Sun West Deploys AngelAi to Eliminate Underwriter Bias in Mortgages

How Artificial Intelligence Can Help Eliminate Discrimination in Financial Services — Ai4

Ai4

Mortgage discrimination stems from human underwriters rejecting complex minority files to avoid administrative fatigue, a structural failure automated algorithmic underwriting eliminates by processing unstructured documentation deterministically.

AI-assisted summary of How Artificial Intelligence Can Help Eliminate Discrimination in Financial Services — original source. Check the original for full context.

Source published: 2026-08-22

Detailed notes

## Background Published on YouTube by Ai4 on August 22, 2026, this fireside chat features Sun West Mortgage CEO Pavan Agarwal and Forbes contributor Tor Constantino exploring how artificial intelligence can help eliminate discrimination in financial services. The session discusses practical frameworks for responsible AI deployment, auditable decision-making, explainability, and the safeguards needed to expand access without replicating historical lending disparities.

## TL;DR Pavan Agarwal explains how AngelAi, developed and deployed within Sun West Mortgage since 2018, leverages automated decisioning to remove human bias and dramatically reduce error-driven loan denials among minority and veteran borrowers. He illustrates that traditional underwriting disparities stem largely from processor fatigue when handling complex, non-traditional documentation rather than conscious malice, making tireless AI compute cycles naturally fairer. To comply with regulatory standards and safeguard consumer privacy, the platform uses deterministic, real-time traceable audit logs and operates entirely within private data centers rather than third-party APIs. Ultimately, the discussion highlights that while unadjusted models risk reflecting historical industry prejudices, neutral algorithmic architectures that strip demographic markers can effectively dismantle systemic redlining.

## Atmosphere & style The dialogue carries a pragmatic, insider-driven tone that balances high-level algorithmic theory with the grounded reality of commercial banking operations. Constantino and Agarwal maintain an engaging and collaborative fireside dynamic, weaving personal anecdotes about immigrant family struggles and early coding with rigorous technical debates on model collapse and synthetic data. Watching the session feels like an authoritative yet candid briefing on financial reform, shifting from sober critiques of banking gatekeeping to a passionate, data-backed optimism for algorithmic equity.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=4s) AngelAi has operated in commercial production since 2018, processing over $40 billion in transactions and serving more than 200,000 families through automated underwriting. - [01:06](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=66s) The platform evolved over decades, progressing from 1980s rule-based algorithms and 1990s database vectorization to the first automated reverse mortgage engine and conversational AI. - [02:44](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=164s) Analysis of multi-year banking data revealed stark decline error rates in manual underwriting: 1 in 2 Black applicants, 3 in 8 Hispanic applicants, and 2 in 3 Veterans denied should have been approved. - [02:44](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=164s) Underwriting bias often stems from cognitive fatigue rather than conscious malice, as human processors reject complex, non-traditional files to reduce their manual workload. - [05:48](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=348s) Interacting with an AI removes the fear of human condescension and judgment, contributing to consumer trust in AI lending platforms exceeding 70%, compared to 19% for traditional bankers. - [08:05](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=485s) Developing the technology in-house with $500 million in internal bank capital avoided the short time horizons of venture capital, allowing multi-decade iteration in live production environments. - [09:43](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=583s) Preventing model collapse and drift requires strictly banning synthetic training data and relying on dedicated data engineering teams to curate empirical, real-world inputs. - [12:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=724s) Regulatory compliance with the CFPB and FDIC requires deterministic, real-time audit logs for every decision step, rather than backward-looking post-hoc reprompting. - [13:56](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=836s) Automating the evaluation of unstructured alternative credit data (such as VantageScore) eliminates the manual bottlenecks that typically lead human underwriters to reject non-traditional applicants. - [13:56](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=836s) Running AI workloads entirely within dedicated, private data centers prevents proprietary consumer financial records from being harvested by external third-party foundation models. - [17:39](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=1059s) Modern redlining frequently occurs because traditional underwriters lack the specialized CPA-level knowledge needed to parse complex income for gig workers and small business owners. - [17:39](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=1059s) A Loyola University study showed that large language models replicate legacy banking disparities until racial identifiers are removed from the prompt inputs, resulting in equitable decisions.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=4s) Introductions and the Origin of AngelAi

