The Race To Reinvent Mortgage Origination

The national average cost to originate a mortgage climbed to $11,800 in the first quarter of 2026, up $800 from the final quarter of 2025, according to the Mortgage Bankers Association’s Quarterly Mortgage Bankers Performance Report.

That figure remains below the record highs of 2023, when production expenses surpassed $13,000 per loan, but it is still well above historical norms. From the first quarter of 2008 through the first quarter of 2024, loan production expenses averaged $7,472.

The industry’s nearly $12,000 problem has no single cause. In a higher-rate environment, lenders are spreading fixed costs across fewer loans. Vendor expenses have also risen, including the price of credit reports, which has increased several times in recent years. But Freddie Mac’s Cost to Originate Study points to a much larger expense: personnel.

Compensation for loan originators, processors, underwriters, and other production employees accounts for the largest share of lenders’ origination costs. That has made labor-intensive parts of the mortgage process a prime target for automation.

Lenders have invested heavily in artificial intelligence and other technology designed to streamline operations and reduce labor costs. Although the industry has made progress automating the borrower-facing application process and certain back-office tasks, core functions such as processing and underwriting remain heavily dependent on people — and expensive because of it.

Now, a new generation of mortgage technology companies is taking aim at that cost structure. Copperlane, Pylon, Ralo, and bevri.ai are pursuing different approaches, from AI loan processors and programmable mortgage infrastructure to direct-to-consumer platforms designed around a nearly push-button experience. Each is betting that advances in AI can finally automate work that older mortgage technology left largely untouched.

Together, these companies are testing whether AI can do what decades of mortgage technology have struggled to accomplish: materially reduce the number of people, systems, and manual steps required to close a loan — and finally put a dent in the industry’s nearly $12,000 cost problem.

Copperlane’s ‘Infinite Processors’

Copperlane co-founders Athan Zhang and Brianna Lin came to mortgage from artificial intelligence, although both grew up with parents who worked in housing finance. Athan’s parents both worked for Fannie Mae and Freddie Mac, while Lin’s father worked for the FHFA, giving them a top-down view of mortgage risk without conditioning them to accept the industry’s existing workflows.

That outsider perspective, Zhang said, allows Copperlane to apply technological advances emerging elsewhere before they become widely used in mortgage.

“We learn from other people in other industries also building AI, and that gives us a lot of insight into what the type of changes we’ll probably see in mortgage, too,” Zhang said. “And so we’re able to implement these things before it even appears in mortgage, and that gives us a big advantage.”

From that vantage point, the founders saw what Zhang described as a significant “data variance problem”: inconsistencies in borrower information that make it harder to assess risk accurately. They also saw the expensive consequences when defects travel downstream, including loans that lenders must repurchase from the secondary market.

When Zhang and Lin examined what caused those defects, they frequently traced them back to routine human errors made while files were being assembled. That led them to Penny, an AI employee designed to automate many of the repetitive tasks traditionally performed by loan processors.

“We realized that a lot of the smaller tasks that loan officers do day-to-day could almost easily be automated with an agent like Penny, who understands the context of the borrower [and] their story,” Lin said.

Unlike software that automates a single task, Penny is designed to remain with the borrower throughout the origination process. Her memory of the borrower and the loan file allows her to use information gathered earlier in the process when completing later tasks, rather than treating each interaction as a separate transaction.

“We decided that having an AI employee would be a more effective solution if Penny is there for borrowers from starting an application all the way to submission to underwriting,” Lin said. “Because she’s able to use her memory and context of this borrower to make more informed decisions along the way.”

Copperlane pushes back against describing Penny as another point-of-sale system or software platform. The company instead presents her as a digital employee with the knowledge of a skilled processor and the capacity of “more or less infinite processors.”

Penny performs two distinct roles: an internal staff assistant for loan officers and other lender employees, and an external assistant that communicates directly with borrowers.

Inside a lender’s Microsoft Teams or Slack workspace, Penny can provide borrower updates, send reminders, follow up with relevant parties, and flag potential problems, including anomalies in income or credit documents. Loan officers can also ask her conversational questions about guidelines or give her documents to analyze.

Penny can prepare recommendations, including preapproval findings, but sends them to a loan officer for review before they reach the borrower.

In her consumer-facing role, Penny communicates with borrowers through text or email. She can collect documents, answer routine questions, assist with income calculations, and identify errors earlier in the process. Automating that work is intended to reduce the document chasing and repeated back-and-forth that consume processors’ time and delay loan files.

However, Penny is not a licensed loan officer. She cannot quote rates or deliver final recommendations directly to borrowers. If a borrower requests a rate quote, she is programmed to defer the question to a human loan officer.

