featured_image: https://cdn.prod.website-files.com/6552eaa01d33c702ff7d6daa/69940e0020bc02a20b9a3440_67a21fbfb8101dfe096d8c58_AdvisorLabs-Banking-Graphic.jpeg

How One Credit Union Boosted Loan Processing Volume 70% With Custom AI

Forum Credit Union, a $2.3 billion institution based in Indianapolis, had a problem that will sound familiar to most credit union executives. Loan volume was growing. Staff wasn't. And every attempt to speed up processing hit the same wall: manual document handling, inconsistent data extraction, and underwriting workflows that depended on institutional knowledge locked inside people's heads.

Their solution wasn't a new SaaS lending platform. It was custom AI, built around Forum's specific loan products, document formats, and underwriting policies. The result was a 70% increase in loan processing volume without adding headcount.

This article breaks down what that actually looked like, why off-the-shelf tools couldn't get them there, and what credit union leaders should understand about AI process automation in lending before investing in the wrong solution.

The Problem With Loan Processing at Credit Unions

Credit unions process loans differently than banks. That's not a marketing statement. It's an operational reality that shapes every technology decision you make.

Your loan products have unique terms, rate structures, and eligibility criteria. Your underwriting policies reflect your field of membership, your risk appetite, and decades of institutional learning. Your document requirements vary by loan type, and the formats members submit range from pristine PDFs to photos of crumpled pay stubs taken on a phone.

Then there's the core system. Whether you run Symitar, Corelation, DNA, or something else entirely, your core is the gravitational center of your technology stack. Every loan touches it. Every integration has to work with it. And every vendor who promises a seamless connection is, at best, oversimplifying.

Forum Credit Union lived this reality every day. Their loan officers spent hours on document classification alone, sorting through income verification, tax returns, bank statements, and employment letters to determine what they had, what they needed, and what format it was in. Data extraction was manual. An officer would open a document, find the relevant numbers, and key them into the system. Underwriting decisions required pulling information from multiple sources and applying policies that varied by loan type.

None of this was broken. It worked. But it didn't scale. And when loan demand increased, the only option was to hire more people or accept longer processing times. Neither was sustainable.

Why Off-the-Shelf AI Tools Fall Short

The obvious answer to this problem is to buy a SaaS AI lending platform. Dozens of vendors offer them. They promise faster processing, automated document handling, and AI-powered underwriting. Some of them are genuinely good products.

But here's what credit union executives discover when they actually try to implement one: the tool doesn't know your business.

A SaaS lending platform is built for the broadest possible market. It handles standard document types in standard formats with standard underwriting logic. That works if your processes are standard. At most credit unions, they aren't.

Forum Credit Union ran into three specific problems when evaluating off-the-shelf solutions.

Integration with legacy systems. Forum's core system had specific data structures, API limitations, and workflow requirements. SaaS tools expected data in their format, not Forum's. Getting the two to talk to each other wasn't a configuration change. It was a custom development project, and the SaaS vendor wasn't going to do it. This is one of the most common failures in credit union AI implementation - the integration piece gets underestimated every time.

Document format variability. Forum's members submitted documents in dozens of formats. Their internal documents had specific layouts that had evolved over years. A generic document classification model could handle a standard W-2, but it struggled with Forum's specific loan application forms, their particular bank statement layouts from regional institutions, and the handwritten notes that loan officers added to files.

Underwriting policy alignment. Forum's underwriting policies weren't a simple decision tree. They involved conditional logic based on loan type, member history, collateral type, and dozens of other variables that reflected Forum's specific risk philosophy. A SaaS tool's underwriting engine would need extensive customization to match, and most vendors don't offer that level of configurability. The regulatory dimension adds another layer of complexity - underwriting decisions must align with fair lending requirements and NCUA expectations, a topic we cover in depth in our guide to NCUA-compliant AI.

This isn't a knock on SaaS platforms. They solve real problems for institutions whose processes align with what the software expects. But when your processes don't align, you're left with two bad options: change your processes to fit the tool, or spend months on custom integration work that the vendor isn't designed to support.

Forum chose a third option.

What Custom AI Actually Looks Like

When we talk about custom AI for credit union lending automation, we're not talking about building a general-purpose lending platform from scratch. We're talking about building targeted AI models that do specific things within your existing workflows.

