How lenders are running real loans on AI agents

Earlier this year, our customers began using our AI agents in production, on real loans, with real borrowers, in the same LOS their teams use every day.

Over the last three months, usage has grown 988%. Agents are scrubbing new applications, verifying employment, reviewing credit, making underwriting decisions, clearing conditions, and reviewing closing packages, with teams stepping in to handle escalations.

Why does usage matter? It's proof they're working. Our customers have no commitment or obligation to use our agents. It isn't us selling an idea or making promises about what agents will be able to do someday. It's customers trying an agent, seeing it work, and trying more. Many, many more.

Here's what we're seeing across those customers.

Where lenders are starting

Most lenders start where the workflow is most contained (e.g. refi): fewer outside dependencies, less branching, data largely already in hand. That makes it the fastest, lowest-risk way to prove what agents can do.

From there they expand, taking on more tasks and more loan types as each one proves out. The sequencing is practical: start with the work that returns the most and that the team can stand up fastest, and let each agent that holds up make the case for the next.

How lenders bring agents online

The most useful reframe we've seen from customers is that bringing an agent online looks more like training a person than deploying software.

Teams instruct agents in plain English. They tell an agent what to do, how to do it, and what it can and can't use, and they tell it when to escalate to a person.

Anyone can do this. In practice, teams give the job to the people who already document how the work gets done: business analysts and product owners. Zero coding is required.

And all of it happens inside Vesta. There's no separate system to integrate, coordinate, or debug when something breaks. Almost any task that can go to a person can go to an agent instead, using the same workflow, the same permissions, and the same queue.

What's proving out in production

Three things have become clear to lenders now that agents are doing real volume:

Capacity is becoming elastic. Mortgage demand rises and falls with interest rates, which has always meant hiring whiplash for lenders. As more of the work runs on agents, a lender leans less on hiring and cutting to match those swings.

Accuracy improves. Lenders report that agents are more accurate than people. An agent follows the rules, doesn't get tired, and doesn't cut corners at the end of a long day. It can also retain far more context about a loan than a person can hold at once: every applicable guideline, every document in the file, and every decision made so far.

Every decision is traceable. Each action an agent takes is logged with its reasoning and the documents behind it. People don't always leave good notes; an agent leaves a complete record every time. If an examiner asks how a loan was handled, the answer is already in the file.

Where lenders set the limits

Lenders decide what their agents do. They set what each agent can touch, what it can act on, and when it has to stop and hand a loan to a person.

A lot of where that line falls is policy, not technology. An agent is capable of declining a loan on its own, but most operators won't allow that. They want a person reviewing every adverse decision, so they set the boundary there. It's a decision the operator makes, not a limit of the technology.

Where this goes

All of this work has had an unexpected impact on the people doing it.

Teams that have spent years working under the constraints of legacy systems are genuinely energized about what they can do with agents. The work is finally in their hands. Much of what once required an engineering project, an operator can now do themselves: describe a task to an agent and put it into production directly, instead of filing a request and waiting in line.

For the first time, the people closest to the work are the ones setting the pace. Not vendors, not engineering roadmaps, but the people who run loan operations. They decide how much work to hand to the agents, watch and assess what they think the agents are capable of, and iterate and improve their own factories. When we close the gap between the frontier of technology and the people with industry expertise, our customers reach new levels of speed, efficiency, and excitement about the future.