Evidence Traceability in AI-Assisted Accounting Summaries for Small Enterprises

Abstract

Generative artificial intelligence can produce fluent accounting narratives without demonstrating that every figure and explanation follows from valid records. This conceptual article develops an evidence-traceability framework for small-enterprise summaries. It separates authorised sources, deterministic calculation, narrative generation, and human approval, and links material claims to source versions, periods, and verification status. Selected risk-management and data-governance publications provide contextual foundations; operational controls and propositions are developed analytically. No software evaluation, experiment, company dataset, or accuracy estimate is reported. Hypothetical cases distinguish arithmetic errors from mismatched periods and unsupported causal explanations. The article proposes a claim-evidence register, a review sequence, change controls, and a synthetic-data experiment for prospective testing. Three propositions concern detection of source and period errors, the distinct effect of calculation separation on numerical accuracy, and the balance between review burden and information reliability. The contribution is a testable control architecture rather than a claim that AI improves financial reporting. Application requires validation with authorised data, explicit treatment of uncertainty, and protection of unnecessary personal and commercial information. The framework does not replace accounting standards, professional judgement, or the responsibility of the person approving the summary.