Pennymac and New American Funding buy into Vesta, the AI agent startup they already use for mortgages

"Many of our customers start an AI agent with a person approving its work, then let it handle a share of loans on its own, then expand," Yu said ^[8]. A loan operations manager switching Vesta on next week starts at the first of those steps. The product as pitched is bigger: people deploying a swarm of agents to finish tasks faster, with each customer choosing which tasks to assign ^[6].

By Yu's own account, users put a person behind the agent and widen its remit as it earns approval. They do not buy the swarm and switch it all on. A US mortgage takes around 40 days to close and costs around $11,000 per loan, according to TechCrunch ^[14]. "Most of that cost is human labor, and a major bottleneck in the timeline is just waiting for a person to get to reviewing your loan," Yu said ^[15]. An agent whose every output sits in a person's approval queue leaves that bottleneck where it was. Days come off the close at the second step, when the agent handles loans alone ^[8].

Neither of Yu's growth figures separates those steps. Revenue and loan volume both rise whether a person approves every agent action or none of them ^[7][9]. Vesta did not disclose revenue in dollars, or how many loans its agents handled without a person signing off. If Yu's share figure is measured in annual loan volume, the market he is counting is above $2 trillion a year ^[20].

Three of the lenders paying for Vesta now own part of it. Three customers put money into the round alongside Citi Ventures and Andreessen Horowitz ^[2]. The $30 million round is about 35% of everything Vesta has raised ^[19].

For the manager deciding which tasks move to the second step, I'd sort them on two questions. The first is whether an agent's error would land as an underwriting decision the lender has to defend, since Yu said the lender stays responsible whatever software it runs ^[11]. The second is whether a reviewer can check the agent's output faster than a person could complete the same task by hand. Tasks low on liability and quick to check should move first. Each one takes a person out of the queue Yu blames for the wait ^[15]. Underwriting calls that are slow to check should stay supervised longest. The tradeoff sits in that same quadrant: a person on every high-liability file protects the lender's audit position and gives up most of the labour saving that justified buying the software. The measures that show whether the rollout worked are the share of files closed with no human approval and days to close against the 40-day figure ^[14].