<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0"><channel><title>AI Business Research</title><link>https://www.jedsonpinto.com/ai-research</link><description>New research using or studying AI/LLMs in business</description><lastBuildDate>Wed, 11 Nov 2026 14:37:39 +0000</lastBuildDate><item><title>This Machine Picks Stocks</title><link>https://doi.org/10.2139/ssrn.7682274</link><guid isPermaLink="false">doi:10.2139/ssrn.7682274</guid><category>finance</category><category>instrument</category><description>Short-term stock selection among S&amp;P 500 companies using public news, technical indicators and fundamentals; the sample period and the number of trades are not stated in the abstract. Unnamed LLMs encode these inputs into directional calls, and an expected-value layer picks trades from observed gains, losses and hit rates, tested in time-ordered out-of-sample backtests. Several model configurations earn large positive returns that persist over a long test window, but the abstract gives no magnitude and the author flags overfitting risk in the backtests.</description></item><item><title>Large Language Models in Asset Pricing: A Literature Review</title><link>https://doi.org/10.2139/ssrn.7683739</link><guid isPermaLink="false">doi:10.2139/ssrn.7683739</guid><category>finance</category><category>method</category><description>Research that turns financial text, from news and earnings calls to disclosures, fund reports, macro communication and climate narratives, into asset pricing signals; the number of papers covered is not stated. No new model is run; the review links LLMs to earlier finance text analysis and machine learning in asset pricing, then covers return prediction, ESG, fund beliefs, bonds, options, currencies and agents. It names prompt sensitivity, look-ahead bias, unstable models, training-data leakage, weak interpretability and trading costs as the main limits, and calls for realistic out-of-sample tests; no quantitative synthesis is reported.</description></item><item><title>Large Language Models and Ottoman Court Records: Long-Term Change and Regional Comparison</title><link>https://doi.org/10.2139/ssrn.7683624</link><guid isPermaLink="false">doi:10.2139/ssrn.7683624</guid><category>economics</category><category>instrument</category><description>385 transcribed Ottoman court register volumes from twelve districts, turned into 160,395 entries covering the period after regular tax surveys ended around 1600. Rules pull formulaic fields while unnamed LLMs classify documents, participant roles, and legal outcomes; checks use hand coded samples and three published Istanbul studies, with no accuracy figure in the abstract. The data recover the Istanbul patterns, show the decisive oath peaking in the early 1600s across courts, and find foreign silver reached large shares sooner in Antep, Amasya, and Kayseri than in the capital.</description></item><item><title>When Do Synthetic Participants Reproduce Human Behavior? Evidence by Target, Method and Domain</title><link>https://doi.org/10.2139/ssrn.7681498</link><guid isPermaLink="false">doi:10.2139/ssrn.7681498</guid><category>economics</category><category>method</category><description>Review of published studies that use large language models as synthetic survey and experiment respondents, organised by fidelity target, improvement method, and task domain; number of studies not stated. No single model is named; the paper compares prompt only methods with approaches needing human data or weight access, and notes that whether models read and reason like people is seldom tested. Synthetic respondents match opinion survey averages but miss dispersion, treatment effects, and structural parameters; added prompt detail shows diminishing returns, methods combine subadditively, and the authors propose salience capture plus reporting rules.</description></item><item><title>The Price of De-Correlation: Quantifying the Agreement-Cost Trade-Off in Heterogeneous Large Language Model Ensembles</title><link>https://doi.org/10.2139/ssrn.7681570</link><guid isPermaLink="false">doi:10.2139/ssrn.7681570</guid><category>finance</category><category>method</category><description>54,000 model calls predicting the direction of 100 equities over 12 dates, plus a preregistered replication on 537 expert clinical decision items. Three five agent LLM committees with rising provider diversity are compared; models are not named; agreement is measured by kappa under an independent draw protocol and errors are scored against correct answers. Provider diversity cuts agreement by kappa 0.336 and error correlation by phi 0.454 at about 4.4 times the cost, with no reliable accuracy gain; exploratory analysis finds homogeneous committees unanimously wrong 1.84 times as often.</description></item><item><title>AI-Assisted Programming and Labor Demand Lindsey</title><link>https://doi.org/10.2139/ssrn.7683628</link><guid isPermaLink="false">doi:10.2139/ssrn.7683628</guid><category>economics</category><category>object</category><description>5,414 multinational firms covering 38 percent of the US programming workforce, with adoption timing tied to predetermined software procurement cycles as a natural experiment. Copilot, an AI coding assistant, is the treatment; the authors use no model for measurement and the abstract does not name the underlying model family. A year after adoption, review induced adopters cut quarterly technical hiring by 29 percent, mostly in junior roles heavy in coding, with no change in separations, total technical headcount, or vacancy duration.</description></item><item><title>LLM Competition and the Aggregator Economy</title><link>https://doi.org/10.2139/ssrn.7683780</link><guid isPermaLink="false">doi:10.2139/ssrn.7683780</guid><category>economics</category><category>object</category><description>Game theoretic model of two LLM providers, each stronger in one feature, and an aggregator selling one subscription to users with heterogeneous preferences; no empirical data are used. No model is used; LLM providers such as OpenAI and Anthropic are the market participants, and users choose a direct subscription, the aggregator, or neither. Aggregators attract users with balanced tastes; in saturated markets they only reshuffle subscribers and help a provider only at a high enough API price, and they alter which feature a provider gains from improving.