Key Moments
What Happens When You Build Software for Agents
a16zKey Moments
Building AI agents into the mortgage industry requires replacing legacy systems entirely, as outdated data architectures prevent true automation and lead to massive inefficiencies.
Key Insights
Vesta replaces the entire core system of record for mortgage lenders, rather than building AI agents on top of existing legacy software.
The US mortgage system, with its 30-year fixed-rate mortgages, is inherently complex due to duration risk, credit risk, and macro sensitivity, making underwriting difficult.
Outdated data architectures in legacy mortgage systems, characterized by flat file databases and 30,000+ fields, prevent parallel processing and hinder automation.
Vesta's AI agents are integrated into an orchestrated workflow, triggered by classical rules rather than relying on individual user prompts, increasing efficiency.
PennyMac, a top five US lender, reported a 25% reduction in back-office operating costs before AI agents were fully deployed, indicating significant gains from Vesta's platform.
The regulatory landscape for mortgages is highly complex, with a patchwork of rules from federal and state entities, as well as government-sponsored enterprises like Fannie Mae and Freddie Mac.
The inherent complexity of the US mortgage system
The US mortgage market is unique globally, primarily due to its widespread offering of 30-year fixed-rate mortgages. This structure, while beneficial for the American middle class, introduces significant underwriting challenges. Lenders must account for 30 years of credit risk and duration risk, meaning they are essentially betting on a borrower's financial stability over three decades. This long-term exposure makes mortgages highly sensitive to macro-economic shifts, as demonstrated by events like the Silicon Valley Bank collapse, which was partly attributed to interest rate risk on long-held assets. The government's role in backing these mortgages, through entities like Fannie Mae and Freddie Mac, further complicates the landscape, leading to a heavily regulated and often bureaucratic environment. This inherent complexity, combined with political sensitivities and historical issues like redlining, creates a system that is both deeply important and incredibly difficult to navigate.
Why layering AI on legacy systems falls short
Mike Yu, co-founder and CEO of Vesta, argues that simply adding AI agents on top of existing legacy mortgage software is insufficient for true automation. He explains that these outdated systems, some dating back to the 1980s and 1990s, possess fundamental flaws, particularly in their data architecture. These systems often use flat file databases with tens of thousands of redundant fields, lacking relational structures. This architecture prevents parallel processing, meaning only one agent or process can access and modify loan data at a time. For instance, automating tasks like calculating income and assets or reviewing purchase contracts simultaneously is impossible. This limitation creates a significant bottleneck, particularly for AI agents that could otherwise operate more efficiently in parallel. Yu emphasizes that to achieve meaningful automation, the core system of record, including its data architecture, must be completely rebuilt.
Vesta's approach: A new system of record and AI orchestration
Vesta's strategy involves replacing the lender's existing core system of record with a modern, cloud-native platform designed from the ground up. This new system provides the underlying compliance, rules, workflow, and integrations necessary for modern mortgage origination. The critical addition is the integration of AI agents, but not as standalone co-pilots. Instead, Vesta orchestrates these agents through classical automation rules. When a specific event occurs in the loan process, such as a borrower signing disclosures, automated rules trigger relevant AI agents. For example, an agent might be tasked with verifying purchase contract terms against disclosures, checking dates, and ensuring the property is within a certain radius of the borrower's employer. This event-driven, orchestrated approach allows for more controlled and reliable automation, which is crucial for building trust with lenders and regulators. This contrasts with older models where users would manually prompt agents, which Vesta found less suitable for enterprise adoption in a regulated industry.
The formidable challenge of replacing core systems
Ripping out and replacing a core system of record in an industry as critical and regulated as mortgages is an immense undertaking. Yu acknowledges that many advised against this approach, suggesting alternative 'wedges' or building AI on top of existing infrastructure. He highlights that incumbents' data architectures are often fundamentally broken, making integration and automation exceedingly difficult. For example, flat file databases with thousands of fields and single-writer limitations prevent parallel processing essential for AI swarms. The complexity extends to integrating with various third-party systems, including Automated Underwriting Systems (AUS) from Fannie Mae and Freddie Mac, which themselves have rigid policies and integration requirements. Furthermore, interfacing with thousands of county-level recording offices, each with unique document formatting rules, adds another layer of difficulty. Vesta's journey, including originating their first loan in December 2022 after nearly two years of development, underscores the significant technical and partnership challenges involved.
