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.
One in twenty.
That’s how many organizations have successfully moved AI into production.
Financial institutions have been exploring the technology, launching pilots, and testing use cases, but very few have crossed the finish line. While 80% of organizations are experimenting with AI and roughly 40% have launched pilots, only 5% have moved into production.
The question is no longer whether AI works. It’s why so few organizations are realizing its value.
That was the focus of a recent MeridianLink® webinar, Operationalizing AI with Specialized Agents, where Mitchell Swanson, Senior Product Manager, and Brandan Sisola, Director of Product Management, explored what separates successful AI deployments from stalled initiatives.
Their conclusion: Production-ready AI depends less on the model itself and more on where it’s embedded, what data it can access, and how well it’s aligns with real business workflows.
Keep reading for the key takeaways, and catch the full conversation in the webinar recording.
The budget is there. The plan usually isn’t.The budget is there. The plan usually isn’t.
Adoption looks healthy on the surface. About half of banks and nearly two-thirds of credit unions have deployed generative AI in some capacity, and more than 80% of lenders plan to increase their AI budget.
Fewer than one in five have a plan for how they’ll spend it.
That’s the honest state of things. Institutions know they should be using this technology. They’re far less certain where it belongs in their operation. Budget without a defined use case produces pilots that quietly expire.
Enterprise software spending tells the same story from the other direction. Spend on AI-native SaaS grew more than 100% year over year, and nearly 80% of IT leaders are putting more budget specifically toward AI. Your peers are already placing these bets. The question isn’t whether AI arrives in lending—it’s whether your foundation is ready when it does.
Point it at the most expensive bottleneck
Ambition without a target is where most of this goes sideways. So start where the money is.
In mortgage, it costs roughly $12,000 on average to originate a loan. The spread between lenders is the more interesting number: about $16,500 per loan in the bottom quartile against $6,900 in the top. That’s a $9,600 gap between institutions doing the same work. Scale explains part of it. Automation explains a lot more.
Asked to name the single biggest bottleneck driving that cost, MeridianLink customers voted for document chaos.
The loop is familiar. An underwriter writes a condition in technical shorthand; the borrower reads it and has no idea what’s being asked. Phone calls come in to the processor and the loan officer. Staff spend hours rewriting requests into plain language, each person doing it a little differently. The borrower uploads the wrong document, or an incomplete one. Then, the rework begins, and the loop runs again.
Every pass costs staff time and borrower patience. It’s a well-defined, high-volume, expensive problem, which makes it a far better first AI target than a general-purpose assistant nobody quite knows how to use.
Why bolt-on AI in lending stalls
Plenty of vendors will sell you an answer to this. Most of them are wrappers.
Over the past couple of years, a wave of tools has arrived that puts an interface around a general-purpose model and sells it as purpose-built for lending. Those tools can be genuinely useful for narrow jobs—a customer service chatbot is a reasonable example. What they can’t do is see your application.
A bolt-on tool sitting on top of your loan origination system might receive a data payload, but it doesn’t have the full context of the application model. It doesn’t know where in the lifecycle the loan sits, what’s missing, or how your institution has configured its workflow. It answers the question in front of it without understanding the file it belongs to.
There’s a longer-term cost, too. If your technology partner isn’t investing in the architecture that agents will eventually run on, you inherit a tech stack expiration date. Competitors building on that foundation pull ahead on automation while you’re still integrating point solutions.
Designed around the way lenders work
That architectural work is what MeridianLink Intelligence is built on. It’s a family of AI agents (internally, we call it Millie) embedded natively inside our loan origination system, with no new integrations to stand up.
The first agent is Doc Agent, and it takes on the document chaos problem in three places.
- It translates conditions into borrower-friendly requests. Instead of a processor rewriting an underwriter’s jargon by hand, Doc Agent generates the document request automatically and filters out the conditions that don’t concern the borrower at all. If the condition says the loan officer needs to sign the 1003, the borrower never sees it. The processor reviews the output, edits anything they want, sees the reasoning behind each decision, and sends with one click.
- It pre-screens documents on arrival. Every uploaded document gets checked for relevance, completeness, and recency before a human opens it. Is this a pay stub or a photo of the family dog? Are all the pages of the bank statement here? Is this last month’s or last year’s? The processor gets a readout the moment they open the file: acceptable, needs review, or reject, with the reason attached. It runs overnight, so a document uploaded at 2 a.m. is already screened by morning.
- It extracts and compares the data. This is where it separates from OCR. Optical character recognition reads pixels and reports what it sees. Doc Agent reads the document, understands what the values mean, and compares them against what’s in the application. A bank statement showing $6,600 against an app record of $10,000 surfaces immediately, side by side, along with the address fields nobody filled in. One click updates the 1003. Every change writes to a log and into the standard audit history, so you end up with more visibility per document than manual entry provides today.
Consumer lending is similar, but with a different interface. Most administrators already configure readable stipulations, so condition translation isn’t the gap there—document extraction and comparison are, and that work is underway now.
Responsible AI starts with control
In a regulated environment, speed and trust have to go hand in hand. That’s why the rollout is deliberately staged.
Doc Agent keeps a human in the loop. It surfaces what it found and recommends what to apply, and a user decides. Full automation is the direction, and institutions get to move at a pace their teams and examiners can accept.
The infrastructure sits inside the same environment as your other MeridianLink products, so your data and your customers’ data stay in the bubble you already trust. Guardrails address hallucination, jailbreaking, manipulation, and bias, and an internal AI governance council reviews what gets built. With state-level AI regulation expanding, drawing a clear line between what the model recommends and what a person decides isn’t optional.
What’s coming
Doc Agent is in beta testing now, with general availability for MeridianLink® Mortgage in Q4. Consumer is in development, with rollout aimed at early 2027 across direct, indirect, and the rest of your loan types.
After that, the agents start working together. A task agent will read Doc Agent’s extraction and act on it. For example, it may spot a large deposit on a pay stub, create the letter of explanation condition automatically, attach the document when it arrives, and clear the condition if you’ve permissioned it to. A virtual loan officer assistant will handle borrower status questions inside the point of sale.
Be the first to see what’s next as new AI agents roll out across our MeridianLink® One platform.
