AI is finally fixing the hard parts of mortgage origination

Mortgage origination has always been a data problem, disguised as a people and paper problem. Dozens of people — loan officers, processors, underwriters, closers — all move information through a process that was never truly designed for them. Every day, they re-enter the same data into different systems, resolve the same exceptions, and reconcile the same PDFs.

‍

That friction has a cost. According to Freddie Mac, the average cost to originate a loan has climbed more than 35% in the past three years, now exceeding $11,600 per loan. Most of that isn’t pricing, compliance, or rate volatility — it’s labor.

‍

AI is finally attacking the right part of the problem

For the first time, AI systems can interpret and structure the data that drives origination. Models can read income documents, extract paystub fields with human-level accuracy, triangulate data across sources, and reason about whether two data points refer to the same borrower or asset.

‍

That’s the quiet revolution happening right now: turning unstructured mortgage data into structured, trustworthy inputs that workflow systems can actually automate the full process.

‍

This isn’t just about slapping a foundation model on top of a mortgage LOS. It requires building out specialized AI pipelines, orchestrating multiple models for their various strengths, architecting the LOS to be used by models instead of humans, and a full blown stack of evaluation and observability tools.

‍

When you use AI to perfect the loan data, downstream automation becomes straightforward. Every condition can be cleared automatically, every income calculation can be verified in real time, every step of the workflow can be tracked end-to-end.

‍

The new foundation lenders need

Unfortunately, AI doesn’t just implement itself. Here’s the part that’s less glamorous but more important: AI only works if the infrastructure beneath it does.

‍

You can’t bolt an intelligent system onto a 20-year-old LOS and expect it to pick up the right context, leverage the LOS’s archaic tools, or feed the clean feedback loops you need for agents to keep improving.

To make AI effective, lenders need:

• Unified, logical, and well-annotated data models so a non-domain specific model can understand and make sense of the data and of each loan
• Robust rules and workflow engines that trigger automation dynamically and tighten the “search space” a model has to traverse to understand what the next steps might be
• Observability and versioning so every model’s output is traceable and explainable to compliance.
• Functions and tools designed for agents, not for people so that each activity in the origination process can actually be invoked programmatically, not just by a person clicking a button

Those are architectural problems — and they’re solvable. But they require treating the LOS not as a UI, but as a data and workflow platform capable of supporting intelligent agents over time.

‍

Why this matters now

For years, lenders invested in digital front ends and borrower experience. A lot of the hope was that by structuring the data at the front of the process, gathering verified data from many sources, and avoiding documents, we could achieve the promised land described above: perfect data makes automation simple.

‍

But very few loans ever realized this vision. There is so much data to collect, and this automated vision requires so much perfection, that maybe a handful of loans have come in digitally, hit on every single verified data source, had clean data from the initial intake, been able to waive the collateral coordination via GSE programs, and realized this automated vision. And every loan with even a single manual step then needs a human who is used to underwriting the entire loan to look at it, creating swirl, double checking of the system, and a lot of extra cost.

‍

Intelligence on tap, available for single digit dollars per million tokens, changes this equation. Rather than painstakingly trying to upgrade the entire ecosystem and change consumer and salesperson behavior, we can just let AI execute the hundreds of small operational steps that never got automated because the data wasn’t structured or consistent enough. The models are more capable at this rote, long-tail work than most people, and this will unlock the holy grail of origination: a completely touchless process.

‍

Where this is heading

In the short term, AI will compress complexity — reducing hundreds of small, manual judgments into a few streamlined, system-driven decisions that humans review. Teams will still oversee the process, but far fewer hands will touch each loan.

‍

Over the next few years, that oversight itself will evolve. Once the system can reliably read every document, validate every field, and clear conditions automatically, the human fulfillment layer disappears. The LOS becomes a closed-loop system — data in, loan out — with human expertise applied only where true exceptions arise.

That’s where we’re heading: a world where origination runs end-to-end on software, and the humans who once moved files now design, monitor, and improve the systems that do.

‍