How to automate mortgage document fraud detection using AI
Detect mortgage fraud that is invisible to the human eye.
Document fraud detection is one of the growing applications of AI in mortgage operations.
In this issue, I’ll show you how you can apply AI fraud detection to your mortgage operations.
I designed a solution to demonstrate how to automate the detecting of fake mortgage documents and fraud using one of the emerging AI fraud detection tech providers.
This solution design consists of the 3 parts:
- Business case → What problems AI fraud detection can solve in mortgage operations
- Solution → How to automate mortgage document fraud detection using AI
- Tech providers → What AI tech providers are available, and what’s my choice
I hope it will give you enough insight into AI fraud detection to gauge if it is something that can benefit your mortgage operations and how to apply it if it is.
Business case for AI fraud detection in mortgage operations
It is easy to get caught up in all the shiny innovations and waste resources on a tech that might not add value to the company.
So, before diving into a solution, I think it’s crucial to understand what problems in mortgage operations we’re trying to solve.
Below, you can find my analysis of the following:
- What job do we need to get done, and why it matters
- The old way of document fraud detection and its limitations
- The new way of document fraud detection and its advantages
What job needs to be done with AI
Mortgage lenders receive thousands of loan applications.
They make lending decisions based on the documents representing applicants’ employment, income, assets, and debts.
The job to be done is to ensure that documents provided by the borrower are authentic to prevent fraudsters from getting loans they don’t qualify for.
The document types that we need to identify the authenticity of:
- Identity documents
- Bank statements
- Pay slips
- W2s
Why document fraud detection even matters
Mortgage documents are highly susceptible to fraud. Fraudsters can inflate their bank balance and hide evidence of bad spending to access larger loan amounts.
When loans are given to those who don’t qualify, they possess a significantly higher risk of defaults and buybacks.
So, if a lender or broker is negligent in preventing document fraud, it results in the following:
- Financial losses (defaults, buybacks)
- Increased Financing, Insurance, and Compliance Costs
- Loss of business relationship
- Legal issues
What’s the old way of document fraud detection
Like most processes worth automating, there’s already a manual solution in place to get the job done.
From my research, the old way of mortgage fraud detection works like this:
- First-line mortgage workers review documents and try to assess the authenticity of documents
- If they suspect fraud, they escalate the case for investigation either by the Financial crime department (if available) or by peers
- Financial crime departments or peers rely on their expertise and experience to judge whether this document is authentic.
- To confirm the authenticity of documents, teams highly rely on outreach to the issuing institution of the documents.
What are the limitations of manual document review
The manual document review process has several limitations:
- It is limited only to what the human eye can see.
- It is highly subjective and dependent on the expertise of the person reviewing the document.
- Initial review and investigation in case of escalation takes a lot of time.
- It’s a people-intensive process that’s hard to scale up/down to respond to volume fluctuations.
These limitations lead to significant operational challenges, including:
- Fake documents being undetected
- Wasting time on the investigation of authentic documents
- Longer loan cycles and turnaround times
- Added friction and poor borrower experience
- Increased operational costs
What’s a new way of fraud detection
The new way of fraud detection is to leverage AI to automate the process.
AI fraud detection products run 100s of checks on every document. And in seconds, they give you a risk score and reasoning for the score.
It’s hard to say what exactly checks the fraud detection tech runs as each tech provider might have different approaches.
But there are common layers of checks that I think almost every fraud detection product relies on:
- AI document tampering detection
- ML fraud detection
- 3d party validation
AI document tampering detection uses computer vision to identify forged pixels in documents that might be invisible to the human eye.
ML fraud detection relies on extracting data from the documents and running it through models trained on historical fraud data to make predictions.
3d party validation relies on cross-validating details extracted from the documents with data from their databases and from 3d party providers.
What are the advantages of AI fraud detection
AI fraud detection unlocks a few new capabilities that weren’t possible with the old approach:
- Detect fraud that’s invisible to the human eye
- Systematicly run 100s of objective checks on every document.
- Get risk analysis within seconds
- Detect fraud without human intervention
- Scale up/down without dependency on human resources
These new capabilities drive positive business outcomes like:
- More fraud detected (the ones that were missed before)
- Less time is wasted on the investigation of authentic documents
- Faster initial review and case investigation time
- Better borrower experience
- Lower labor costs
Solution to automate mortgage documents fraud detection using AI
With WHY we’re doing what we’re doing out of the way, let’s dive into HOW to do it.
Below is an overview of the automated mortgage fraud detection solution using AI.
Each section below has a more detailed dive into each solution node.
Upstream integration
The first problem to solve is that our fraud detection system has no access to documents.
