Academic protocol (PHP-AIO) identifies four systemic risks—tacit knowledge loss, resilience reduction, regulatory exposure, and institutional capital degradation—that standard ROI misses when autom...
PA-DSL handles realistic audit scenarios where human labels contain errors and only some items receive expert review, improving automated classifier accuracy in accounting data labeling.
Research paper argues that AI agent failures stem from weak context engineering—poor instructions, tools, and guardrails—not agent architecture alone, with implications for reliable accounting auto...
Researchers propose TRACE, a typed schema for recording AI agent reasoning traces to ensure auditability and accountability in accounting workflows, addressing a critical gap in making AI agent dec...
New method compiles standard operating procedures into machine-readable pseudo-code for LLM agents, improving compliance by up to 16 points across six models—critical for safety-critical accounting...
PREF-Gate framework applies provenance-constrained evidence fusion to relational fraud detection, addressing validity issues in neighborhood risk assessment for financial graph analysis.
Research paper presents constraint-aware hierarchical search method for fine-grained classification under regulatory rules, addressing customs tariff, export control, and equipment coding tasks whe...
GRID, a new grammar-constrained decoding engine, ensures LLM-generated SQL is syntactically valid, role-aware, and audit-compliant—addressing enterprise security and governance requirements for dat...
Researchers propose harness-engineering framework that transforms prototype LLM applications into traceable, auditable agent architectures with deterministic behavior, source boundaries, and reprod...
New research paper introduces CAGE-1 evaluation framework for auditing autonomous AI agents in enterprise workflows, addressing control, authorization, policy compliance, and accountability challen...
MindBridge argues that traditional audit sampling—the practice of reviewing representative transaction samples to infer population health—becomes obsolete when autonomous finance systems enable 100...
Technical deep-dive on four failure modes where AI agents lose write operations without logging errors—a critical reliability gap for accounting systems handling financial transactions.
TaxCalcBench, a new open-source evaluation framework, measures whether AI agents can accurately complete tax filing tasks—addressing a critical gap in benchmarking AI for accounting workflows.
IPA survey finds 70%+ of accounting firms boosting AI budgets, but evidence of ROI remains sparse—highlighting disconnect between investment confidence and proven outcomes.
Eticas proposes open infrastructure for operationalizing AI audits across 74 competing risk taxonomies, moving beyond naming risks to executing measurable, defensible tests against real systems.
Researcher proposes meta-benchmarking framework organizing 452 public LLM benchmarks into 41 work activities and 38 banking domains, showing that general-purpose leaderboards miss domain-specific d...
Researchers propose active learning strategy for cascaded table detection and structure recognition in business documents, reducing costly fine-grained annotation burden while maintaining pipeline ...
SpreadsheetBench 2 evaluates AI agents on end-to-end spreadsheet tasks (generation, debugging, visualization) used in business financial modeling and reporting—moving beyond isolated cell operation...
Academic research reveals that LLM-compressed financial data loses decision-critical fidelity, with errors potentially amplifying across agentic systems—a risk for automated investment analysis.
Meridian releases Spreadsheet Arena, an open benchmarking platform co-developed with Cornell, CMU, and Scale AI to evaluate LLM performance in generating spreadsheet workbooks through blind pairwis...