41% Lower Origination Costs, 2x Conversion: How Better.com's AI Loan Agent Betsy Automated 1.89M Mortgage Calls
Better.com's Challenge
Better is an AI-native mortgage lender that has funded over $110 billion in loan volume. Mortgage origination depends on state-licensed loan officers whose time is the primary constraint on throughput: a borrower calling to check eligibility, get a rate, or lock a loan requires a licensed consultant regardless of how routine the enquiry. Better's ambition to operate as a high-volume, low-cost originator was structurally limited by this headcount dependency. A loan officer spending time on eligibility checks and rate inquiries — questions with deterministic answers given a borrower's data — is a loan officer not working on the complex cases that genuinely require human judgement. Automating voice interactions in a regulated financial services environment introduced a further constraint: any AI agent handling mortgage conversations must operate under explicit controls. Pricing and eligibility decisions cannot be delegated to a general-purpose language model — they must be computed by authorised systems and communicated by the agent, with no opportunity for unsanctioned decisions. Better needed a voice layer capable of sustaining natural, trusted borrower conversations at scale while remaining auditable and compliant within a state-regulated lending framework.
How ElevenLabs Solved It
Better built Betsy — a generative AI voice loan assistant — by layering ElevenLabs Agents on top of Tinman, Better's proprietary full-stack mortgage platform. The architecture separated concerns explicitly: Tinman executes all pricing, eligibility, and rate-lock computations, returning results to ElevenLabs, which communicates them to the borrower. No unsanctioned decisions pass through the voice layer. ElevenLabs replaced an earlier Speech-to-Speech model after Better determined that a Speech-to-Text → LLM → Text-to-Speech pipeline produced more natural interaction and lower latency. Betsy uses a professionally licensed, consented cloned voice for brand consistency. The system handles inbound and outbound calls — guiding borrowers through eligibility checks, pricing, and rate locks entirely by voice — and can execute dozens of Tinman tool calls within a single conversation without interrupting conversational flow. Deployment began in June 2025; volume scaled to approximately 100,000 monthly conversations within the year.
The Outcome for Better.com
In 2025, Betsy placed 1.89 million inbound and outbound calls. 35.5% of mortgage inquiries were resolved end-to-end by the agent without any human loan officer involvement. The cost to originate a loan fell 41%. Lead-to-lock conversion doubled. Loan officers recovered more than 1,666 hours of aggregate monthly capacity — time previously consumed by routine eligibility and pricing calls that Betsy now handles independently. Cory Hayden, AI Product Manager at Better, summarised the operational significance: "ElevenLabs gives us the control and reliability we need to handle over 100,000 monthly mortgage conversations." The result demonstrates that regulated AI voice agents — where the voice layer communicates rather than decides — can achieve material automation rates in financial services without introducing the liability risk that unsupervised AI decision-making would carry.
What Better.com Learned
Better's architecture solves the central problem in deploying AI agents in regulated industries: the difference between an agent that communicates and an agent that decides. Betsy does not determine whether a borrower qualifies for a loan or at what rate — Tinman computes those outputs and Betsy reports them. This separation of computation from communication is what makes 35.5% automation achievable in a state-regulated lending environment where an AI system that made unsanctioned pricing decisions would create immediate compliance exposure. The 1,666 hours of monthly officer time recovered is not a headcount reduction story — it is a reallocation story. Loan officers are licensed professionals whose regulatory value lies in exercising judgement on complex applications; consuming that capacity on routine eligibility calls is an allocation decision, not a necessity. The 41% cost-to-originate reduction flows directly from correcting that allocation: when routine enquiries are handled by voice AI and complex cases reach licensed officers faster, origination economics improve on both the cost and conversion dimensions simultaneously. For any financial services firm deploying AI in customer-facing roles, the architecture pattern here — voice layer communicates, core platform decides, audit trail maintained throughout — is the design that makes compliance and automation compatible.
- In regulated industries, the compliance-compatible AI agent architecture separates the voice layer (communicates outcomes) from the decisioning layer (computes them) — this is what makes high automation rates achievable without creating liability exposure.
- 1,666 hours of monthly loan officer time recovered is not headcount saving — it is capacity reallocation. Licensed professionals deployed on routine eligibility calls cannot simultaneously be working on complex applications where their judgement creates value. Automation of the routine is what unlocks throughput at the top of the funnel.
- The 2x lead-to-lock conversion improvement alongside the 41% cost reduction shows these metrics are complementary, not in tension: faster response to borrower enquiries improves conversion at the same time as automation reduces per-loan cost. The bottleneck was officer availability, and removing it improved both.
See how ElevenLabs can help your business address similar challenges.
Affiliate link — we may earn a commission at no cost to you.