Audit & Risk · Case Study

Forensic Expense
Audit System

How I applied AI classification, Benford's Law analysis, and pattern detection to 800+ corporate expense transactions — and delivered a risk report that changed how leadership thinks about expense governance.

Dataset
800+ transactions · 48 cardholders
Timeline
3 weeks build
My Role
Sole analyst & developer
Stack
Python · OpenAI API · Excel · Pandas
The Problem

The organisation used Pleo cards across 48 employees — a mix of department heads, operations staff, and senior management. Over the course of a quarter, 796 transactions were processed through the system, totalling a significant sum of corporate spend.

No systematic review had been done. Notes were inconsistent or missing. Some transactions had no business justification. And there were patterns that, to a trained eye, looked wrong — but nobody had looked.

The Risk

Without systematic review, expense fraud, policy breaches, and wasteful spending go undetected. In an AIM-listed company, this is a governance failure with real consequences — both financial and reputational.

The Audit Framework

I designed a 10-dimension forensic audit framework — each dimension targeting a specific category of financial risk:

01

AI Transaction Classification

GPT-4o classified every transaction by type (hotel, meal, transport, ATM, gift, mystery merchant) and assigned a policy compliance score based on the company's expense policy rules.

02

Benford's Law Analysis

Applied Benford's Law to the leading digits of all transaction amounts. Statistically significant deviation from the expected distribution flags potential manipulation — a standard forensic accounting technique.

03

Velocity Pattern Detection

Identified employees with unusual transaction frequency, clustering of spend near policy limits, or repeated transactions to the same merchant in short windows.

04

Cross-Employee Collusion Analysis

Mapped merchant usage across employees — looking for cases where multiple employees used the same unusual merchant, a known signal of coordinated fraud.

05

Note Quality Scoring

Scored every transaction note from 0–10 on specificity, business justification, and completeness. Identified cardholders with systemic poor documentation — itself a governance risk.

06

ATM & Cash Withdrawal Profiling

Cash withdrawals are the highest-risk expense category — untraceable and difficult to justify. All ATM transactions were profiled individually with manual flags for CFO review.

Sample Risk Findings

The following represents the type of findings the system surfaced (anonymised for this case study):

CategoryFindingRiskAction
ATM Withdrawals4 withdrawals in 9 days, total £840, no notesHIGHCFO review required
Velocity11 transactions to same restaurant in 3 weeksHIGHManager investigation
Benford's LawSignificant deviation in £40–£50 range (near limit)MEDPolicy limit review
Note Quality3 cardholders averaging <2/10 note scoreMEDTraining required
Mystery MerchantUnidentifiable merchant, £220, no category matchHIGHReceipt requested
Cross-employee2 employees, same unusual vendor, different datesMEDMonitoring flagged
The Output
796
Transactions reviewed
48
Cardholders profiled
23
High-risk flags
10
Audit dimensions

The final deliverable was a CFO and audit committee report — structured with an executive summary, risk heat map, individual cardholder profiles, and a recommended action list. Every high-risk item was traceable back to a specific transaction with evidence.

The Impact

The audit directly informed changes to the company's expense policy — stricter note requirements, lower ATM limits, and a new approval layer for transactions above £150. These changes were implemented within 4 weeks of the report being delivered.

What I Learned
PythonOpenAI APIPandasExcelBenford's LawForensic AuditRisk ScoringGPT-4o
CFO P&L App ← Back to Portfolio Power Automate