Since we have no automation without documents, the first step of the solution is pulling files into our system.
You can usually pull all the documents you need via API from LOS or POS.
Sometimes, you might need to capture documents from an Email inbox, DropBox folder, or FTP server folder.
Document classification
To process documents, our systems need:
- Have documents be in single-document files
- Know what document each single-document file contains
The problem is that we obtain Files from upstream integration. And Files aren’t the same thing as Documents.
It’s common for the file to contain multiple documents within.
The solution is to run files through a document classification system that will provide a list of classified single-document files ready for processing.
You can find more info about document classification in this post I wrote.
Document routing
Though fraud detection with AI is much more cost-effective than manual review, it still costs money.
Yeah, that sucks, I know.
Since most of us don’t have unlimited money, we need to be mindful and not throw whatever documents we’ve got into the fraud detection node.
So, using the data from document classification, we’ll route for processing only documents that are most susceptible to fraud.
Document analysis and risk scoring
That’s the part where all the magic happens.
In this step, we feed the routed documents into the AI fraud detection tech (Resistant.ai, in my case).
Note: You can find more about my product choice for this solution in the section below.
Interaction through the API is pretty straightforward:
- Create document processing request
- Upload the file into the request
- Wait
- Get results of the analysis
Resistant.ai runs 500+ checks in under 30 seconds and provides a Risk score and Indicators as an output.
Risk score reflects where the document sits on the scale from Trusted to High Risk. Risk sore value might be one of these 4: Trusted, Normal, Warning, High Risk.
- Trusted → system identified a reason to trust them.
- Regular → system has no particular reason to trust or distrust
- Warning → system has identified likely modifications
- High risk → system has identified modifications that are associated with fraud.
Indicators contain the reasoning of the system for the score it set. Some indicators may be positive, while others may be negative.
- Positive indicators would increase the Trust
- Negative indicators would increase the Risk
Then, the Risk score and Indicators are fed into the next node of the solution.
Decision automation
The output from Resistant.ai is a document risk score and the reasoning for this score.
Humans are still the ones who make the decision whether the document is authentic or fake.
The vast majority of the documents going to be authentic. So, it might make sense to automate decisions for clear-cut cases to save time.
In my solution, I automatically label a document as authentic if it has a Trusted or Normal score and as fraud, if it has a score equal to High Risk.
If the document has a score equal to Warning, I escalate the case and forward the document for further human investigation.
Case investigation
Once a document is flagged for investigation, someone must review it and decide whether it is authentic or fraudulent.
We display the indicators contributing to the document’s score to reduce the time spent on the investigation.
Reasons to trust the documents and reasons why it might be a fraud.
Most of the tech providers have their own UI for this purpose.
I think that in most cases, the information provided by Indicators should be enough to make a decision without being an expert in document fraud detection.
Downstream integration
After either an automated decision or a case investigation, we know whether the document is fake or not.
The last step in the solution is to let downstream systems know this information so they can act accordingly.
It can be adding a tag to a file within a LOS, creating a ticket, or sending an automatic notification to a relevant person.
Tech providers for AI mortgage document fraud detection
I designed this solution around an AI fraud detection product. Below is my analysis of the tech providers available on the market and my preferred choice for this solution.
Tech providers relevant to mortgage lenders
You always have the option to build your fraud detection tech using low-level infrastructure providers like AWS, GoogleCloude, or Azure. Or buy one from the product company.
In my opinion, in most cases, it doesn’t make sense for mortgage lenders to build custom fraud detection tech.
It is a highly sophisticated piece of software. It will cost more, take more time, and have lower accuracy than buying from a tech provider.
So, when selecting tech providers for this solution, I only focused on those offering complete AI fraud detection products.
Here’s the most relevant AI fraud detection tech for mortgage lenders that I came across when doing research for this solution:
My choice of product for this solution
I think any of the 4 products listed above will be able to meet your AI fraud detection needs.
When designing this solution, my eye fell on the Document Forensics product from Resistant.ai.
Though I haven’t been on a call with them, from reviewing their website and documentation, these are the things that contributed to my decision:
- Publicly available API documentation
- Easy to implement Rest API interface
- Fraud detection is the sole focus (I expect higher quality, lower cost)
- Relevant case studies to the mortgage industry
- It only takes 50–80 documents to train the new detection model
- Offer other fraud solutions like behavior analysis and transaction analysis
What’s next?
I hope this post helped you understand how to use AI to automate mortgage document fraud detection.
If you’d like to stay on top of the latest mortgage tech and how it can be applied to mortgage operations, consider joining our mortgage technology newsletter.