Independent brokerages have emerged as a particularly strong use case, the founders said, because smaller shops tend to be more sensitive to staffing expenses.

“So she’s able to help brokerages process more volume without increasing their costs,” Lin said.

Copperlane claims Penny can reduce document chasing by as much as 75%, cut processor workloads by 80%, and enable loan officers to handle more than twice as many loans.

If those results hold at scale, the impact would reach beyond faster closings. By increasing the number of loans existing employees can manage, lenders could spread their personnel expenses across more originations — directly attacking the largest component of the industry’s nearly $12,000 cost to produce a mortgage.

Pylon: The Back-Office Mortgage Rails

Pylon is attacking the industry’s high cost to originate by rebuilding the infrastructure that connects mortgage originators with the capital markets. Rather than begin with the borrower-facing experience and add software to each step, the company worked backward from the secondary market to create one vertically integrated system.

“Over the last few years, we spent time vertically integrating the entire thing,” Hedge said, “starting with the capital markets and working backwards, rather than the originator or borrower and moving forwards.”

Traditional mortgage production depends on multiple software products handling separate parts of the process, with employees often responsible for moving information from one system to another. Pylon replaces that patchwork with standardized application programming interfaces, or APIs, that connect the underlying functions required to manufacture and fund a loan.

Those “mortgage rails” are not intended to operate like a conventional, out-of-the-box loan origination system. Instead, they give lenders the infrastructure to build their own applications around the borrowers they serve. More than 75% of Pylon’s customers use its APIs to develop custom platforms rather than rely on a standard interface, according to the company.

“Our view is that in the future, you’ll just originate a mortgage through an API call, the same way that payments, stocks, bonds, crypto, treasury products — really every other financial product — operate today,” Hedge said.

The connection to origination costs is straightforward: Every disconnected system, manual handoff, and back-office employee adds expense to the production of a loan. Pylon’s model is designed to consolidate those functions into software, reducing the technology and personnel costs that lenders would otherwise recover through higher margins.

That, Hedge said, allows Pylon to pass the underlying cost of the loan through to Main Street originators and borrowers without layering on the same operational expenses carried by more labor-intensive lenders.

“And so that’s been our whole philosophy: How do we sort of open up the true cost of capital to everybody in the market? And the only way you do that is by investing very deeply in software,” Hedge said. “Otherwise, we would have the same cost basis that everybody else would, and we’d have to, you know, add all of those margins in, too.”

Pylon’s broad scope also makes the company difficult to classify using the mortgage industry’s familiar categories. Hedge said industry participants frequently ask whether Pylon is a loan origination system or a point-of-sale platform. Pylon rejects those labels because its infrastructure performs functions that cut across several traditional software categories.

“There are these narrow frames that everybody has to think in,” Hedge said. “But if you come in and re-engineer the entire thing from the ground up or from first principles, you would ignore a lot of these acronyms that people have come up with.”

For now, however, that infrastructure is primarily available to large originators. Earlier this year, Pylon introduced a minimum commitment of 100 loans per month, but the company has made room for select startups, including fellow Y Combinator company Ralo.

“Ralo signed on early, before our current volume requirements were in place,” Hedge said in an emailed response. “Full API access is now typically reserved for large, publicly traded companies, but we make room for a select group of high-growth customers like Ralo. They’ve built an AI-native mortgage brokerage entirely on Pylon’s rails, and we’re happy with their early traction.”

Pylon charges a $10,000 monthly platform fee and an additional $1,250 pass-through fee per originated loan to cover expenses like credit reports, underwriting, processing, and closing.

“Long-term, we intend to serve any U.S. mortgage originator, and we’ll share more about our current customer base soon,” Hedge added.

Ralo, The AI Broker

If Pylon is building the back-office rails for a less labor-intensive mortgage process, Ralo is attempting to build the borrower experience on top of them.

Ralo is an AI-native mortgage brokerage founded by CEO Arjun Lalwani and CTO Helly Shah, two former technology professionals whose experience includes Google, YouTube, and Goldman Sachs. The idea grew out of the founders’ own frustrating homebuying experiences, which they felt involved excessive paperwork, unwanted solicitations, and repeated handoffs between loan officers.

“The problem wasn’t lack of software. It was the system: hidden fees, inflated rates, misaligned incentives. There was a loan officer, processor, and underwriter all increasing the costs of your rate. The overhead was never a feature. It was a tax. And borrowers paid it,” Lalwani said.

The founders became licensed loan originators to learn the mortgage business from the inside, then used that knowledge to train Ralo’s proprietary AI. The company’s stated mission is to “give every borrower the savings the mortgage industry keeps for itself.”

But Ralo remains in its infancy. The co-founders spent much of last year connecting borrowers with loan officers at no cost while learning the industry. The company began originating its own loans in March and raised $2.9 million in seed funding in June to expand the business.

