How to Detect Expense
Report Fraud with AI
Expense report fraud is the most common form of occupational fraud in small and medium businesses. According to the ACFE Report to the Nations, it accounts for 21% of fraud cases — and the median loss per incident is approximately US$26,000 (around S$35,000). For a lean Singapore SMB, a single case can represent months of profit.
The uncomfortable truth is that manual expense review is nearly incapable of catching sophisticated fraud. Human reviewers are fast, but they are also consistent in exactly the wrong ways: they check the obvious things (does the amount look right?), miss the subtle ones (was this receipt printed twice?), and get fatigued long before they finish the pile.
AI expense fraud detection works differently. It does not get tired, does not rely on memory, and can check multiple signals simultaneously across the entire history of claims — not just the one in front of it. Here is how it works in practice.
The Five Most Common Types of Expense Fraud
Before explaining how AI detects fraud, it helps to understand what it is looking for. These five patterns account for the vast majority of expense fraud cases in SMB environments.
Fictitious Expenses
Submitting a claim for an expense that never occurred — often accompanied by a fabricated or altered receipt. AI vision models are particularly effective here: they detect inconsistencies in font rendering, metadata, and document structure that human reviewers miss at a glance.
Receipt font inconsistency, missing GST registration number, document creation timestamp mismatch
Inflated Expenses
The underlying expense is real, but the amount has been altered — a S$35 lunch receipt becomes S$135. AI extracts the original amount from the invoice image and compares it character by character against the submitted amount. Even minor alterations are flagged.
Amount field mismatch between invoice image and submitted value, amount outlier vs. historical category average
Duplicate Submissions
Submitting the same receipt twice — either in the same period or across different periods. This is surprisingly common and easy to miss manually. AI maintains a hash of every processed receipt and flags exact or near-duplicate submissions immediately.
Receipt hash match to previous submission, same vendor/amount/date combination as approved claim
Personal Expenses Misclassified as Business
Submitting personal expenses — groceries, personal travel, family restaurant meals — as business expenses. These are often the hardest to detect without context. AI uses anomaly detection on category, vendor type, day of week, and submission timing to flag suspicious patterns.
Vendor category mismatch (e.g., supermarket as "client entertainment"), weekend submission for office-hours category, amount significantly above category benchmark
Policy Manipulation
Structuring expenses to stay just under approval thresholds or documentation requirements — for example, submitting three S$48 claims instead of one S$144 claim that would require a receipt. AI detects splitting patterns and unusual frequency from the same employee in the same category.
Multiple same-category claims on the same day, amounts consistently just below documentation threshold
How AI Detects These Patterns
An AI expense verification system like Kopi runs a multi-layer check on every submission before the approver sees anything. Each layer targets a specific category of fraud:
Invoice Authenticity Analysis
AI vision models read every attached receipt or invoice image. They extract text via OCR, check for font consistency, validate that amounts and fields are machine-generated (not hand-edited), and look for markers of legitimate Singapore GST invoices — registered business number, correct tax rate, etc.
Amount Cross-Validation
The extracted invoice amount is compared field by field against the submitted claim amount. A mismatch of any size — even S$1 — generates a flag. The system also compares the claim amount against the employee's historical spending in the same category, surfacing statistical outliers.
Duplicate Detection
Every receipt is fingerprinted with a hash of its key fields (vendor, date, amount, document structure). When a new submission arrives, the hash is compared against all historical claims. Exact duplicates are flagged immediately; near-duplicates (same vendor and amount, different date) are surfaced for review.
Vendor and Category Validation
AI classifies the vendor from the receipt into a business category (meals, transport, SaaS, office supplies, etc.) and checks whether the submitted expense category matches. A receipt from Cold Storage filed as "client entertainment" is an immediate anomaly.
Temporal and Contextual Checks
The AI checks submission timing (weekend expenses in office-hours categories, expenses dated in the future, expenses submitted 60+ days after the date), frequency patterns (unusual submission volume from one employee), and splitting patterns (multiple same-day claims that sum to a round number).
What AI Cannot Do (And Why Humans Still Matter)
AI fraud detection is powerful, but it is not a replacement for human judgement — it is a force multiplier for it. There are cases that AI flags as anomalies that are entirely legitimate: a one-off large client entertainment expense, a team member who recently changed roles and now has different spending patterns, or a vendor whose receipts consistently look unusual because of how they format their invoices.
The goal of AI is not to make the fraud decision — it is to surface the right signals so the human approver can make a faster, better-informed decision. A claim that used to take 4 minutes of manual review can be assessed in 30 seconds when the AI has already done the data work and flagged exactly what is unusual.
This is why Kopi's system always delivers a check card to the human approver — never auto-rejects. The AI's job is to preprocess and surface; the approver's job is to decide. For the 80% of clean, routine claims, the decision is instant. For the 20% with flags, the approver has context they could not otherwise have assembled manually.
Prevention as a Side Effect
One of the underappreciated benefits of AI expense fraud detection is deterrence. Employees who know that every receipt will be vision-scanned, that duplicate submissions are caught, and that unusual amounts are benchmarked against their own history — are substantially less likely to attempt fraud in the first place.
The ACFE consistently reports that most occupational fraud is opportunistic: it happens when people believe they will not be caught. A visible, systematic AI check — one that surfaces its reasoning to the approver — changes the perception of detection risk and reduces the incidence of attempted fraud, independent of how many actual fraud cases it catches.
For Singapore SMBs running expense approvals through Lark, adding an AI layer protects the business on two fronts: it catches the fraud that does occur, and it deters the fraud that might otherwise be attempted. Learn more about how AI expense approval works in Lark, or start a free Kopi account to add fraud detection to your Lark workspace today.
Add AI fraud detection to your Lark expense workflow
Kopi checks every submission for fraud signals before the approver sees it. Free for Singapore SMBs in private beta.