</description></item><item><title>Online-Adaptive Fuzzy Cognitive Maps for Neuro-symbolic Financial Forecasting</title><link>https://arxiv.org/abs/2611.13388v1</link><guid isPermaLink="false">arxiv:2611.13388v1</guid><category>finance</category><category>instrument</category><description>Forecasts of the NASDAQ Composite built from two price feeds and macroeconomic indicators; the abstract does not state the sample period or the number of trading days. An unnamed LLM helps assemble a fuzzy cognitive map from published research, tagging each causal edge with a citation; the abstract reports no accuracy check of these edges against a benchmark. The adaptive map fused with an LSTM beat a plain LSTM by 7.6 percentage points on return-sign accuracy; LLM-built center-node maps reached 68.9 to 85.7 percent on state direction.</description></item><item><title>Codifying Expertise: AI Agents and Apprenticeship</title><link>https://doi.org/10.2139/ssrn.7682049</link><guid isPermaLink="false">doi:10.2139/ssrn.7682049</guid><category>economics</category><category>object</category><description>Theoretical model of an expert able to embed tacit know how in an AI agent, a potential apprentice, and buyers who demand two independent suppliers; no data are used. No language model is used; the AI agent is a modelled technology that lowers marginal production cost while each period of use risks leaking the process to rivals. A cheaper AI deployment can stop the master from training an apprentice, because both suppliers would then deploy and cut joint profit, so the master stays the sole supplier of a smaller market.</description></item><item><title>On-Premise LLMs for Operational Excellence: A Lean Six Sigma Benchmark of 15 Open-Weight Models</title><link>https://doi.org/10.5281/zenodo.24082529</link><guid isPermaLink="false">doi:10.5281/zenodo.24082529</guid><category>management</category><category>method</category><description>64 Lean Six Sigma test questions covering 16 methods, all five DMAIC phases and four levels of difficulty, answered by models hosted locally on one NVIDIA DGX Spark machine. Fifteen open-weight configurations from Qwen, GLM, Llama and Mistral at Q2 to Q8 quantization; an LLM judge graded all 960 answers from 0 to 5 against author-approved references, with no judge agreement figure reported. The mixture-of-experts model Qwen3.6-35B-A3B balanced domain knowledge and inference performance best, ahead of larger dense models; architecture outweighed size, and changing quantization level did not significantly shift domain scores.</description></item><item><title>Generative Artificial Intelligence in Human Resource Management: A Critical Systematic Review of Opportunities, Ethical Challenges, Governance, and the Future of Work</title><link>https://doi.org/10.5281/zenodo.24076243</link><guid isPermaLink="false">doi:10.5281/zenodo.24076243</guid><category>management</category><category>object</category><description>Literature on generative AI and large language models in human resource management across the employee lifecycle; the number of studies reviewed and the search period are not stated. No model is applied; the review is organised around four research questions on HR opportunities, legal and ethical risk, the future of work and governance, and refines a six-stage governance framework. Low-risk, reversible tasks such as drafting and summarising show the most augmentation potential, while hiring rankings, pay, promotion and dismissal need much tighter controls; no quantitative estimates are given.</description></item><item><title>State of AI Product Management 2027: Evidence from 1,082 Product Manager Job Postings</title><link>https://doi.org/10.5281/zenodo.24072645</link><guid isPermaLink="false">doi:10.5281/zenodo.24072645</guid><category>management</category><category>object</category><description>1,082 unique product manager job postings at 165 companies, collected on 9 November 2026 through the Greenhouse, Ashby and Lever public job-board APIs. No language model is used; postings count as AI-focused through a keyword rule on title and body, and the abstract reports no validation of that rule. The author founded a related course platform. 45.8% of all postings count as AI-focused, yet just 15.1% flag AI in their titles; among US postings disclosing pay, AI-focused roles have a median midpoint of USD 227,500 versus USD 198,000, about 15% more.</description></item><item><title>When AI thinks about culture: comparing language-based and empirical measures of individualism–collectivism</title><link>https://doi.org/10.1057/s41267-026-00913-3</link><guid isPermaLink="false">doi:10.1057/s41267-026-00913-3</guid><category>management</category><category>method</category><description>Country-level individualism-collectivism scores from five LLMs compared with a recently validated index built from nationally representative surveys; the number of countries is not stated. Five unnamed LLMs produce country estimates that are benchmarked against the survey index; correlation reaches r = 0.874, and smaller and older models align less closely. Western societies are rated too individualist and Confucian East Asian societies too collectivist, deviations that look systematic rather than random and track Hofstede&#x27;s original rankings.</description></item><item><title>ShopMind: A Mental World Model for Personalized and Realistic E-commerce User Simulation</title><link>https://arxiv.org/abs/2611.11897v1</link><guid isPermaLink="false">arxiv:2611.11897v1</guid><category>management</category><category>agent</category><description>Online shopping user simulation; the abstract does not state the dataset, platform, sample size, or period used in the experiments. ShopMind models how simulated shoppers&#x27; beliefs, uncertainties, and intentions evolve, then uses synthetic behavior and mental state pairs to train a model that infers latent states; the base LLM is not named. Reports gains of 18.0 percent in personalization and 19.6 percent in behavioral realism over the best baselines, plus better consumption value measurement and click through rate prediction.</description></item><item><title>Steering Emergent Behavior in Multi-Agent LLM Systems via Information Networks</title><link>https://arxiv.org/abs/2611.11714v1</link><guid isPermaLink="false">arxiv:2611.11714v1</guid><category>economics</category><category>agent</category><description>Groups of ten identical LLM agents in repeated group decision games over 1,080 episodes and six games covering coordination, anticoordination, social dilemmas, and discovery. A budgeted manager adds directed links that let agents see peers&#x27; past actions; the abstract does not name the model, and the authors frame results as one controlled LLM population. More peer information helps or is neutral when private and collective goals align but amplifies harm under free riding; in the public goods game no information was best, and barbell wiring was worst.