Ralo uses proprietary AI to automate most of the work that follows an application, including tasks traditionally handled by processors and other back-office employees. Lalwani said digital prequalification has become nearly instantaneous, but Ralo wants to extend that one-click experience through full loan approval.

That requires automating the more difficult work of verifying employment, analyzing tax returns, and determining whether a loan meets investor requirements. It also means improving application intake so errors and missing information are identified before they create repeated follow-up, manual review, and processing delays.

“I think that an intelligent layer that constantly tells the borrower how they should correct their application so they can qualify for a loan quickly is very key,” Lalwani said. “We want to move towards that intelligent application intake, and we’re making a lot of progress there.”

Ralo currently uses self-reported borrower information to provide preliminary pricing before those details have been fully verified.

“So we are using all self-reported information from a consumer to actually give you what the rate you qualify for,” Lalwani said.

After the application is submitted, Ralo’s technology automates much of the file processing that would otherwise require back-office labor. The company combines that borrower-facing system with Pylon’s infrastructure, allowing the two startups’ technologies to automate opposite ends of the transaction.

“We have pretty much automated the entire role of a processor,” Lalwani said. “Pretty much … all the backend stuff has been automated. The thing that we’re working towards now is we are the ones advising a lot of, like, the borrowers and really understanding what kind of advice they need at what point of the journey.”

Because its AI handles much of the processing work, Ralo’s current origination operation is run by its two co-founders. Lalwani believes the model can substantially reduce the time between application and closing.

On the industry’s typical 45-day closing timeline, Lalwani said, “They should happen at least twice or three times faster, and we, to a large extent, have started to see that, and we feel very confident that post application intake everything will be automated in the next few years."

That reduction in manual labor is also central to Ralo’s cost-saving argument. Lalwani said early testing found that the company’s cost to originate was already at least four times lower than the industry average of $11,800. As Ralo scales, he expects those costs to decline further, allowing the brokerage to pass more of the savings on to borrowers.

The same economics could substantially change the role — and compensation — of a loan officer.

Ralo expects its loan officers to move away from document collection and file processing and toward advice, relationship management, and technology development. With AI performing more of the routine work, the company envisions a single loan officer eventually overseeing hundreds or even thousands of loans.

Those loan officers would also help train the technology. By teaching the AI how to handle unusual borrower circumstances and other difficult edge cases, they would enable the system to perform more of the work independently over time.

Ralo does not expect the human role to disappear entirely. Loan officers would remain responsible for building trust with real estate agents, helping borrowers make complicated financial decisions and providing reassurance when problems arise, such as an unexpectedly low appraisal. In that model, the loan officer becomes less of a file manufacturer and more of a strategic adviser and human point of contact.

Ralo’s founders believe that shift also makes the industry’s traditional commission structure increasingly difficult to justify. 

“We do think the loan officer making 100, 200, 300 basis points per loan is not sustainable in the long run, because a lot of the work that loan officers do today will be done by AI,” Lalwani said. “What I would rather have is to pay them a great base salary and make sure that their performance is directly correlated to how satisfied the customer was.”

If the model works at scale, Ralo would not eliminate the loan officer so much as separate the value of human advice from the cost of manufacturing a mortgage. Technology would handle the repetitive work, while loan officers would be paid for the judgment, trust, and guidance that automation cannot yet provide.

Bevri.ai’s Data-Driven Mortgage Platform

Jonathan Haddad’s vision for bevri.ai grew out of two very different experiences in mortgage lending. In 2016, while he worked for Rocket Mortgage, he witnessed the company’s massive “Push Button, Get Mortgage” launch — which happened to fall on Super Sunday. “The system crashed — it was that busy,” he recalled.

Haddad said the technology automated so much of the application process that the loan officers on his team had little left to do during the intake phase. They came into the office and “pushed a button” to submit locked loans to underwriting because their licenses were legally required to move the files forward. With the system handling much of the paperwork, individual loan officers could manage volumes rarely seen in the broker channel.

“The clients were finishing their application online, my LOs were walking in because they had the license, and then I was paying these loan officers simply to submit a loan to underwriting,” Haddad said. “I watched this happen 10 years ago, and now you fast forward to today, these opportunities are now here for the broker channel, and you’re able to afford it.”

Haddad built bevri.ai to give independent mortgage companies access to the automation and performance data that have traditionally required the resources of a major retail lender. The all-in-one platform is designed to help brokers complete more applications, identify problems earlier, and understand where they are losing borrowers throughout the sales process.

At the same time, Haddad does not want bevri to become another closed system. Its open application programming interface, or API, allows loan officers to connect their existing customer relationship management platforms, point-of-sale systems, and vendors directly to the platform.