For Forum Credit Union, the project focused on three areas.

Automated Document Classification

The first problem to solve was sorting. When a member submits a loan application, it often arrives as a single PDF containing multiple document types, or as a batch of separate files with inconsistent naming. A loan officer might receive a 40-page PDF that contains pay stubs, tax returns, a driver's license, bank statements, and a signed application, all in one file.

The custom AI model was trained on Forum's specific document types. Not just generic categories like "income verification" or "identity document," but Forum's exact document taxonomy, including internal forms, specific third-party documents they commonly receive, and regional variations.

The model learned to classify documents with over 95% accuracy across Forum's document types. When a new loan application arrived, documents were automatically sorted, labeled, and routed to the correct stage of the workflow. What used to take a loan officer 15 to 20 minutes per application now happened in seconds.

Intelligent Data Extraction

Classification tells you what a document is. Extraction tells you what's in it. This is where custom AI process automation delivers its biggest advantage in credit union lending.

A generic OCR or extraction tool can pull text from a document. But pulling the right data from the right fields and mapping it to the right places in your system requires understanding your specific document layouts and data requirements.

Forum's custom extraction models were trained on their actual documents. The AI learned where to find gross income on the specific pay stub formats their members commonly submitted. It learned how to handle the variations in tax return layouts across different years. It learned to extract property values from appraisal documents in the specific formats Forum's approved appraisers used.

Critically, the extraction models were built to map directly to Forum's core system fields. Data didn't need to be reformatted or manually transferred. It flowed from the document into the system in the format the system expected, reducing both processing time and data entry errors.

Accelerated Underwriting Decisions

Document classification and data extraction are the foundation. Underwriting acceleration is where the real volume improvement comes from.

Forum's custom AI didn't replace underwriters. It prepared everything an underwriter needed to make a decision and flagged applications that met specific criteria for expedited review.

The system applied Forum's actual underwriting policies, not a generic risk model. It calculated debt-to-income ratios using Forum's specific methodology. It evaluated collateral using Forum's valuation standards. It flagged exceptions based on Forum's exception criteria, not industry defaults.

For straightforward applications that met all policy requirements, the AI prepared a complete underwriting package with a recommendation. The underwriter reviewed the package, confirmed the recommendation, and approved the loan. What used to take 45 minutes of analysis and data gathering took five minutes of review.

For complex applications, the AI still did the heavy lifting of document processing and data extraction. It flagged the specific areas that needed human judgment, so the underwriter could focus their expertise where it mattered instead of spending time on data entry and document hunting.

Why the 70% Improvement Happened

The 70% increase in loan processing volume wasn't one dramatic improvement. It was the compound effect of eliminating friction at every stage of the process.

Document classification went from 15 to 20 minutes to seconds. Data extraction went from manual entry to automated population with human verification. Underwriting preparation went from 45 minutes of gathering and calculating to five minutes of reviewing a pre-built package.

Multiply those savings across every loan application, every day, across the entire lending team. The math adds up fast.

But the reason the improvement was 70% and not 10% or 15% comes down to one factor: the AI was built for Forum's specific workflows. It didn't require loan officers to change how they worked. It didn't force data into a different format. It didn't apply generic logic that needed manual overrides.

It fit. And because it fit, adoption was immediate. Loan officers didn't resist it because it didn't disrupt their process. It just removed the tedious parts.

This is the core argument for custom AI loan processing: the improvement is proportional to how well the solution matches your actual operations. A generic tool might give you a 10% to 15% improvement because it only addresses the parts of your process that happen to match its design. A custom solution addresses all of it.

What This Means for Your Credit Union

Every credit union's loan processing is different. That's not a problem to solve. It's the result of decades of building products and policies that serve your specific membership.

Your loan products have terms and structures that reflect your members' needs. Your underwriting policies balance risk and access in ways that align with your mission. Your document workflows have evolved to handle the reality of what your members actually submit. Your core system - whether it's Symitar, Corelation, DNA, or another platform - has been configured over years to support your specific operations.

A SaaS tool doesn't know any of that. It can't. It's built to serve thousands of institutions, which means it's optimized for none of them.