Tor Constantino opens the fireside chat by introducing himself as an artificial intelligence columnist for *Forbes* and a consultant across AI, quantum computing, and blockchain. He introduces his "partner in crime," Pavan Agarwal, highlighting his long-standing expertise in financial services. Agarwal notes that his background spans 45 years in the financial services sector, culminating in the creation of AngelAi. He points out with a laugh that AngelAi went into full commercial production back in 2018—well before the public explosion of ChatGPT. Since its rollout, the platform has processed more than $40 billion in transactions and assisted over 200,000 families, leveraging automated artificial intelligence decision-making and direct banking execution to systematically eliminate discrimination in lending.

- **Production Timeline:** AngelAi has been operational in full production since 2018. - **Scale and Impact:** Over $40 billion transacted and more than 200,000 families served through automated AI decisioning.

---

[01:06](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=66s) From 11-Year-Old Programmer to End-to-End Autonomous Underwriting

Agarwal explains that he is a second-generation financial services technologist who began programming at age 11, solving esoteric mathematical problems on early personal computers within his family's business. His early work focused on building algorithms designed to automate mortgage and lending decision processes. While he began writing early AI algorithms in the 1980s, the hardware and computing power of that era were insufficient for live production environments.

A major technical breakthrough occurred during the 1990s when his team successfully vectorized both the underlying database structures and the decision-making pipeline. By the 2000s, this work evolved into the industry's first automated AI underwriting engine for reverse mortgages—a milestone Agarwal jokes about, noting that he is now only a few years away from qualifying for a reverse mortgage himself. By 2018, these decades of research coalesced into a unified, end-to-end conversational AI interface. Borrowers can engage in an empathetic chat where they outline their financial situation, goals, and personal dreams, while the AI dynamically engineers and structures a bespoke financial solution.

```mermaid flowchart TD accTitle: Evolution of AngelAi subgraph S1["1980s: Foundational Logic"] A["11-Year-Old Coder solving esoteric math"] --> B["Early lending automation algorithms limited by compute power"] end subgraph S2["1990s-2000s: Vectorization & Underwriting"] C["Vectorization of databases & decision pipelines"] --> D["First automated AI underwriting for reverse mortgages"] end subgraph S3["2018-Present: Full-Stack Production"] E["Empathetic conversational interface"] --> F["End-to-end automated loan engineering & execution"] end B --> C D --> E ```

---

[02:44](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=164s) Historical Lending Disparities and the Root Cause of Underwriter Bias

Addressing Constantino's concern that AI models might ingest decades of systemic redlining and amplify historical inequities at scale, Agarwal stresses that "AI is like any other tool, it's only as good as the people behind it." As first-generation immigrants, Agarwal and his family witnessed firsthand how traditional banking practices routinely mistreated and excluded minority communities. To quantify the scale of the issue, his team analyzed seven to eight years of government mortgage data from the major commercial banks, exposing stark baseline rejection errors:

Agarwal demystifies how bias enters manual underwriting, explaining that it rarely stems from blatant, conscious discrimination where an underwriter maliciously rejects a file based on race. Instead, borrowers from underserved backgrounds frequently present complex, non-traditional financial documentation that demands significant manual labor, cross-linking, and mental energy. Human underwriters naturally prioritize the "easier deal" to maximize throughput, opting to decline complicated files simply to make the administrative burden go away. By contrast, an AI engine executes neutral compute cycles regardless of file complexity, treating a file requiring cross-referencing across 10,000 documents with the exact same objective diligence as a file with one single document.

---

[05:48](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=348s) Eradicating Judgment and Rebuilding Consumer Trust

When investigating why marginalized communities avoid traditional mortgage applications—often turning instead to predatory options like payday loans—surveys consistently reveal a profound fear of mistreatment and judgment by bank personnel. Applicants carrying past medical debt, collections, or difficult life circumstances find it excruciating to watch a loan officer's facial expressions and body language in real time while exposing their financial vulnerabilities.