Navigating the regulatory and political landscape
Mortgages are deeply intertwined with government policy, primarily due to the US government's commitment to 30-year fixed-rate mortgages. This political backing translates into a dense web of regulations from entities like the Consumer Financial Protection Bureau (CFPB) and state-level regulators. Government-Sponsored Enterprises (GSEs) like Fannie Mae and Freddie Mac issue extensive, often unyielding, underwriting guidelines. While many of these rules are based on government-developed Automated Underwriting Systems (AUS), which Vesta integrates with, the overall environment is risk-averse. Yu notes that while automating back-office tasks is generally less contentious, the role of AI in customer-facing positions, like sales agents, remains a gray area with unclear regulatory guidance. A specific challenge Vesta faces is the VA's portal for Certificate of Eligibility, which currently requires a user account, hindering automation for VA loans until the VA provides a system account solution.
The impact of AI and the future of mortgage origination
Vesta's initial strategy, even before the rise of advanced AI agents, was to automate 80% of mortgage origination work using improved computer vision for document extraction, classical rules, and workflow configuration. The advent of powerful LLMs and AI agents significantly accelerated this vision, moving towards 95% and eventual 100% automation. PennyMac, a top five lender, reported a 25% reduction in back-office operating costs prior to the full deployment of Vesta's AI agents, illustrating the platform's efficiency gains. Yu estimates Vesta is currently around 75% of the way towards full automation, with the last 25% being the most challenging due to specific regulatory hurdles and the diffusion of technology adoption among lenders. He anticipates that while the core technology for automation is largely in place, widespread adoption will take time. The long-term vision is to drastically reduce the time to close a mortgage, aiming for a future where the process is significantly streamlined and less costly.
Market dynamics and the SaaS apocalypse
Despite the broader market downturn for SaaS companies, Yu remains bullish on Vesta's position, particularly as a vertical SaaS provider in a highly regulated industry. He argues that the 'SaaS apocalypse' is a false dichotomy, as AI agent companies are fundamentally also enterprise software companies. The key differentiator for vertical SaaS, like Vesta's focus on mortgages, is that customers tend to use a larger portion of the product compared to horizontal SaaS, where customers might only utilize 10%. This deeper integration and value delivery make vertical solutions more resilient. Yu also notes that during periods of high interest rates and lower origination volumes, lenders become more receptive to implementing new technologies to reduce costs, creating a window for Vesta to gain market share. He believes that even with conservative forecasts, the baseline volume of purchase mortgages in the US remains substantial, ensuring a fundamentally sound business.
Motivation and the importance of mortgages
Yu is driven by the belief that the mortgage industry, despite its current disarray, plays a crucial role in supporting the American middle class and overall economic stability. He sees Vesta as uniquely positioned to bring AI and automation to this vital sector, a task he feels few others are equipped or motivated to undertake. He likens Vesta's situation to his earlier days at Blend, where the opportunity to make a significant impact in a complex, underserved market was compelling. For Yu and his team, the challenge of transforming a foundational industry and making homeownership more accessible is a powerful motivator. He emphasizes that while companies like Anthropic may be at the forefront of AI research, Vesta's mission is about delivering practical, impactful AI solutions to specific industries, creating value where a strong need and a clear opportunity exist. This focus on impact within a critical sector, rather than purely theoretical AI advancement, fuels his dedication.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
Common Questions
Vesta provides the core system of record and back-office capabilities for mortgage lenders. It aims to automate the entire mortgage origination process using AI agents, replacing manual tasks and integrating with existing systems.
Topics
Mentioned in this video
A company providing core system of record and back-office capabilities for mortgage lenders, incorporating AI agents to automate processes.
