Can AI Reliably Process Critical Data? A 10‑Minute End‑to‑End Test with Four Excel Tables

Test Overview

The author conducted a hands‑on test of the Office Xiaoxuan desktop AI assistant using a simulated university administration scenario. Four Excel files—student roster, classroom attendance, leave approvals, and warning rules—were placed in a local folder and processed from file ingestion to final interactive HTML dashboard.

Data and Task Definition

The four tables contain:

Student roster : 321 records with college, grade, and enrollment status.

Classroom attendance : 8,137 sign‑in rows.

Leave approvals : 92 rows with various approval statuses.

Warning rules : 10 cleaning, correction, and alert rules.

Key challenges include mismatched student IDs, extra spaces in sign‑in status, approved leaves that appear as “absent”, and records for students who have withdrawn but still have attendance entries. The goal was to define statistics, clean anomalies, perform cross‑table verification, and produce results without altering the original files while preserving source and justification for every change.

Using Office Xiaoxuan

The desktop client was chosen because its local agent can read multiple files within an authorized directory and keep all outputs in the same workspace, enabling a continuous workflow that mimics real office tasks. The assistant offers a “plan mode” that first outlines file relationships, identifies risks, and confirms the execution plan before any processing begins.

Execution Steps

Plan mode : The AI listed how the four files relate, highlighted potential mismatches, and waited for user confirmation.

Data cleaning : It detected anomalies such as attendance marked “absent” that matched an approved leave, missing student IDs, and stray whitespace, then applied the defined cleaning rules.

Cross‑table verification : The assistant matched records across all tables, ensuring each change retained its original source and justification, and that pre‑ and post‑cleaning counts aligned.

Result generation : Cleaned Excel files and an interactive HTML dashboard were produced. The dashboard can be previewed inside the software, viewed fullscreen, or shared with others.

Transparency and Cost

Each dialogue round displays token consumption, making the cost of the AI interaction transparent. The tool awards “points” per conversation; users can monitor their balance and purchase additional packs if needed.

Additional Capabilities

Beyond spreadsheet processing, the assistant can connect to external data sources such as Feishu and Obsidian, offers a customizable desktop pet for a companion feel, and provides a “skill center” where users can create reusable skill packages that combine local agents with expert modules.

Verification and Limitations

After the automated pipeline, the author recommends re‑checking results: open a new tab for an independent verification conversation to avoid context interference, optionally run two separate conversations and then perform manual sampling. For high‑risk data (attendance, finance, etc.) the final judgment must remain human‑driven.

Personal Takeaways

The test demonstrates that AI can stitch file reading, relationship understanding, data cleaning, cross‑table verification, and delivery into a single ten‑minute workflow. It is especially useful for administrators, teachers, and operators who regularly handle multi‑sheet reports and need a transparent, locally‑executed assistant rather than a simple text‑generation tool.