Vesta成功完成了3000万美元的B轮融资，旨在进一步推广用于处理抵押贷款业务流程的AI智能体，包括信贷审批环节。参与本轮投资的包括三家已在使用该软件的贷款机构。不过，同样的软件若在欧洲市场推广，则可能因违反《人工智能法案》而面临高风险；而《通用数据保护条例》也早已对此类自动化决策作出了限制。
Vesta has raised a $30M Series B to expand AI agents that handle mortgage origination work, including underwriting decisions, with three customer lenders among the investors. The same software sold in Europe would be high-risk under the AI Act, and GDPR already limits automated decisions of this kind. Vesta has raised a $30M Series B to put AI agents inside mortgage lenders.
此次融资由Conversion Capital牵头，除三家现有客户贷款机构外，花旗风险投资以及安德森·霍洛维茨基金也参与了投资。Vesta首席执行官迈克·余向TechCrunk透露，过去一年市场需求“呈爆发式增长”，公司营收增长了12倍。截至目前，Vesta累计融资额已达8500万美元，不过其在市场中的份额仍不足5%。
Conversion Capital led the round, and three lenders already using the software invested alongside Citi Ventures and Andreessen Horowitz. Chief executive Mike Yu said demand had “exploded in the last year” and revenue is up twelvefold, he told TechCrunch. Vesta has raised $85M in total and puts its market share below 5%.
据HousingWire报道，使用该平台的贷款机构每年发放的抵押贷款总额超过1000亿美元。目前，Vesta平台上约40%的业务流程已由AI智能体及自动化工作流负责处理。这些自动化流程的应用范围远不止文书工作：贷款机构利用这些智能体进行文档审核与数据验证，部分机构甚至将其用于信贷审批、条件确认以及贷款结案材料的审核。
Lenders on its platform originate more than $100B of mortgages a year, HousingWire reported. About 40% of tasks running through Vesta are now handled by agents and automated workflows. Those tasks reach further than paperwork.
成本优势是此类技术备受青睐的主要原因。在美国，完成一笔抵押贷款流程平均需要40天时间，相关成本约为1.1万美元，其中大部分为人力支出。多年来，美国抵押贷款银行协会的数据也显示，抵押贷款办理成本始终高于1万美元。
Lenders use the agents for document review and data validation, and some for underwriting decisions, condition clearing and closing package review. The appeal is cost. A US mortgage takes around 40 days to close and roughly $11,000 to produce, mostly labour.
迈克·余指出，Anthropic公司开发的Claude Sonnet 4.5模型起到了关键推动作用。他表示，该模型能够长时间严格遵循预设指令，完美契合抵押贷款业务流程的复杂需求。
The Mortgage Bankers Association has put origination costs above $10,000 for years. Yu credits one model.
不过，自动化信贷决策也伴随争议。2021年《The Markup》的一项研究显示：在控制17项财务因素后，美国贷款机构拒绝黑人申请者的概率仍比拒绝白人申请者高出80%。迈克·余强调，贷款机构仍需对所使用的任何软件生成的信贷决策承担全部责任，所有操作及其背后的决策逻辑均会被详细记录以供审计。
He named Anthropic’s Claude Sonnet 4.5 as the breakthrough, saying it followed configured instructions over the long stretches that mortgage work demands. Automated lending decisions carry a record. The Markup found in 2021 that US lenders were 80% more likely to deny Black applicants than comparable white ones, after controlling for 17 financial factors.
欧洲并不会将此事完全交由供应商自行处理。《人工智能法案》将用于评估自然人信用状况的系统归类为高风险系统，而房贷审批显然属于此类系统。
Yu says responsibility does not move. Lenders remain accountable for underwriting whatever software they use, he said, and every action and the reasoning behind it is logged for audit.
这一分类意味着相关方需承担法定义务而非仅遵循指导性建议。供应商必须完成合规性评估与注册程序；而系统使用者则需安排具备相应资质的人员实施人工监督，留存至少六个月的操作日志，并开展基本权利影响评估。
Europe does not leave that to the vendor. The AI Act puts systems that evaluate the creditworthiness of natural persons in its high-risk class, and mortgage approval sits squarely inside it.
这些义务的履行期限此前曾被推迟。根据欧盟第2026/1744号法规，原定于今年8月生效的要求被延至2027年12月2日。届时将由新设立的“人工智能办公室”负责监管，该机构有权对违规企业处以相当于其年营业额3%的罚款。
That brings obligations rather than guidance. Providers face conformity assessment and registration, while deployers must assign human oversight to competent staff, keep six months of logs and run a fundamental rights impact assessment. Those duties were pushed back.
不过有一条规则早已生效。《通用数据保护条例》第22条明确规定：任何仅由自动化系统作出的、可能产生法律或其他重大影响的决定均属违规，而拒绝发放房贷正是典型的此类案例。
Regulation (EU) 2026/1744 deferred them from August this year to 2 December 2027, enforced by an AI Office that can fine 3% of turnover. One rule is not waiting.
实际上欧洲已有类似实践。今年6月，柏林的Taktile公司成功筹集1.1亿美元用于开发商业贷款审批系统。由于该条款仅适用于自然人，因此上述系统并不受其约束。
GDPR Article 22 already restricts decisions made solely by automated means where they carry legal or similarly significant effects, and a refused mortgage is the textbook case. Europe already has a version of this. Taktile in Berlin raised $110M in June for agents that underwrite business loans, which point 5(b) does not reach because it covers natural persons.