</description></item><item><title>Coevolution in generative AI content ecosystems: behaviour-risk feedback, governance chains and local stability boundaries</title><link>https://doi.org/10.21203/rs.3.rs-11310134/v1</link><guid isPermaLink="false">doi:10.21203/rs.3.rs-11310134/v1</guid><category>management</category><category>object</category><description>Theoretical evolutionary game among content producers, platform governance units, and users of generative AI content; the model is dimensionless and uncalibrated, with no empirical sample. No language model is run; generative AI enters as producers&#x27; ability to rework risky content after moderation, alongside finite moderation capacity, provenance retention, and trust memory. Weakening several governance channels together raised risk and cut trust and quality more than single channel losses; tenfold faster trust memory shifted the stability boundary toward weaker feedback.</description></item><item><title>Which local businesses do AI tools recommend? Answer stability, cross-tool agreement and cited sites for four trades in ten US cities</title><link>https://doi.org/10.5281/zenodo.24074945</link><guid isPermaLink="false">doi:10.5281/zenodo.24074945</guid><category>management</category><category>object</category><description>Buyer questions on four trades (dentists, plumbers, restaurants, accountants) across ten US cities, asked twice per tool on 30 and 31 October 2026, yielding 1,197 usable answers. ChatGPT, Claude, and Perplexity were queried through their APIs with web search on; this is a descriptive audit run by a firm that sells AI visibility audits, with no accuracy benchmark. A repeat query named only 56 percent of the same businesses on average, the top pick matched in 44 percent of pairs, and just 7 percent of businesses were named by all three tools.</description></item><item><title>LEDGER: Deterministic Checks and Auditable Evaluation for Business AI Systems</title><link>https://doi.org/10.21203/rs.3.rs-11639610/v1</link><guid isPermaLink="false">doi:10.21203/rs.3.rs-11639610/v1</guid><category>management</category><category>method</category><description>A vendor&#x27;s evaluation platform tested on its own business intelligence assistant, using 97 held out scenarios built from the company&#x27;s data, questions, and rules. Deterministic checks, an LLM judge, and sampled human review score DeepSeek V4 Flash against five unnamed proprietary models and GLM 5.3 given identical tools and definitions. Self hosted DeepSeek V4 Flash passed 96 of 97 scenarios while rivals passed 94 to 97, and the rivals cost an estimated 3 to 54 times more per correct answer.</description></item><item><title>Generative AI in Tourism and Hospitality: A Systematic Review</title><link>https://doi.org/10.3390/su182111405</link><guid isPermaLink="false">doi:10.3390/su182111405</guid><category>management</category><category>object</category><description>Systematic review of 125 peer reviewed journal articles on generative AI in tourism and hospitality, drawn from Web of Science and Scopus for 2022 to June 2026 under PRISMA 2020. Generative AI is the topic of the reviewed studies; the authors use no model and code each article by method, theory, application, benefits, constraints, and personalization mechanism. SEM and scenario experiments dominate, with TAM, UTAUT, and trust theories used most; the review sorts four benefit and four constraint categories and proposes a framework linking GenAI traits, tourist responses, and implementation.</description></item><item><title>Generative AI Reliance and Perceived Investment Decision Quality Among Saudi Arabian Retail Investors: The Sequential Roles of Perceived Reliability, Confidence, and AI Overreliance</title><link>https://doi.org/10.3390/jrfm19110908</link><guid isPermaLink="false">doi:10.3390/jrfm19110908</guid><category>finance</category><category>object</category><description>Survey of 497 retail investors in Saudi Arabia who had used generative AI for investment activities; the abstract does not state the survey period. Generative AI is studied as a decision aid rather than used by the researcher; links among reliance, perceived reliability, confidence, overreliance, and GenAI financial literacy are estimated with PLS SEM. Reliance relates positively to perceived decision quality, but a negative indirect path runs through reliability, confidence, and overreliance; literacy weakens the reliability to confidence link. Effect sizes are not stated.</description></item><item><title>The impact of explorative generative AI usage on construction project professionals&#x27; innovative work behavior: a technology affordance actualization perspective</title><link>https://doi.org/10.1108/ecam-06-2026-1186</link><guid isPermaLink="false">doi:10.1108/ecam-06-2026-1186</guid><category>management</category><category>object</category><description>Structural equation model fitted to 239 valid survey responses from construction project professionals in the Chinese construction industry; the survey period is not stated. Generative AI use is the explanatory variable rather than a research tool; no model is named, and explorative usage is linked to knowledge acquisition and then to innovative work behavior. Explorative GenAI use raises knowledge acquisition, which in turn lifts innovative work behavior, and the effect holds only when organizational support is high; coefficients are not stated in the abstract.</description></item><item><title>An Algorithmic Framework for Hybrid Retrieval and LLM-Based Filtering in Pharmaceutical Semantic Catalogue Search</title><link>https://doi.org/10.3390/a19110977</link><guid isPermaLink="false">doi:10.3390/a19110977</guid><category>management</category><category>method</category><description>Catalogue search for call center staff at Bidafarma, a Spanish pharmaceutical wholesaler carrying products from over 30 distributors in three categories, tested on 14,997 production queries collected November 2024 through July 2026. An unnamed LLM filters Elasticsearch candidates only on semantically complex queries picked out by a routing classifier; on the production queries the full system reaches 99 percent accuracy. Average response time is 2.57 seconds with a p95 of 7.4 seconds; the LLM is invoked on 35 percent of queries and removes 24 percent of lexical retrieval candidates.