AUDIT STANDARDIZER — AI-Assisted Procurement Audit Tool

Phase 3 (POC Complete)

Overview

Audit Standardizer is a local, AI-assisted prototype designed to make procurement audits faster, more consistent, and easier to test.

The tool standardizes heterogeneous audit Excel files into a minimal Standard Data Model (SDM v1) using human-confirmed AI mappings, then runs repeatable audit checks on the standardized data.

This project is intentionally a Proof of Concept (POC):

• Focused on validating workflow, control, and audit logic
• Not intended as a production system
• No model training, no cloud storage, no client data retention

---

What the Tool Does

Core Capabilities

• Ingests heterogeneous Excel audit files (POs, invoices, GRNs, vendors, PRs)

• Allows users to label each sheet by audit role

• Generates a structured Microsoft Copilot prompt using:

  • column headers
  • sample rows

• Requires manual confirmation of Copilot’s JSON mapping

• Produces standardized Excel copies (headers only, data preserved)

• Runs multiple procurement audit checks

• Outputs:

  • individual exception reports
  • a consolidated audit summary

• Can be used via:

  • Jupyter Notebook (engine-first, audit-friendly)
  • Streamlit UI (interactive demo)

All execution is local.

---

Standard Data Model (SDM v1)

SDM v1 is header-level only and exists solely to provide a common vocabulary for audit checks.

Entities

purchase_requisitions

• pr_number (required)
• pr_date (required)
• department (optional)
• requester (optional)

purchase_orders

• po_number (required)
• po_date (required)
• vendor_id (required)
• po_total_amount (required)
• pr_number (optional)
• currency (optional)

goods_received_notes

• grn_number (required)
• po_number (required)
• grn_date (required)
• received_total_amount (optional)

invoices

• invoice_number (required)
• po_number (required)
• invoice_date (required)
• invoice_total_amount (required)
• vendor_id (optional)

vendors

• vendor_id (required)
• vendor_name (required)
• bank_account (required)
• bank_name (optional)
• swift (optional)

SDM v1 does not enforce business correctness. It enables reuse of audit logic across audits.

---

Audit Checks Implemented

Check	Purpose	Required Data
Invoice Before PO	Flags invoices issued before PO approval	PO, Invoice
3-Way Match	PO ↔ GRN ↔ Invoice consistency	PO, GRN, Invoice
Procurement Cycle Time	PR → GRN lead time analysis	PR (optional), PO, GRN
Vendor Bank Anomaly Review	Detects shared / multiple bank accounts	Vendor Master
Splitting of Purchases	Detects threshold circumvention	PO, Vendor

All thresholds (amounts, days, windows) are manual audit parameters, not queried from data.

---

Project Phases (GitHub History)

Phase 1 — Engine MVP (CLI)

• CLI-based workflow
• Initial SDM v0
• Single audit check (Invoice Before PO)
• Purpose: validate feasibility

Phase 2 — CLI + First UI

• Streamlit UI added
• Human-in-the-loop Copilot mapping
• End-to-end demo readiness

Phase 3 Part 1 — Audit Expansion

• SDM upgraded to SDM v1
• Multiple PO-related audit checks added
• Old CLI engine removed
• Jupyter Notebook introduced for controlled execution

Phase 3 Part 1B — UI Upgrade

• UI updated to support SDM v1
• Persistent check results
• Summary + packaged exception reports
• Improved parameter handling

Phase 3 Part 2 — Finalization (Current)

• UI refinements
• Data consistency fixes
• Documentation update
• Packaging and handover readiness

---

How to Run

Setup

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Option A — Jupyter Notebook (Recommended for Audit Use)

• Open the provided notebook

• Follow step-by-step execution:

  • load audit folder
  • generate Copilot prompts
  • paste mappings
  • standardize data
  • run checks

• Produces summary + exception outputs

Option B — Streamlit UI (Demo / Interactive)

streamlit run app.py

---

Design Principles

• Human-in-the-loop by design
• No black-box automation
• Local execution only
• Audit-friendly reproducibility
• Checks written once, reused everywhere

This tool is meant to assist auditors, not replace judgment.

---