Interacting with an AI interface provides an emotional and psychological "hard reset" for the applicant. The machine evaluates cold facts without judgment, furrowed brows, or condescension. Agarwal references a think-tank study highlighting a massive divergence in institutional credibility: consumer trust in traditional bankers stands at a dismal 19%—a figure Agarwal notes is lower than public approval of Congress—whereas consumer trust in AI platforms exceeds 70%. Deploying fair, accurate, and predictable conversational technology directly addresses the emotional barrier preventing underbanked populations from entering the financial mainstream.

---

[08:05](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=485s) The Structural Advantage of In-House Bank Development over Venture Capital

Constantino notes the unusual pedigree of AngelAi, which was incubated directly inside Sun West Mortgage Company rather than launched as a venture-backed Silicon Valley startup. Agarwal asserts that the technology could never have survived the standard venture capital lifecycle. Traditional VC funds operate on short horizons and run out of patience within a couple of years, whereas building industrial-grade, bias-resistant financial AI required a continuous multi-decade commitment.

Many enterprise financial institutions sit on vast datasets that are worth more than the companies themselves, yet fail to extract value from them. By funding the platform internally—reinvesting $500 million of his own capital—Agarwal was able to align data science directly with live banking operations. The engineering team deployed modules step-by-step into the bank's production workflow, generating immediate commercial returns and validating compliance with fair-lending mandates before expanding the architecture into a fully autonomous, end-to-end platform in 2018.

- **Capital Investment:** $500 million in self-funded capital without reliance on outside venture funding. - **Development Feedback Loop:** Incremental modules were immediately deployed into live banking operations for validation and direct commercial returns.

---

[09:43](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=583s) Guarding Against Model Drift, Synthetic Data Collapse, and Uncertainty

Turning to technical governance, Constantino references an article he published in August 2024 regarding a *Nature* study on "model collapse"—the phenomenon where generative AI models trained recursively on their own synthetic outputs degrade into gibberish within five generations. Agarwal confirms that preventing model drift and collapse requires uncompromising data integrity. AngelAi strictly prohibits synthetic or AI-generated training data within its core decision engine, maintaining dedicated teams of data engineers who curate and verify raw empirical inputs daily.

In addition to securing multiple patents around predictable and precise algorithmic decisioning, the company recently open-sourced a research paper addressing real-world data uncertainty. While a human can distinguish between a cat and a dog after seeing only one or two examples, traditional AI architectures require millions of images to achieve similar categorization. Agarwal's team is focused on embedding analogous "intuition" into financial models, allowing the AI to navigate messy real-world variations and resolve discrepancies without hallucinating or defaulting to historical prejudice.

[12:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=724s) Explainability, Traceability, and the Black Box Dilemma

When an automated mortgage decision dictates whether an applicant can purchase a home, simple rejection notices fall short of true transparency. Tor Constantino asks what explainability ought to look like for denied borrowers, pressing beyond technical explanations tailored for data scientists or regulators. Pavan Agarwal points to guidance issued roughly two years earlier by the Consumer Financial Protection Bureau (CFPB) and the Federal Deposit Insurance Corporation (FDIC). The regulators established that while financial institutions are permitted to deploy artificial intelligence across banking operations and consequential decision-making, the technology cannot be treated as a "black box."

Crucially, Agarwal stresses that true explainability requires real-time traceability as the decision is being formulated. Simply submitting a backward-looking reprompt to ask an engine why it made a choice "doesn't count." Instead, the system must log every deterministic step in sequence. For Agarwal, building auditable decision logs into AngelAi was not merely a matter of satisfying regulatory mandates, but of protecting balance-sheet capital: with $40 billion entrusted to the platform, the institution itself demands complete clarity on why any loan is approved or declined.