</description></item><item><title>Trust and Deception in AI-Generated Advertising: The Double-Edged Role of AI Realism, Disclosure, and Consumer AI Literacy</title><link>https://doi.org/10.21203/rs.3.rs-10571464/v1</link><guid isPermaLink="false">doi:10.21203/rs.3.rs-10571464/v1</guid><category>management</category><category>object</category><description>Between subjects experiment with 224 adult consumers who saw one AI generated ad for an invented supplement brand, in a 2 by 2 design crossing realism and disclosure. Generative AI content is the stimulus rather than a research tool; no model is named, and conditional process analysis with PROCESS Model 3 tests moderation by measured AI literacy. Greater realism raises advertising and brand trust, and disclosure amplifies that gain; realism also raises perceived deception most when AI use is hidden, though the three way interaction is not significant.</description></item><item><title>AI-Enabled Cryptocurrency Scams and Investor Vulnerability: Evidence on Scam Sophistication, Adoption, Financial Access and Losses</title><link>https://doi.org/10.21203/rs.3.rs-11383134/v1</link><guid isPermaLink="false">doi:10.21203/rs.3.rs-11383134/v1</guid><category>finance</category><category>object</category><description>Cryptocurrency fraud studied with World Bank figures, survey responses and public blockchain analytics, comparing investor losses across eleven countries; the years covered are not stated. Generative AI enters as a feature of scam operations, not as a research tool; scam sophistication, adoption, and account ownership are related to losses through descriptive comparisons and Spearman correlations. Scam operations using AI tools took in roughly 4.5 times the revenue of other operations; loss prevalence ranges from 1 to 36 percent, adoption is unrelated to losses, and account ownership correlates at minus 0.74.</description></item><item><title>Easy listing, hard selling? How AI-source disclosure in second-hand product descriptions shapes purchase intention</title><link>https://doi.org/10.1080/0267257x.2026.2760826</link><guid isPermaLink="false">doi:10.1080/0267257x.2026.2760826</guid><category>management</category><category>object</category><description>Four online experiments on product descriptions in consumer to consumer second hand markets, contrasted with primary markets; sample sizes and participant sources are not stated. Generative AI as the description source is the manipulated factor rather than a research tool; the abstract names no model and tests mediation through information diagnosticity and seller trust. Disclosing AI authorship lowers purchase intention in second hand but not primary markets, more so for experience goods; the drop shrinks when AI only proofreads or the seller reviews the text.</description></item><item><title>Mind the Gap: Belief-Policy Gaps about an Artificial Bargaining Counterpart in the Ultimatum Game</title><link>https://doi.org/10.21203/rs.3.rs-11409231/v1</link><guid isPermaLink="false">doi:10.21203/rs.3.rs-11409231/v1</guid><category>economics</category><category>object</category><description>108 university students give incentivised forecasts of an LLM&#x27;s offers and acceptance choices as proposer and responder in an Ultimatum Game, compared with 1,000 simulated plays of each scenario. The abstract does not name the model; it plays the game in simulations, and its choices change with temperature, prompt wording, the numeric scale used, and the inference setup. Proposer forecasts show no mean bias, but 85.2 percent of subjects expect a weakly monotone acceptance rule the model does not follow; 62.0 percent see $500 as optimal where the model&#x27;s policy favours $250.</description></item><item><title>Reading the City: Urban Sentiment, Technological Disruption, and the Exporting and Performance of High-Growth Firms in China</title><link>https://doi.org/10.31235/osf.io/9vm3s</link><guid isPermaLink="false">doi:10.31235/osf.io/9vm3s</guid><category>management</category><category>instrument</category><description>Simulated survey of 2,700 firms, 540 of them high growth, across 25 Chinese cities, built to mimic the design of the World Bank Enterprise Survey rather than drawn from real respondents. An unnamed large language model classifies text into city and industry sentiment scores on economic development and technological disruption; the abstract reports no check against human coding. Development sentiment rises with export intensity, disruption sentiment follows an inverted U pattern, and both moderate export returns to productivity and growth across 72 multiverse specifications; no magnitudes stated.</description></item><item><title>Agent4Crypto: A Dual-Reflective and Market-Feedback-Aligned LLM-Based Multi-Agent System for Cryptocurrency Trading</title><link>https://arxiv.org/abs/2611.12269v1</link><guid isPermaLink="false">arxiv:2611.12269v1</guid><category>finance</category><category>agent</category><description>Cryptocurrency trading on BTC, ETH, and XRP using price, technical indicator, news, and chart inputs; the sample period and data frequency are not stated in the abstract. An LLM multi agent system with three trader styles whose weights a PPO gating policy learns from reflection on realized trades and counterfactual contribution; the base model is not stated. Reports higher total return and Sharpe ratio than traditional, machine learning, reinforcement learning, and earlier LLM trading baselines; the abstract reports no magnitudes.</description></item><item><title>When Divergence Falls: Generative AI and Information Environment</title><link>https://doi.org/10.2139/ssrn.7681797</link><guid isPermaLink="false">doi:10.2139/ssrn.7681797</guid><category>accounting</category><category>object</category><description>Financial analysts&#x27; earnings forecasts and research reports before and after the public release of ChatGPT; the abstract states no sample size or period. A classifier flags AI assisted content in analyst reports, and textual similarity of reports on the same firm quarter is measured; the abstract reports no accuracy figure for the classifier. After ChatGPT, forecast divergence drops sharply, most where reports show more AI help, and consensus accuracy improves; analysts&#x27; errors appear more correlated, and firm investment responds less to Tobin&#x27;s Q.</description></item><item><title>Monetization of Generative AI: Advertising, User Engagement, and Market Outcomes</title><link>https://doi.org/10.2139/ssrn.7681558</link><guid isPermaLink="false">doi:10.2139/ssrn.7681558</guid><category>economics</category><category>object</category><description>Metro level and country level mobile app usage data on ChatGPT around the introduction of advertising in the United States, with other countries serving as controls. No language model is used as a research tool; ChatGPT is the platform studied, with Claude and Gemini as comparison apps in a difference in differences design. ChatGPT weekly active users fall 5.4% after ads launch, more so in metros with larger prior AI user bases; US time spent and subscription revenue also decline relative to controls.