amplifying it.
So what: if this generalizes, expert-quality dilution alongside novice lift is a service-quality risk not visible in headline throughput gains.
Extendsestablished
David Autor, Tanya Rodchenko, Josh Martin, Zanna Iscenko, Scott Strand, David Pearl, Melissa Ferere
Pre-registered RCT, high evidence
In the United States (133 lawyers across 11 intellectual property (IP) law firms), a pre-registered three-month RCT shows a custom assistant improves on-task quality while used and produces durable unaided gains on a post-intervention judgment task, concentrated among senior lawyers, with no average lasting improvement for juniors.
So what: if durable gains accrue mainly to seniors, uneven learning is a talent-pipeline risk for professional firms.
Also Notable
Tensionestablished Automated Assignment and Bureaucrats' Productivity: Evidence from a Field Experiment in Peru Alipio Ferreira, Sandra Cermeño-León: A field RCT finds algorithmic case assignment reduces timely completion by 5 percentage points and disrupts complexity-to-experience matching, with no quality gain, suggesting discretionary human matching mattered in this setting.
Newsuggestive AI Persona, Service Consumption, and User Intent Entropy: Field Experimental Evidence from an LLM Platform Junjie Li, Xiaofan Li, Lauren Xiaoyuan Lu, Yiwei Wang, Bruce Yang: In a platform field experiment, a warmer, relational persona increased sessions (+8.1%), duration (+10.6%), and intent entropy (+5.8%, a measure of how varied requests are), pointing to engagement shifts that could affect monetization.
Tensionsuggestive Better Technology, Worse Motivation: Generative Artificial Intelligence’s Mediocrity Trap Yvonne Chen, Jie Gong, Jin Li, Zibo Zhao: With professional illustrators, AI use sped early work and may flatten effort–quality returns and shorten workflows, raising a risk of weaker long-run craftsmanship incentives.
Newsuggestive From Forecast Skill to Restoration Outcomes: Operational Value of AI Early Warning for Distribution Outages Dongrui Jiang, Jochen Garcke, Dongrui Jiang: Decision-oriented evaluation suggests modest outage reductions (~6%) despite weak links between standard predictive metrics and operational value.
Newsuggestive What Firms Sell: Product Exposure to Generative AI * Tianhao Song: In this sample, around ChatGPT’s launch, firms with AI-substitutable products had declines in revenue, gross profit, and hiring, indicating a product-substitution channel distinct from task automation.
Newdescriptive Same Text, Different Numbers: The Divergence of LLM-Based Measures Hamid Boustanifar, Sasan Mansouri: Corporate-text scores vary markedly across large language models (LLMs) (mean rank corr ~0.52); ensembles stabilize ranks but not levels, warning that model choice can flip empirical results.
Extendssuggestive Complementary Capabilities and the Returns to Artificial Intelligence: Management, Digital Public Infrastructure, and Firm Growth Mohammad Mahbubur Rahman, Abiola Adeyemo: AI exposure is associated with larger wage bills and higher skill intensity where digital public infrastructure and management are strong, with no clear labor-productivity effect.
Newframework The AI Rent Who Captures the Gains When Machines Take over Offshored Tasks? Krupali Prakashkumar, Tejas Soni: A calibrated trade-in-tasks model suggests AI providers capture 85–99% of gains, raising the skill premium and leaving limited surplus for compensation.
Newframework Public AI Wealth Funds and the Economics of Shared Ownership John DeVadoss: A calibrated framework suggests public AI wealth funds can broaden participation but primarily redistribute and may not fix monopoly distortions; early dividends may be modest.
Newframework After Attention: The Economics of Markets Where the Buyer Is an Agent Sudhir Vissa, Venkata Manikanta Sangaraju: In the model, even modest agent adoption can shift seller strategies toward persuasion and higher concentration, and large language models may be susceptible to unverifiable claims.
Newdescriptive Large Language Models for Traffic Signal Management: A Systematic Review of Control Authority, Deployment Evidence, and Policy Implications KEXIN LIU, Saeed Asadi Bagloee, Majid Sarvi: Despite many lab studies, no sustained, monitored field handover is documented; unsupervised control lacks evidentiary support.
Newdescriptive Off-Grid and Microgrid Power Supply for AI Data Centers: A Techno-Economic Comparison Benchmarked to Illinois Jin Jo: With monetized interconnection delays, on-site gas plus batteries is least cost for a 600 MW AI data center in this Illinois benchmark; it is also the most carbon-intensive option.
Newdescriptive Automated Alpha Discovery Under Rigorous Out-of-Sample Validation: A Forensic Study of Trading Signal Search Tanvi Khatu: Under strict validation, in this study no price/volume-based signals survive, tempering claims about automated or LLM-guided alpha discovery.
What Moved
Contested & Watch
Methods Spotlight
Large-scale field RCT with full decision delegation (P1): delegating live airline pricing to an agent provides rare, causal evidence on agentic AI in revenue management, with heterogeneity by uncertainty to diagnose mechanisms.
Dual-Simultaneous Scaling (DSS) for matrix balancing (P13): a vectorized algorithm that minimizes Kullback-Leibler (KL) divergence, converges to a unique fixed point, and reports 2x–680x speedups, enabling larger input-output, transport, and optimal-transport applications.
Conditional diffusion scenarios inside Wasserstein-DRO dispatch (P26): combining generative scenarios with distributionally robust optimization, in simulations reduced cost volatility and tail risk in high-renewable microgrids.
