Academic papers, industry whitepapers, and research on AI in accounting, audit, and financial reporting — from LLM applications in audit to algorithmic tax compliance and explainable AI in finance.
Study reveals reasoning AI models (GPT-6 Sol, Gemini 3.8 Flash) fail at true randomness in selection tasks critical for auditing and resource allocation, exhibiting threshold and divisibility biase...
Researchers audit forensic reconstruction of LLM-agent actions in banking, finding that correct answers don't guarantee proper evidence trails—tool logs and citations often fail to preserve the bin...
FinVector-Market-4B demonstrates that LoRA-adapted 4B parameter model achieves 91.3% JSON validity and 82.7% calculator-expression correctness on structured financial benchmarks, suggesting efficie...
Hans Schabert & Christoph Peters propose a Model Context Protocol approach that sacrifices agent autonomy for step-by-step process visibility—critical for organizations needing audit trails and pre...
Research reveals language models like Llama-3.1 rely on demographic stereotypes rather than financial facts when giving investment advice with incomplete information, raising risks for personalized...
Academic paper argues that auditable claims about AI agents require explicit policy, scope, settlement records, and authorship—addressing EU AI Act Article 12's logging requirements.
Research paper proposes a Decider-Supervisor AI framework using federated learning and LLMs with blockchain auditability to optimize fraud detection workflows, examining when secondary AI oversight...
Researchers introduce a budget-constrained evaluation framework for ML-based journal entry anomaly detection, comparing nine unsupervised detectors against supervised and deep semi-supervised metho...
Researchers find that better name matching doesn't improve company clustering without upgrading the classifier, using multilingual encoders on 3,633 real company identities from GLEIF data.
Academic paper reveals how long-horizon AI agents are vulnerable to indirect prompt injection attacks across multi-step workflows, with implications for auditing AI systems used in accounting conte...
Research shows LLMs cannot reliably decide which verification rules require human oversight vs. automated checking in regulated domains like finance, complicating agent deployment in accounting and...
Researchers auditing an agentic LLM system for label ontology processing discovered that standard label agreement metrics mask critical failures in authorization, completeness, and evidence trackin...
Researchers introduce FinSkillBench, a 2,603-episode evaluation suite testing 9 AI models on financial workflows including portfolio construction and risk management, revealing gaps between knowled...
Academic research examines how algorithmic decision-making transforms the legal concept of 'suspicion' in anti-money laundering and counter-terrorist financing, arguing AI surveillance creates new ...
Digits' ML team re-benchmarked frontier AI models including TypeSafe's Jev against transaction categorization tasks, finding they still underperform outsourced accountants at end-to-end bookkeeping...
Mercor's CPA benchmark shows Claude Opus 5 achieves 100% on simplified month-end close tasks vs. 37% for junior accountants, but only reaches 54.5% on full APEX-Accounting benchmark, highlighting t...
New benchmark evaluates LLMs' ability to handle tax filing tasks end-to-end, revealing capability gaps in real-world tax preparation workflows.
Researchers audit 33 LLM judge configurations against 45,796 worker ratings, finding that high pair accuracy doesn't guarantee agreement on acceptance rates—critical for deploying AI in workplace t...
Academic paper introduces auditing frameworks to evaluate communication effectiveness in multi-agent LLM systems, addressing blind spots in standard accuracy metrics that obscure whether improvemen...
Researchers propose Runtime Assurance Contracts (RAC) to dynamically adjust AI agent autonomy during financial tasks based on real-time evidence, closing the gap between benchmarks and trustworthy ...