</description></item><item><title>Is AI Reviving the Renaissance Man? LLM Adoption and Scientific Novelty in 22 Million Abstracts</title><link>https://doi.org/10.2139/ssrn.7682051</link><guid isPermaLink="false">doi:10.2139/ssrn.7682051</guid><category>economics</category><category>object</category><description>21.9 million journal articles published from 2018 to 2025, with researcher careers compared before and after each author&#x27;s first paper flagged for LLM use. A purpose trained detector flags LLM generated text in abstracts to measure adoption; the abstract reports no accuracy figure for the detector. A model of idea search motivates the test. LLM use appears in 40% of abstracts by October 2025; adoption is linked to 6% more uncommon term combinations, 4.5% more forward disruption and a 9% higher chance of top decile citations.</description></item><item><title>Not Your Average Clever Hans: Latent Representation of Psychological Constructs in Large Language Models</title><link>https://doi.org/10.2139/ssrn.7682438</link><guid isPermaLink="false">doi:10.2139/ssrn.7682438</guid><category>management</category><category>method</category><description>66 psychological constructs, among them moral agency, psychological ownership and perceived goal progress, examined inside 16 open weight LLMs with a multitrait multimethod design. Researchers probe latent layer representations, not only outputs, to test whether silicon sample responses pass through construct states; the abstract names no model family and reports no agreement statistic. Internal representations meet construct validity standards, are shared across developers and model sizes, and cluster by appraisal process; whether activation changes answers depends on closeness to response production.</description></item><item><title>Code-Paper Consistency in Economics and Finance</title><link>https://doi.org/10.2139/ssrn.7681378</link><guid isPermaLink="false">doi:10.2139/ssrn.7681378</guid><category>economics</category><category>instrument</category><description>1,820 articles published in top economics and finance journals, each paired with its analysis code; statements about the implemented analysis are the unit of review. A structured LLM protocol compares each claim with the matching code; the abstract names no model and reports no accuracy figure against human reviewers, only robustness checks on the headline share. 10% of resolved claims show a discrepancy and 49% of articles have a potentially serious one tied to a central result (37% to 56% across robustness checks), most often in sample rules.</description></item><item><title>Removing the Ladder from the Bottom Rung: Agentic AI, the Erosion of Apprenticeship, and the Crisis of Building Professional Judgement</title><link>https://doi.org/10.2139/ssrn.7682720</link><guid isPermaLink="false">doi:10.2139/ssrn.7682720</guid><category>management</category><category>object</category><description>Critical review of 28 recent sources, spanning payroll microdata, job vacancy data, industry telemetry, workforce surveys and randomised experiments, read through Lave and Wenger&#x27;s theory of peripheral participation. No language model is used for analysis; agentic AI and its effect on junior roles is the object, and the disclosure says generative AI only checked APA formatting. Cited evidence puts employment of AI exposed workers aged 22 to 25 at 19% below counterfactual and shows analysts following coin flip agent advice 86% of the time; causation remains unsettled.</description></item><item><title>Counsel ex Machina: The Value and Limits of AI Financial Advice</title><link>https://doi.org/10.2139/ssrn.7681824</link><guid isPermaLink="false">doi:10.2139/ssrn.7681824</guid><category>finance</category><category>object</category><description>A newly assembled set of human conversations with chatbots about personal finance, plus two preregistered laboratory experiments that vary access to AI; sample sizes are not stated. The abstract does not name the model; chatbot answers on equity allocation and mortgage refinancing are compared with recommendations implied by theoretical models, and no accuracy figure is given. Advice tracks theory benchmarks closely; users rarely share personal details yet weight the advice heavily, make fewer reasoning errors and more uniform choices, but show more bias under misleading framing.</description></item><item><title>Direction Instability in an LLM Evidence Synthesizer: A Pre-Registered Two-Wave Study of Recorded Trading Inputs</title><link>https://doi.org/10.2139/ssrn.7681837</link><guid isPermaLink="false">doi:10.2139/ssrn.7681837</guid><category>finance</category><category>method</category><description>Evidence logs captured by a live trading system running on a demo account, replayed in two preregistered waves covering 898 analysed evidence sets and 4,490 analysed model calls. An unnamed hosted model turns each evidence set into a direction vote, queried five times at temperature 0; instability is the Gini-Simpson diversity of outputs, with no ground-truth accuracy test. Instability was lower where the rule already reached consensus (gaps of 0.011 and 0.018) but not significantly; about one third of sets changed consensus outcome across calls, while a local open-weights model agreed fully.</description></item><item><title>From Words to Portfolios: Disentangling Narrative and Reality in Mutual Fund Differentiation</title><link>https://doi.org/10.2139/ssrn.7682021</link><guid isPermaLink="false">doi:10.2139/ssrn.7682021</guid><category>finance</category><category>instrument</category><description>Mutual fund portfolio holdings and prospectus text; the abstract does not state the sample period, number of funds, or market. A transformer-based embedding of holdings is paired with prospectus embeddings from an unnamed pre-trained language model; the abstract reports no validation of either uniqueness measure. The narrative and portfolio uniqueness measures are only weakly correlated; distinctive prospectuses go with higher fees and, after strong returns, more inflows, while distinctive holdings predict alpha only for funds with strong records.</description></item><item><title>Do Modern Difference-in-Differences Estimators Change Published Findings in Business Research?