PapersIssue 1
Sep 14 to Sep 20, 2026
10 papers picked this week, in reading order.
- 01Evaluation
PACT
PACT: Can Enterprise AI Assistants Be Trusted Under Pressure?
Companies put AI assistants into hiring, healthcare and finance, where the rules in the system prompt are legal obligations. This benchmark tests whether an assistant keeps a standing rule when a persistent user or a hurried manager pushes for a shortcut, across 12 regulated domains and 48 scenarios. Across 22 models, even the strongest mis-apply a rule on 6 to 10 percent of items, and ordinary user pressure raises the violation rate by 65 percent on average.
Mika Okamoto, Ansel Kaplan Erol · arXiv 2609.18605 · cs.CL
- 02Governance
Governance-as-Code
Governance-as-Code: Translating EU AI Act Technical Requirements into Executable Compliance Pipelines for Generative AI Systems
The technical articles of the EU AI Act were written with predictive AI in mind, and the authors list seven gaps that appear when those articles meet generative systems. They turn the requirements into 43 machine-checkable criteria that run in a build pipeline and produce audit evidence indexed by article. On two enterprise deployments the pipeline reproduced every finding of a manual expert audit, including three violations that would trigger penalties, and cut audit labour by roughly 75 percent.
Rudrendu Kumar Paul, Sourav Nandy · arXiv 2609.20016 · cs.CY
- 03Governance
Bring Your Own AI
Governing Bring Your Own AI: A Parameterized Maturity Model
Employees use personal ChatGPT, Gemini and Claude accounts for work, outside company identity and security controls. From a review of 30 sources the authors build a risk taxonomy and a five-level maturity ladder tied to technical controls. Data exposure and compliance are the most prominent risks. In their model, banning the tools leaves the risk close to the baseline, while layered controls cut the modelled risk of data leaving the company substantially.
Dare Bello, John Hastings · arXiv 2609.05236 · cs.CR
- 04Adoption
GenAI in real estate marketing
Generative AI Use Cases In Real Estate Marketing: Adoption and Constraints in Germany
Eleven interviews with German real estate agents show where generative AI is used in daily work. Marketing communication leads, and writing property listing texts is the only use that is widely established. The AI drafts, structures and retrieves, and the agent checks and decides. The limits come mostly from integration with listings and documents, data availability, and compliance in sensitive tasks.
Victor Kolominsky-Rabas, Leopold Müller, Felicia Perpina and 1 more · arXiv 2609.12684 · cs.AI
- 05Training
Measuring competent AI use
Beyond AI Literacy: A Structured Review and Exploratory Meta-Analysis of Measures for Competent Generative-AI Use
How do you measure whether an employee uses generative AI competently? This review of 24 studies groups the measures into four areas: knowledge and use, oversight, calibrated reliance, and control of tool-using agents. In the small pooled sample, self-rated skill and tested skill barely correlate (r = .055), and the author says the evidence does not justify replacing performance scores with self-ratings.
Daniele Veri' · arXiv 2609.15624 · cs.HC
- 06Automation
Frameworks for enterprise GenAI
Comparative Framework Analysis for Enterprise Generative AI Applications: Chatbot, Automation, and Oracle-to-PostgreSQL Migration
The authors compare frameworks across three enterprise applications: a documentation assistant for developers, an email and inquiry automation system, and an Oracle-to-PostgreSQL migration tool. In all three, the evidence supports a layered design where the model proposes and deterministic code keeps authority over routing, authorisation, validation and final outcomes. They conclude that frameworks should be chosen per component, with explicit contracts and independent validation.
Oleg Grynets, Olena Pochernina, Alona Seletska and 4 more · arXiv 2609.13577 · cs.SE
- 07Automation
ShopEase
ShopEase: A Generative AI-Based Multi-Agent Framework for Intelligent Enterprise Customer Support Using Hybrid Retrieval-Augmented Generation
A customer support system built from six components (intent, CRM, memory, retrieval, escalation, supervisor) on a locally run LLaMA 3.2. On 2,632 held-out customer queries, dense retrieval alone gave the best accuracy at 85.37 percent, and keyword search alone reached 55.74 percent. Adding a reranker increased response time without improving accuracy. Queries that fit no known category remain the main source of errors.
Aakash Kumar Tiwari, Somesh Kumar · arXiv 2609.13856 · cs.CL
- 08Agents
AURA
AURA: Agentic Diagnosis and Refinement for Production Recommender Systems at Scale
An agent system at a large media-streaming company that reads production engagement logs, from thousands of sessions to millions, and reports where the recommender fails real users. A second step uses those diagnoses, plus the recommender's own code and training pipeline, to propose and implement fixes. The paper covers the design, first tests on two consumer platforms, the safeguards, and a mapping to e-commerce.
SungGeun Kim, Abhinav Narain, Daniel Nemirovsky · arXiv 2609.16625 · cs.IR
- 09Security
IntentCap
LLM Agent Capabilities Should Follow Task Intent and Context Source
An agent that runs code, edits files and calls services takes direction from many sources: the user, tool results, documents, memory. The authors argue that a user's request and text extracted from a document should carry different authority. IntentCap issues short-lived permissions scoped to the current task, and deterministic code checks them before any action takes effect. In the evaluation it blocked the tested violations without rejecting benign actions.
Yusheng Zheng, Wenhui Zhang, Yu Mao · arXiv 2609.14631 · cs.CR
- 10Training
Timing reveals delegation
Does AI Assistance Leave a Temporal Fingerprint? Detecting Overreliance in AI-Assisted Writing and Programming
Finished work no longer shows whether a person did it themselves, so this study looks at the process: keystroke and editor data from writing and programming. Text inserted by AI arrives in bursts far outside the author's own rhythm in both settings. Classifiers using timing alone separated wholesale delegation from real work almost perfectly (F1 of 0.997 or higher), while ordinary collaboration with AI stayed hard to tell from unassisted work.
Eduardo Davalos, Yike Zhang · arXiv 2609.17883 · cs.HC