Stop arguing about architecture: 5 ways to evaluate AI in your LOS

By Mitch Swanson, Sr. Product Manager

Let me start with something that might sound harsh: the conversation about AI in loan origination systems is broken.

Everyone is debating whether the LOS was “built on AI” or if the vendor “added AI later.” They’re arguing about codebase age, proprietary models, and architectural decisions. It sounds smart. It feels like a shortcut to the truth.

But I’ve been on both sides of this technology. I was a loan processor managing dozens of files at once. Now I lead AI products at MeridianLink^®. And I can tell you with certainty that the architecture argument is not what predicts whether AI will actually help your team close loans faster.

The real evaluation criteria are different. And they’re probably not what your vendor is emphasizing.

The problem with how we’re talking about AI in lending

I started in the mortgage industry in 2012 as a loan officer assistant, then a processor. I know what it feels like to pull conditions at 5 p.m. on a Friday, chase documents from borrowers, and explain to someone for the third time why a bank statement with a missing page doesn’t work.

I’ve lived the specific friction that AI is being positioned to solve. And I know how easy it is to get distracted by the wrong questions.

Today, I spend my time building AI for document processing at MeridianLink®. One thing I’ve learned in that transition is that the marketing conversation around AI has almost nothing to do with what actually determines whether the technology works.

The current debate is about platform architecture. It’s appealingly simple. It sounds like a clean technical shortcut. But it’s not the question that matters.

Here are the five questions that do.

1) Does the AI make your staff faster, or does it just create different work?

This is the question that matters most. And it’s the one almost nobody asks. 

When AI vendors demo their technology, they lead with accuracy metrics. The model correctly identified the document 94% of the time. Precision was 96%. These numbers sound impressive. They’re technically meaningful. They tell you almost nothing about whether the AI will actually make your loan processors faster. 

Think about it: it’s entirely possible to build a highly accurate AI that creates more work, not less.  

Here’s a real scenario:   

A processor gets a document. They need a yes or no answer: is this acceptable or not?  

If the AI accuracy is high, but the output is a multi-paragraph analysis that the processor has to read through to extract the conclusion, you haven’t solved the problem. You’ve moved it. The processor now has to parse AI-generated text instead of looking at the document. Or even worse, they read paragraphs and then still have to look at the document.  

A loan processor managing 30 to 50 active files doesn’t have time for that. They need a clear signal. Accept or reject. Specific reasons. A fast path to the next action. Seconds matter when you’re closing loans by end of day. 

The workflow is where domain expertise actually shows up. Not in which foundation model the vendor is calling, but whether they understand the time pressure, the mental model of the processor, and what matters in that moment. 

You can’t design that from a whiteboard. You have to know the seat. 

2) What happens when the AI is wrong, and can you prove it? 

Fannie Mae’s Lender Letter LL-2026-04 changed the game on this one. And I don’t think it’s gotten enough attention. 

It requires lenders to hold their technology vendors to the same AI governance standards that lenders themselves must meet. 

Let that sink in. Your vendor’s AI governance posture is now your compliance exposure. 

This is not theoretical. Non-deterministic software making suggestions that affect loan files creates a regulatory footprint. You need to prove what the AI did, why it did it, and when a human overrode it. This audit trail is not optional. It’s now required. 

When I was processing loans, compliance was the thing that slowed me down but also kept the institution out of trouble. Every experienced loan processor intuitively understands this tension. AI doesn’t change the tension. It raises the stakes. 

Here’s what to look for when evaluating AI governance: 

• Does the vendor have a formal AI governance body?
  Not a marketing page. An actual committee with a charter, cross-functional membership, and regular meetings.
  
• Is the human-in-the-loop?
  In lending, the AI should suggest. Staff should decide. If the AI acts autonomously on anything that touches a loan file, you need to know exactly where that autonomy boundary is.
  
• Is every AI action logged?
  Can those logs be compiled into an audit-ready report per loan? Do you know how the system reviewed the document, evaluated it, wrote data to the URLA, or when a user overrides the output?
  
• Is the AI scoped away from credit decisioning?
  There’s a meaningful difference between AI that evaluates document quality and AI that influences loan approval. The boundary should be explicit, not implied. Look up Massachusetts vs Earnest Operations LLC if you want to understand why this matters.
  
• Has the framework been reviewed against LL-2026-04 specifically?
  The policies and procedures for development, implementation, use, and maintenance of any AI/ML must have risks mitigated.

These are governance questions, not architecture questions. They favor vendors who have been through regulatory cycles—TRID, the URLA/ULAD transition, years of CFPB and ECOA scrutiny. Not vendors who launched recently with a clean codebase and a thin compliance track record. 

Bring your compliance officer to the demo. The questions they ask will tell you more than any feature comparison. 

3) If they claim proprietary AI, what data trained it? 

Here’s something you hear a lot in lending tech: “We have proprietary AI trained specifically for lending.” 

I build AI for lending. Let me tell you what that claim actually requires. 