</title><link>https://doi.org/10.2139/ssrn.7682393</link><guid isPermaLink="false">doi:10.2139/ssrn.7682393</guid><category>management</category><category>instrument</category><description>Published staggered-adoption difference-in-differences specifications in business-school journals that maintain journal-wide public replication archives, covering 355 specifications drawn from 40 papers. Unnamed large language model agents, following Hull (2026), find eligible specifications, reproduce the TWFE estimate, and re-estimate it with two heterogeneity-robust estimators; the abstract reports no accuracy check on the agents. Among comparisons passing the leverage screen, TWFE and modern estimates correlate at 0.61 and 0.70, reverse sign 25.1 and 18.0 percent of the time, and modern standard errors run 1.30 to 1.61 times larger.</description></item><item><title>Inside the Trading Loop: Markets, Models, and Moves with LLM Agents</title><link>https://arxiv.org/abs/2611.11657v1</link><guid isPermaLink="false">arxiv:2611.11657v1</guid><category>finance</category><category>agent</category><description>A broker recorded paper trading account, which ran with deterministic fallbacks, from March 11 to August 28, 2026, plus retrospective replays of LLM enabled configurations. LLM analyst agents read company news and macroeconomic inputs and pass structured reports to a coordinator that decides cash reserves and stock weights; the abstract does not name the model or report validation. The account returned 34.5 percent against 14.4 percent for SPY, with a deeper maximum drawdown; a constant cash control performed about as well as stance based sizing.</description></item><item><title>THEATRON: A Benchmark for Evaluating Persona Persistence in Long-Horizon LLM Decision-Making</title><link>https://arxiv.org/abs/2611.11487v1</link><guid isPermaLink="false">arxiv:2611.11487v1</guid><category>finance</category><category>method</category><description>A financial replay benchmark where models see shared daily market evidence and their own account state over multimonth trajectories; markets, assets, and period are not stated. Four unnamed LLMs take behavioral finance personas through zero shot, one shot, or chain of thought injection, judged on within persona stability, cross persona separation, and action collapse. Personas shift behavior relative to a no persona baseline, but the shifts often fail to form coherent policies, and apparent separation can reflect stereotyped action collapse; no magnitudes are given.</description></item><item><title>FinCutBench: A Continually Updated Benchmark for Measuring the Temporal Boundaries of LLM Financial Knowledge</title><link>https://arxiv.org/abs/2611.11588v1</link><guid isPermaLink="false">arxiv:2611.11588v1</guid><category>finance</category><category>method</category><description>A monthly refreshed set of 9,000 Chinese and English multiple choice questions on Chinese and US market facts, dated January 2023 to September 2026, across four categories of financial information. Unnamed LLMs answer the questions, and a fitted monthly accuracy profile dates each model&#x27;s knowledge boundary, with resampling used to gauge how stable that date is; the abstract gives no accuracy figures. Boundary dates vary by model; afterward some models drop in US asset ranking and earnings recall, while none of the eight paired forecast error tests reaches significance at the 0.10 level before adjustment.</description></item><item><title>When Memory Is Not Enough: What Drives Robustness in AI Decision Systems? Evidence from Sequential Financial Decision-Making</title><link>https://doi.org/10.3390/electronics15215181</link><guid isPermaLink="false">doi:10.3390/electronics15215181</guid><category>finance</category><category>agent</category><description>Sequential trading decisions on Apple, Tesla, and NVIDIA over 2021 to 2023, used as a test bed for AI decision systems that combine language models, retrieved memory, and rule based execution. Targeted ablations vary memory access, coordinator choice, retrieval quality, randomness, and execution modules; the abstract names no specific model and judges behavior by risk adjusted and downside performance. Memory alone does not yield robust behavior; the full setup, pairing relevant memory with bounded model coordination and constrained execution, shows stronger risk adjusted and downside results, with no magnitudes stated.</description></item><item><title>Pemanfaatan Tools AI (Artificial Intelligence) Generatif dalam Meningkatkan Produktivitas dan Budaya Kerja Karyawan Gen-Zdi Era Digital</title><link>https://doi.org/10.62710/87wts837</link><guid isPermaLink="false">doi:10.62710/87wts837</guid><category>management</category><category>object</category><description>Qualitative study based on in depth interviews with five Generation Z employees from several industry sectors; the abstract does not state the country or the interview period. Tools such as ChatGPT, Claude, Microsoft Copilot, Grammarly, and BRIBrain are the subject of study; the researchers use no model themselves, and findings come from informant accounts. Informants report faster work on repetitive and basic analytical tasks and a more adaptive work culture, while manually double checking outputs and masking company data; no magnitudes are reported.</description></item><item><title>AI as a Junior Assessor: Can Large Language Models Score IFRS S2 Climate Disclosures? Evidence from Nigerian Upstream Oil and Gas</title><link>https://doi.org/10.5281/zenodo.24046192</link><guid isPermaLink="false">doi:10.5281/zenodo.24046192</guid><category>accounting</category><category>method</category><description>FY2025 reports of three Nigerian upstream oil and gas firms (Seplat Energy, Oando, Aradel Holdings), scored on 15 IFRS S2, SASB and NUPRC climate indicators against a published human benchmark. DeepSeek V4.1 Flash, Gemini 2.5 Flash, GPT-4o and Claude Sonnet 4.5 gave 0 to 3 scores without the coder&#x27;s specific criteria; agreement with the human coder is measured by quadratic weighted kappa and error size. Single Claude and DeepSeek runs reached within firm kappa of 0.53 and 0.51, and a three run DeepSeek median 0.64; pooled, a naive firm level rule matched or beat every single run.