Leaking Millions, Manufacturers Enlist AI to Stop Supply-Chain Fraud
- Nearly $350 million are lost worldwide to supply-chain fraud in manufacturing companies.
- Examples of supply-chain fraud include duplicate invoices, parts delivered that do not match specifications, and kickbacks from vendors.
- By using artificial intelligence (AI) and machine learning, computers can learn to spot supply-chain fraud by using massive amounts of data.
A typical corporation loses 5% of its annual revenue to fraud. That’s a painful truth—more so because the perpetrator is often a trusted employee who knows the company’s vulnerabilities. For manufacturers, many points targeted by fraudsters lie in the corporate supply chain.
Common Supply-Chain Fraud Scenarios
Fraud poses a threat throughout a manufacturer’s procurement process. Materials come to the shop floor and then disappear. There are duplicate purchase orders or duplicate invoices for the same order. Fraud appears in labor: people signing in but not showing up. Fraud in employee travel expenses is endemic across industries. Committing fraud is often collaborative between an employee requisitioning products or services and a vendor, with the employee getting a cash kickback for allowing the vendor to overcharge.
“Every company has a dollar amount above which every purchase has to go through a thorough investigation,” says Utkarsh Kansal, product manager for Falcon Assurance Navigator (FAN), a procurement fraud-detection system marketed by FICO in San Jose, CA. “The vendor and the purchaser agree to split the requisitions so that each invoice comes in under the threshold for scrutiny. That’s a common fraud scenario.”
Of the $35 billion in annual procurement among manufacturers globally, 0.5%–1% (up to $350 million) is lost to fraud, Kansal says.
Fraud may even affect manufacturing parts or ingredients, says Tim Shinbara, chief technology officer of the Association for Manufacturing Technology (AMT), a trade association. “If the part delivered is passed off as being within the original specifications, but it actually isn’t compliant, that may be a deliberate misrepresentation,” Shinbara says.
How Machine Learning Can Help Risk Management
An experienced human expert can sense, through process or intuition, when a deal looks wrong. But the sheer volume of transactions makes AI a better solution. A manufacturer may have tens of millions of supply-chain purchases yearly—far too many for a risk-management team to monitor accurately but which an AI model can track and analyze.
In banking and financial services, fraud experts have been applying AI to fraud detection for decades; only recently have manufacturers applied these technologies to supply-chain transactions.
Fraud-detection systems are based on machine-learning models—systems that are not programmed conventionally but “trained” to discover patterns in large volumes of data. The data consist of tens or hundreds of thousands of procurement records, some known to be associated with fraud.
FICO’s product is based on Falcon, a neural network–based fraud-detection model. FAN embeds the model in a platform that monitors procurement. It integrates with enterprise resource planning (ERP) systems, sprawling software packages that monitor corporate supply chains and report to management on costs, capacity, production scheduling, inventory, sales, shipping, and just about everything else. Purchase requisitions, contracts, purchase orders, invoices, expense reports, and other documents flow through the ERP system.
The conventional process starts with the purchase requisition, purchase-order issuance, and the receipt of an invoice. A lot of procurement happens through purchasing cards—credit cards authorized for procurement transactions up to a cash limit.
AI Training Methods to Identify Fraud
Although there are predictable, recurring procurement scams, fraudsters consistently seek new angles. AI offers two distinct approaches to detecting and flagging fraud: Data scientists call them supervised and unsupervised models.
Fraud experts train the supervised models to look for known scams. The supervised model is exposed to a training set of procurement records (perhaps three months of transactions), including records for normal transactions and fraudulent ones. The AI is directed to look for specific “features” of the transactions relevant to identifying fraud. There may be hundreds of relevant features or only a few dozen. The model generates a score, much like a credit score, indicating the likelihood that a new transaction is fraudulent.
Once trained and tested, the model can analyze real transactions. If it sees something suspicious, the model triggers an alert for a human fraud expert to take a closer look and start an investigation if needed.
Frauds evolve. Even known scams change subtly over time, making them harder to spot. A human auditor may begin to see transactions that should have been flagged but that the model gave low fraud-likelihood scores. Supervised models require periodic retraining on fresher data to pick up these evolutions.
Novel fraud techniques crop up regularly, so a complete detection system incorporates a second type of algorithm: an unsupervised model. The unsupervised AI also is trained on a large set of procurement-transaction records—not to look for specific fraud types, but to develop an image of what procurement normally looks like (typical transaction volume, types of deals, cash value, frequency of each type of transaction, and so forth).
After training, the unsupervised model looks for outliers—deals that for unspecified reasons don’t look right. The model scores each new transaction, but this score measures how far the transaction deviates from the norm for this type of purchaser, product, and vendor. Deals with high deviance scores merit a closer look.
Personas and Networks
Machine-learning models can analyze transactions individually, tracking what is normal for each purchasing agent. For very large numbers, however, that may be computationally cumbersome and impractical. Some models group individuals into complex psychographic archetypes, called personas. These models score purchasing activities based on what is normal for the persona, not the individual.
Emerging forms of AI applied to fraud detection include cognitive computing approaches that can spot networks of fraudulent activity, recognizing and mapping links between purchasers and other entities, including known fraud actors.
How Manufacturers Can Implement AI to Stop Supply-Chain Fraud
AI developers are taking different approaches toward supply-chain fraud detection. FICO has applied machine learning to fraud detection since the 1990s. Newer vendors have focused on niche applications, such as Atlanta-based Oversight’s software to track employee-purchase-card abuse. Other players, such as Inspectorio (Minneapolis) and Sight Machine (San Francisco) monitor manufacturing-specific processes and might have usage scenarios in procurement fraud.
“A lot of the AI vendors in this relatively new space come from the pick-and-place robotics world and are looking for applications further upstream in the manufacturing process,” Shinbara says.
One potential obstacle is operational scale. Small or midsize enterprises comprise a large proportion of the world’s manufacturing capacity. An AI project usually represents a six- to eight-figure investment—feasible for a large manufacturer with a global supply chain but beyond the means of a small shop.
But there may be a small-scale alternative. A custom model typically is built from the corporation’s own transaction data; a large manufacturer will have enough transactions to train an effective model. In other industries such as financial services, AI vendors have developed so-called “consortium models” from data collected from multiple, comparable companies. Smaller institutions can share a consortium model, hosted in the cloud, at a relatively modest cost. For midsize manufacturers with similar procurement processes, such a shared model could open the door to AI-based fraud detection.
This article has been updated. It originally published in November 2019.