Building your own AI model infrastructure requires four core pillars:  

1. Specialized compute (GPUs/TPUs) 
2. High-speed storage (NVMe) 
3. Low-latency networking 
4. And orchestration software 

Whether building a private local server or renting cloud hardware, you need a dedicated hardware and software stack to process data, train the model, and serve inferences.  

For those that decide to purchase the hardware, that could mean 3-7 months to just acquire the GPU hardware. Once they are able to get hands on the hardware, they must be ready for large price tags to achieve the scalability of SaaS. The capital expenditure is measured in hundreds of thousands of dollars. Also, hardware depreciates fast. Keeping up with foundation model improvements means continuous reinvestment. As for renting the GPUs, that price tag could run you into $1M+ per year at scale. 

There’s a reason OpenAI, Anthropic, and Google spend billions on this. The economics only work at horizontal scale serving millions of users. 

For a vertical SaaS company serving one industry? The math is hard to digest. 

Then there’s the data problem. The one that actually keeps me up at night. 

Training a lending-specific model requires a massive corpus of training data. But in regulated lending, consumer consent requirements make it extremely difficult—arguably impossible in many jurisdictions—to use production loan data for model training without explicit borrower consent. And most institutions haven’t collected that consent on loans originated in the past. 

This is not a technical hurdle. It’s a regulatory wall. 

So, what does “proprietary AI” actually mean in this market? 

It’s harder to put “application layer” on a marketing slide than “proprietary AI.” But it’s what determines whether the system actually works when a Processor is staring at a blurry paystub at 5 PM on a Friday. 

Next time someone tells you they have proprietary AI trained on lending data, ask two simple questions. What data was it trained on? And do you have consumer consent for that use? The answers will tell you whether you’re hearing a technical claim or a marketing claim. 

4) How deep is the integration ecosystem? 

When I was processing loans, I never thought about the LOS as a product. It was just the system. I opened it every morning, lived in it all day, and closed it at night. 

What I didn’t appreciate until I crossed over to the product side is how much invisible infrastructure sits behind that system. 

A lender doesn’t just buy an LOS. They buy into an ecosystem. Credit bureaus. AUS engines. Pricing engines. Doc prep. Title. AMC. Fraud detection. CRM. POS. Core banking systems. And hundreds more. 

At MeridianLink, that’s over 300 production-certified partners. 

Each integration represents years of edge-case handling and compliance validation. The kind of work that doesn’t show up in a demo but absolutely shows up when something breaks in production. 

Here’s the thing about AI in lending: it’s the visible tip of the iceberg. 

It’s functionality is what you see in the demo. It’s on the feature comparison slide. But the integration ecosystem, the compliance infrastructure, the multi-product coverage—that sits below the waterline. And when you’re choosing an LOS, what sits below the waterline is what determines whether your operation actually runs. 

I’ve noticed something interesting: when a newer vendor announces a single new integration partnership, it’s headline news. For an established platform, that’s a normal week. That gap tells you something important about where each vendor is in build maturity to support your business. 

So, ask how many production-certified integrations does this platform have? How long did the most recent integration take to build? And does the AI cover your full lending footprint, or will you be maintaining separate systems? 

5) Does the AI cover your full lending operation or just one channel? 

This cost is routinely underestimated. 

When you choose an AI solution that covers only one lending channel, you’re not just choosing which vendor AI to use. You’re choosing to maintain separate systems of record. Separate compliance frameworks. Separate vendor relationships. 

If your institution does consumer and mortgage lending, a platform that only supports one—however impressive its AI—creates a new fragmentation problem.  

The platform decision isn’t just about what the AI can do. It’s about whether the platform can support your full lending footprint as one connected operation. 

Map what you need today and where you expect your lending business to be three to five years from now. Then ask: Can this platform support that vision as one operation, or will it leave you managing separate systems?

The real platform decision 

An LOS isn’t just the technology you use. It’s the foundation everything else sits on. 

The AI layer is modular. It can be upgraded, replaced, and expanded. But the system of record, the compliance infrastructure, the integration ecosystem? Those can’t be easily replaced. Those decisions lock you in for years. 

That’s the real platform decision. Not which vendor has the newest code. But which one has the foundation your AI will run on. 

The five criteria in this article—operational workflow efficiency, compliance governance, data transparency, integration depth, and multi-channel coverage—are designed to help you move past the architecture debate and focus on what actually matters.  

Will this AI solve real problems in your operation? Is the foundation solid enough to support it? 

The answers won’t be obvious from a demo or a feature comparison slide. 

They require digging into workflows. Asking your compliance officer hard questions. Understanding not just what the AI can do, but how the entire platform supports your lending operation. 

Choose the foundation carefully. Everything else sits on top of it. 

See how MeridianLink puts AI to work the way lenders need with a stronger platform foundation and embedded intelligence. 

The materials available in this article are for informational purposes only and not for the purpose of providing legal advice. You should contact your own advisors with questions regarding the AI in lending content herein. The opinions expressed in this article are the opinions of the individual authors and may not reflect the opinions of MeridianLink, Inc.