</description></item><item><title>The AI Auditability Dilemma: A Comparative Analysis of Retrieval-Augmented Generation and Fine-Tuning Large Language Models in Credit Risk Management</title><link>https://doi.org/10.5281/zenodo.24053429</link><guid isPermaLink="false">doi:10.5281/zenodo.24053429</guid><category>finance</category><category>method</category><description>Conceptual comparison of two ways to adapt LLMs for credit risk management; the abstract reports no dataset, sample, or study period. Parameter efficient fine tuning and retrieval augmented generation are assessed on factual accuracy, data freshness, cost and explainability; the abstract names no model and reports no empirical test. Argues fine tuning suits credit memo formatting but fails audit traceability while RAG keeps source trails, and proposes a hybrid of a small fine tuned model with GraphRAG; no magnitudes are reported.</description></item><item><title>Avanço da IA Generativa na Geração de Ideias para o Comércio Exterior</title><link>https://doi.org/10.5281/zenodo.24060048</link><guid isPermaLink="false">doi:10.5281/zenodo.24060048</guid><category>management</category><category>object</category><description>Exploratory qualitative study of business model creation in foreign trade, based on 2023 to 2026 publications and official Brazilian trade sources; the article is written in Portuguese. Claude, ChatGPT, Gemini and Copilot are reviewed as aids for generating and checking international business models; the abstract reports no experiment or validation of their outputs. Usefulness is judged high for market and value proposition ideas, moderate for comparing models, and low for customs, tax classification and non tariff checks; a five step GERAR-VALIDAR protocol is proposed.</description></item><item><title>AI-Focused Promotional Positioning and Client Attribution of Responsibility for Accounting Firm Service Failures</title><link>https://doi.org/10.33423/r4evn841</link><guid isPermaLink="false">doi:10.33423/r4evn841</guid><category>accounting</category><category>object</category><description>A between-subjects experiment in which participants read about an accounting firm that prepared a tax return containing errors; the abstract does not state participant count or pool. No language model is used by the researcher; the manipulation is whether the firm&#x27;s marketing stresses generative AI tools or traditional services, with the error&#x27;s cause left unclear. Participants shown AI-focused marketing blame the firm more, credit it with more control over preventing errors, and expect errors to recur; the abstract gives no effect sizes.</description></item><item><title>Generative Artificial Intelligence Governance in Procurement: Skills Transformation, Decision-Making Accountability and Ethical Issues</title><link>https://doi.org/10.47577/tssj.v89i1.13858</link><guid isPermaLink="false">doi:10.47577/tssj.v89i1.13858</guid><category>management</category><category>object</category><description>Open-ended questionnaire answered by 40 professionals working in procurement, IT, legal, compliance, and retail or supply chain, coded with reflexive thematic analysis. Generative AI is the subject of study rather than a research tool; the abstract names no specific model and reports no coder agreement statistic. Respondents value GenAI for preparatory document work but refuse to let it make legally or economically binding calls alone, citing confidentiality, hallucinations, and blurred accountability; a governance matrix tiered by risk is proposed.</description></item><item><title>Hybrid multimodal model for predicting the probability of winning in competitions of the Presidential Grants Fund</title><link>https://doi.org/10.5281/zenodo.24054145</link><guid isPermaLink="false">doi:10.5281/zenodo.24054145</guid><category>management</category><category>instrument</category><description>1,420 projects entered in Presidential Grants Fund competitions in the Orenburg region from 2017 to 2026, split 80/20 into training and test sets. YandexGPT scores each project, and the score joins expert ratings, Yandex.Wordstat search interest, and VK social signals in a logistic regression; the abstract reports no separate accuracy for the LLM scores. The full hybrid model reaches a ROC-AUC of 0.83, compared with 0.70 for base expert scoring and 0.74 for logistic regression on raw features.</description></item><item><title>Artificial Intelligence and the Future of Human Work: Global Governance, Workforce Transformation, and Strategic Pathways for Industry 5.0</title><link>https://doi.org/10.5281/zenodo.24044253</link><guid isPermaLink="false">doi:10.5281/zenodo.24044253</guid><category>economics</category><category>object</category><description>A review of 62 scholarly and institutional sources, combined with analysis of AI governance documents issued from 2016 to 2026, including a focus on developing economies like Bangladesh. No language model is used by the author; generative AI is the subject, examined for its effects on employment, productivity and the treatment of workers under governance rules. Cited evidence suggests AI reshapes tasks more than it removes jobs: the ILO puts about 24 percent of global jobs in exposed occupations and 3.3 percent in the top group, and trials show 14 to 56 percent task gains.</description></item><item><title>BuySideFlow: A Point-in-Time Benchmark for LLM Agents in Investment Research</title><link>https://arxiv.org/abs/2611.12966v1</link><guid isPermaLink="false">arxiv:2611.12966v1</guid><category>finance</category><category>method</category><description>408 practitioner-derived Chinese-language investment research tasks covering equities, funds, fixed income and macroeconomics, built on a frozen snapshot of a licensed production financial database. Nine LLM agent configurations write SQL and Python workflows, scored against executable reference answers plus expert audits of 900 runs for point-in-time errors; the underlying models are not named. Result accuracy runs from 16.2% to 36.8%, and of 380 audited runs that passed result checks, 193 (50.8%) still broke point-in-time rules such as disclosure cutoffs.</description></item><item><title>Pengaruh Generative Ai Marketing Dan Creator Trust Terhadap Purchase Intention Pada Tiktok Shop</title><link>https://doi.org/10.62379/jebd.v4i2.5316</link><guid isPermaLink="false">doi:10.62379/jebd.v4i2.5316</guid><category>management</category><category>object</category><description>Questionnaire survey of 31 TikTok Shop consumers selected by purposive sampling; the abstract does not state the country or the survey period. No language model is used by the researchers; generative AI marketing is a survey construct, tested with creator trust in a multiple linear regression run in SPSS. Generative AI marketing shows no significant effect on purchase intention (p of 0.705), while creator trust is positive and significant; together the two explain 50.8 percent of the variance.