Government-sponsored enterprise involved in the US housing market. Vesta has designed data verification programs for them.
A company whose CTO's insight about data architecture being key to replacing core systems was influential.
Government-sponsored enterprise involved in the US housing market. Vesta has designed data verification programs for them.
Mikeu's former employer where he interned as a trader. Its probabilistic thinking heavily influenced his approach.
A company whose event-driven monitoring approach inspired Vesta's orchestration of AI agents.
A digital mortgage company where Mikeu interned and later worked for four years, gaining significant experience in the mortgage space.
A major bank that was an early customer of Blend. Mikeu's work on their mortgage application's accessibility compliance was a key early project.
One of the major banks that Blend worked with during its early days.
An AI research company mentioned as an example of a company where smart people could work, contrasting with Vesta's mission.
Company that Blend's founders reportedly left to start their mortgage company.
A top-five US mortgage lender and a major Vesta customer that reported significant cost savings after implementing Vesta's solution.
A company whose agent product Vesta used early on, which Mikeu found instrumental in understanding agent capabilities.
Silicon Valley Bank, mentioned as an example of a bank holding mortgages that faced problems due to duration risk.
A horizontal SaaS company that people complain about but rarely migrate from, used as a point of comparison for vertical SaaS.
Company that ran a Super Bowl ad promoting online mortgage applications, influencing the market's adoption of digital mortgages.
Department of Justice. Had a judgment against Wells Fargo regarding ADA WCAG compliance.
Venture capital firm and host of the podcast, with Mikeu's company Vesta being in their portfolio.
Consumer Financial Protection Bureau, one of the regulators overseeing mortgage lenders.
Office of the Comptroller of the Currency, a regulator for banks that also oversee mortgage lenders.
Department of Veterans Affairs. Its portal for certificate of eligibility for VA loans presents a challenge for full automation due to system account limitations.
The VA, responsible for VA loans and whose portal's limitations hinder full automation of VA loan eligibility checks.
The US Department of the Treasury, which received funds from Vesta for integration resources.
Media outlet mentioned for its mortgage price tracker, indicating the commodity nature of mortgage pricing.
An automated underwriting system from Fannie Mae that Vesta needed to integrate with.
Mentioned as a programming language that might be used for 'vibe coding' an LOS, contrasted with Vesta's approach.
Amazon Web Services, used as an analogy for how cloud infrastructure costs are viewed, contrasting with token pricing models.
Mentioned alongside DataDog in the context of event-driven monitoring that inspired Vesta's agent orchestration.
A service that provides employment and income verification. Vesta has integrated with it.
An AI model mentioned as being not smart enough to build agents initially, but later enabling them.
A company in the life sciences vertical, used as an example of vertical SaaS.
Mentioned in the context of building an LOS without proper documentation or established processes.
An AI model that was a significant unlock for enabling meaningful work with agents.
A model mentioned for its pricing structure (10 cents per million input tokens), suggesting competition in the model layer.
A country whose mortgage market, characterized by 5-year terms, is contrasted with the US 30-year fixed-rate mortgages.
State mentioned for its custom fields in county recording offices, posing a data modeling challenge.
A country mentioned as potentially having 30-year fixed-rate mortgages, unlike most other nations.
City where Mikeu worked full-time at Blend during his senior year of college.
Large Language Models, mentioned in the context of not being initially perceived as world-dominating in 2020, but developments in computer vision were clear.
Technology with significant developments observed around 2020, relevant to mortgage origination tasks like document extraction.
Artificial intelligence agents that Vesta uses to replace manual back-office work in mortgage origination.
More from a16z Deep Dives
View all 65 summaries
40 minHow Valon Is Rebuilding Mortgage Servicing for the AI Era
25 minHow Microsoft is Adapting to the AI Era
22 minCybersecurity in the Agentic Era | Deep Dives with a16z
23 minDatadog CISO on Securing AI Agents at Scale | Deep Dives with a16z
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free