Custom AI flips that equation. Instead of adapting your processes to fit a tool, you build AI that adapts to your processes. The models train on your documents. The extraction maps to your data fields. The underwriting logic reflects your policies. The integration connects to your core system as it actually exists, not as a vendor wishes it existed.

The Path From Pilot to Production

If you're considering AI process automation for your credit union's lending operations, the path Forum followed is instructive.

Start with a specific problem. Forum didn't try to automate everything at once. They started with document classification because it was the most time-consuming manual step and the easiest to measure. Pick the bottleneck that costs you the most time and start there. For a framework on selecting and scoping your first AI project, see our 90-day implementation guide.

Use your actual data. The custom AI models were trained on Forum's real documents, not synthetic data or industry samples. This is what makes custom AI accurate for your specific use case. Your documents, your formats, your variations.

Build for your core system. Integration isn't an afterthought. The AI should be designed from day one to work with your core system's data structures, APIs, and workflow requirements. This is where most SaaS implementations fail, and where custom solutions succeed.

Keep humans in the loop. Forum's AI didn't replace loan officers or underwriters. It handled the mechanical work so that humans could focus on judgment, member relationships, and exception handling. This approach is both more effective and more practical for regulatory compliance.

Measure everything. Forum tracked processing time per application, error rates, underwriter review time, and total volume before and after deployment. Clear metrics justified the investment and identified areas for further optimization.

The Real Cost of Waiting

Credit unions that delay AI adoption in lending aren't just maintaining the status quo. They're falling behind. Member expectations for speed are increasing. Competing institutions - both credit unions and fintechs - are processing loans faster. And every month you spend manually classifying documents and keying in data is a month your lending team could have spent on higher-value work.

The 70% volume improvement Forum achieved didn't require a massive technology overhaul. It didn't require replacing their core system or retraining their entire staff. It required building AI that understood their specific operations and integrating it into the workflows they already used.

That's the opportunity for AI process automation in credit union lending. Not a rip-and-replace transformation, but a targeted, practical application of AI to the specific bottlenecks in your lending process.

Get Started With a Pilot

Every credit union's path to AI-powered loan processing will look different because every credit union's operations are different. That's exactly the point.

The right first step is a focused conversation about your specific workflows, your core system, your document types, and your underwriting policies. From there, you can scope a pilot that targets your biggest bottleneck and delivers measurable results within weeks, not months.

Book a free consultation to scope an AI loan processing pilot built around your credit union's specific workflows. We'll map your current process, identify the highest-impact automation opportunities, and outline a pilot that fits your technology environment and your budget.

Your members are waiting. Your lending team is ready. The technology exists. The only question is whether you build it around your credit union, or try to fit your credit union around someone else's software.

Frequently Asked Questions

How long does it take to implement custom AI for loan processing?

Most credit unions can have a working pilot in production within 60 to 90 days, starting with a single high-impact area like document classification. Full deployment across the lending workflow typically takes three to six months depending on the number of loan products and complexity of your underwriting policies.

Is custom AI more expensive than a SaaS lending platform?

The upfront investment for a custom pilot is often comparable to the annual license cost of a SaaS platform. The difference is that a custom solution delivers higher ROI because it addresses your specific bottlenecks rather than forcing you into a generic workflow. Forum Credit Union's 70% volume increase translated directly to revenue growth without adding headcount.

How does custom AI handle regulatory compliance in lending?

Custom AI is built with your compliance requirements baked in from the start. The system applies your institution's underwriting policies, maintains full audit trails for every decision, and keeps humans in the loop for final approvals. This approach aligns with NCUA guidance on responsible AI adoption, which emphasizes explainability and human oversight.

What data do we need to get started?

You need historical loan application documents, your underwriting policy documentation, and access to your core system's data structures. The AI models are trained on your actual documents and workflows, so the more representative the training data, the more accurate the system. Most credit unions have more than enough data from the past 12 to 24 months to build an effective model.

Will our loan officers need to learn a new system?

No. The AI is designed to integrate into your existing workflows, not replace them. Loan officers continue using the tools and processes they already know. The AI handles the manual, repetitive steps - document sorting, data extraction, package preparation - so officers can focus on member relationships and judgment calls.