</description></item><item><title>Track or Follow? A Jev-Compatible Multi-Agent Trading System When Web News Lags the Market</title><link>https://arxiv.org/abs/2611.10414v1</link><guid isPermaLink="false">arxiv:2611.10414v1</guid><category>finance</category><category>method</category><description>Web news items about listed firms, matched to announcement-window abnormal returns and to issuer 8-K filings; the abstract does not state sample size or period. A multi-agent trading system reads news through the Jev typed interface using an unnamed pinned LLM; its event-type labels agree with 8-K filings at an AUC of 0.833. News judgments correlate weakly with abnormal returns (weekly Spearman 0.119), but 72 percent of earnings items appear after the filing, and timed from the first post-publication trade the link drops to 0.023.</description></item><item><title>Large language models for tech mining and technology opportunity analysis: A diffusion of innovation perspective on detecting evolutionary stages of innovative practices using ChatGPT</title><link>https://doi.org/10.1016/j.techfore.2026.124946</link><guid isPermaLink="false">doi:10.1016/j.techfore.2026.124946</guid><category>management</category><category>instrument</category><description>Full texts of 129,566 scientific articles published from 2019 to 2024, used to map technology, process and industry configurations across 11 enabling technologies for technology opportunity analysis. ChatGPT-4o runs inside a hybrid pipeline pairing engineered prompts with a rule-based expert system; it is compared with expert human coders, near-graduate students, a tuned BERT classifier and rival LLMs, but the abstract reports no accuracy figure. The workflow yields 43,632 practices; supervised learning on them flags plausible vacant configurations as opportunity windows, though the abstract gives no effect sizes or counts of opportunities found.</description></item><item><title>TRACKING GENERATIVE AI ADOPTION IN A CULTURALLY DYNAMIC SETTING: DAILY GOOGLE TRENDS EVIDENCE FROM THE UAE</title><link>https://doi.org/10.5281/zenodo.24036649</link><guid isPermaLink="false">doi:10.5281/zenodo.24036649</guid><category>economics</category><category>object</category><description>Daily Google Trends series on public search interest in ChatGPT in the UAE, April 2024 to May 2025, with attention to Ramadan and other calendar effects. ChatGPT is the object of study rather than a tool; the author applies STL decomposition, GARCH volatility models, and anomaly detection, then compares SARIMA, Prophet, N-HiTS, and random forest forecasts. Interest rises over time with a workweek cycle, spikes around model releases and holidays, and dips during Ramadan; the random forest forecasts best, with a MAPE of 10.65 percent.</description></item><item><title>Open to Interpretation: How Frontier AI Developers Practice Openness through Licensing</title><link>https://doi.org/10.31235/osf.io/m5hsa</link><guid isPermaLink="false">doi:10.31235/osf.io/m5hsa</guid><category>management</category><category>object</category><description>An original longitudinal dataset of 1,313 foundation model releases by twelve developers in the US, China, France, and Canada, spanning January 2023 to October 2026. Foundation models are the subject rather than a tool; the study classifies developers&#x27; release and license terms, and the abstract reports no language model used in the analysis. Openness attaches to model portfolios in four flagship configurations; open flagships use bespoke revenue-targeted terms, and all US developers block unpermitted inspection of their top model, unlike others except Baidu.</description></item><item><title>From AI Literacy to AI Fluency</title><link>https://doi.org/10.35542/osf.io/49bmu</link><guid isPermaLink="false">doi:10.35542/osf.io/49bmu</guid><category>management</category><category>object</category><description>A conceptual article on higher education, motivated by agentic AI automating entry-level knowledge work and reshaping corporate hierarchies into diamond shapes; it uses no sample or data. No model is used or evaluated; the authors propose a two-stage pedagogy of deliberate friction inside coursework followed by experiential learning such as clinical-style internships and client projects. Argues that basic AI literacy falls short and that assessment should shift from finished work to students&#x27; ability to evaluate and own AI outputs; the abstract reports no empirical results.</description></item><item><title>Assessing human–AI collaboration in management education: Prompting behaviour, communication patterns, and reliance trajectories</title><link>https://doi.org/10.1016/j.ijme.2026.101581</link><guid isPermaLink="false">doi:10.1016/j.ijme.2026.101581</guid><category>management</category><category>object</category><description>Metadata only. Follow the link for the paper.</description></item><item><title>Definitions, Label Names, or Demonstrations? Auditing What Governs LLM Annotation</title><link>https://doi.org/10.48550/arxiv.2611.09457</link><guid isPermaLink="false">doi:10.48550/arxiv.2611.09457</guid><category>other</category><category>method</category><description>Three political-text datasets labelled with researcher-supplied category definitions, used to test what actually drives LLM annotation of social-science text; sample sizes are not stated. A counterfactual audit swaps definitions across output labels to separate adherence to supplied definitions from the original label mapping, and also tests fine-tuning and retrieval of examples; the models are not named. Models seldom follow swapped definitions and fine-tuning reinforces the old mapping; plain k-nearest-neighbour labels recover 74% to 93% of the retrieval gain on two tasks, so retrieval mostly passes neighbour labels through.</description></item><item><title>What Matters for LLM Persona Simulation: Signal Form in Cold-Start E-Commerce Advertising</title><link>https://doi.org/10.48550/arxiv.2611.08840</link><guid isPermaLink="false">doi:10.48550/arxiv.2611.08840</guid><category>management</category><category>method</category><description>One e-commerce campaign run on Google Ads, split into five demographic segments, used to forecast ad performance before launch when no historical logs exist yet. Persona-conditioned LLM simulations of user responses, with ablations over demographic anchors, banner-blindness cues and prompt wording, scored against observed campaign results; the specific LLMs are not named. The best setup posts MAE 24.9 and MAPE 26.1%, with error 51.3% below the best non-LLM baseline and 77.3% below an LLM given explicit CTR anchors.</description></item></channel></rss>