“The philosophy for me moving forward is the same philosophy that I believe companies like ChatGPT are doing, which is this open concept,” he said.

Bevri begins by attempting to convert more prospective borrowers into completed applications. According to Haddad, company data show that consumers who begin their applications within the platform complete them 92% of the time.

“Right now, the data that we have indicates that if your consumer starts inside a bevri, there’s a 92% chance they complete the application in full,” Haddad said.

Its proprietary AI assistant, BevriLLM, then communicates directly with borrowers and guides them through the application process while the loan officer is unavailable. The system can identify problems in uploaded documents, such as an expired pay stub, and immediately ask the borrower to correct them. Haddad said the platform can catch those errors “within minutes of the client uploading it.”

Identifying problems at intake is intended to reduce the repeated follow-up and manual review that slow down a loan file. But bevri’s broader value proposition extends beyond automating individual tasks. The platform consolidates data that are typically scattered across numerous vendor portals, giving brokers a clearer view of how borrowers move from lead to closed loan.

Haddad said brokers cannot improve their performance without tracking the key performance indicators behind their operations. That includes how many leads result in a credit pull, how many credit pulls become closed loans, and where the company is losing time, money, or prospective borrowers.

“We don’t have that on the broker space. Some of your largest shops do, but that’s because they have the money to do it,” Haddad said. “Our community has never had the ability to be specific. They don’t know what they need to work on. And one of the biggest reasons is you have data in 47 different portals.”

That emphasis on transparency also applies to third-party companies using the platform. Haddad said lead providers and other vendors operating through bevri must disclose their conversion metrics, allowing loan officers to compare which sources generate activity and which ultimately produce funded loans.

“That is what’s going to make us different,” Haddad said.

Bevri uses that data not only to show loan officers how they are performing, but also to tell them what to do next. The platform analyzes each loan officer’s database and produces a daily scorecard comparing that performance with peers. Its agentic AI then assigns specific actions intended to improve conversion rates.

Newer loan officers can also study the scripts and follow-up strategies used by higher-performing members of the community. The goal is to turn the habits of top producers into measurable practices that other loan officers can adopt, rather than leaving each originator to determine independently what works.

“I’m not going to be nice about your conversion numbers,” he said. “We are going to be real, we’re going to be direct, and we’re going to tell you, here’s where you’re dropping the ball, here’s what other LOs in your peer group are doing that you are not doing, and this is what you need to do to execute on.”

The Eventual Transition

The founders describe the transition less as the disappearance of mortgage professionals than a redistribution of their work. AI would handle more document collection, file review, data entry, and routine follow-up, while the people surrounding the loan move closer to the borrower.

Copperlane’s founders believe loan officers should retain ownership of the customer relationship even as Penny assumes more of the work behind it.

“I’d say one of the first things we learned when we entered the industry is it’s built on relationships,” Brianna Lin said. “And that’s one thing that AI will never be able to take over. One thing we always say is the human — the loan officer should always be the one who owns the relationship.”

That vision would also change the processor’s role. Instead of spending the day collecting documents and moving files forward, processors could manage larger pipelines and concentrate on exceptions requiring experience or judgment.

Ralo CEO Arjun Lalwani offers a more disruptive version of that transition. In Ralo’s model, loan officers would spend less time reading files and chasing paperwork and more time advising borrowers at a scale that would be difficult under the industry’s current structure.

“One way is to frame it as, ‘Hey, your job is being automated,’” Lalwani said. “The other way to frame it is like you actually get to do the thing that is the most exciting part, which is talking to humans and helping humans out, and you can do it at a scale that no other loan officer could do it before.”

Pylon founder and CEO Trent Hedge expects some mortgage professionals to move from production roles into customer-facing positions as software replaces more of the work involved in manufacturing a loan.

“What we’re starting to see is the people are shifting from the back office to the front office,” Hedge said. “So if you eliminate a lot of the manual process of manufacturing a mortgage, and you replace the factory assembly line with these standardized rails, as we’re doing, then you actually have more volume to deal with. Because, again, now you’ve made it cheaper to get a mortgage, so more people can get a mortgage. So now you have to take those people from the back office and put them in the front office so they can help customers. And we actually think that’s a good shift.”

For bevri.ai founder Jonathon Haddad, technology would not eliminate processors but expand the volume each one could handle.

“So when your processor logs in, there’s always going to be space for a processor, in my opinion, but hey, if your processor’s doing 30 units a month, we want them to be able to do 60 units a month,” Haddad said. “This is all about doubling capacity.”

For loan officers and processors, the question is no longer whether AI will change the mortgage process, but which parts of their work they can hand off to technology — and how they will use the time it gives back.