An Explainable and Audit-Aware Analytics Platform with Grounded AI Assistance: Design and Evaluation on Retail Order Data

Business data analysts typically switch between a number of tools: one to clean a file, another to chart the data, a third to fit a model, and a more recent tool, a chatbot, to summarise the data. Every hand-over is losing information on changes that have occurred and from where a figure is from. This paper outlines and reviews one web platform that holds these steps in check and holds each step accountable. Reversible data cleaning (dry-run preview, cell-level audit log), same filter model for dashboards, order-level estimation of profit and sales (requires a baseline), anomaly screening, model explanation with scenario analysis, and an assistant based on computed statistics and never inventing a number. It is tested on the 51,290-line Global Superstore dataset with a chronological split of 80/20. Tree ensembles achieve an R² of 0.73 for profit and a reduction in the mean absolute error of approximately 50% compared to the historical mean baseline and, for sales, Random Forest produces an R² of 0.86. A sensitivity study reveals that larger ensembles are not really that much better, and that Gradient Boosting actually gets worse with later data input – as more stages are added, the more worrisome they become. There are also things the system cannot do: it can guess the result of a given order and cannot predict future months; its language component hasn't yet been assessed using a formal faithfulness measure. The system has all 13 automated API tests passing and is deployed publicly.