```mermaid flowchart TD accTitle: Real-Time AI Decision Traceability Flow subgraph Ingestion["1. Loan Processing"] A["Application & Data Input"] --> B["Deterministic AI Evaluation"] end subgraph Audit["2. Traceability Engine"] B --> C["Step-by-Step Decision Log Generated"] B --> D["Immediate Decision Output"] end subgraph Delivery["3. Explainability Output"] C --> E["Consumer Explanation"] C --> F["Regulatory & Institutional Audit"] end ```

- CFPB and FDIC regulatory guidance permits comprehensive AI use provided decisions remain auditable and traceable rather than opaque black boxes. - Post-hoc conversational reprompting ("asking the AI why it did something") is deemed insufficient for compliance. - Sun West Mortgage built native step-by-step logging directly into AngelAi to safeguard more than $40 billion in capital assets.

---

[13:56](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=836s) Alternative Data, VantageScore, and Infrastructure Privacy Risks

Addressing how non-traditional credit data affects marginalized borrowers, Agarwal notes that the Federal Housing Finance Agency (FHFA) approved VantageScore models that incorporate unstructured data to calculate creditworthiness beyond legacy FICO scores. While expanding credit access is the goal, processing messy, unstructured documentation creates bottlenecks in conventional banking environments. In a traditional setting, dealing with unstructured files requires more human staff, increasing processor fatigue and leading to higher rates of arbitrary rejections. Automating the workflow end-to-end through AI removes human frustration from the equation by evaluating structured and unstructured files consistently without manual intervention.

Turning to data privacy, Agarwal expresses deep skepticism toward banks that route sensitive consumer records through third-party conversational AI platforms. If consumer data leaves an institution's dedicated data center, lenders lose visibility into whether proprietary borrower records are harvested to train external third-party models. To eliminate this exposure, Celligence and Sun West maintain their entire model infrastructure inside dedicated, secure data centers. When questioned about fraud risks associated with alternative data, Agarwal acknowledges the theoretical possibility of fabricated documentation from borrowers or brokers, but notes that automated cross-referencing against verified government data sources keeps actual fraud rates exceptionally low across the industry.

- The FHFA's approval of VantageScore enables lenders to leverage unstructured borrower data to assess creditworthiness outside traditional FICO frameworks. - Human review of unstructured files often leads to underwriter fatigue and increased loan denials, reinforcing the operational need for automated end-to-end evaluation. - External API calls to third-party commercial AI tools pose data leakage risks; AngelAi insulates sensitive borrower records by operating fully within private, secure data centers.

---

[17:39](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=1059s) Combating Redlining and Addressing Historical Industry Bias

Agarwal argues that the ability to ingest and analyze unstructured data rapidly serves as a vital safeguard against discrimination and redlining. When major lenders face enforcement penalties for redlining—deliberately avoiding low-income or minority neighborhoods—the underlying driver is often an unwillingness to handle complex income verification. Independent contractors, gig workers, and small business owners present documentation that is notoriously labor-intensive to parse. Human underwriters are essentially asked to possess analytical skills far surpassing certified public accountants (CPAs). Expecting standard underwriting staff to master intricate tax and accounting scenarios places them in an untenable position, resulting in blanket rejections that manifest as systemic discrimination. Deploying an AI engine synthesized with the collective knowledge of hundreds of thousands of CPA-level experts resolves this barrier instantly.

Looking toward measurable benchmarks over a five-year horizon, Agarwal highlights an academic study conducted at Loyola University. The researcher evaluated 1,000 real-world mortgage applications through prominent large language models like ChatGPT and Gemini. When running unadjusted baseline data, the systems mirrored historical lending disparities: for every two Black applicants denied a loan, one should have been approved under objective financial criteria. However, when the researcher removed race identifiers from the prompt inputs, the decision results equalized. Agarwal emphasizes that AI engines themselves are not inherently biased; rather, they reflect the flawed historical lending practices of the banking industry on which baseline models were trained.

- Modern redlining frequently stems from lenders refusing to process complex, non-standard income documentation from gig workers and small business owners. - AI systems trained on specialized accounting and tax rules alleviate the burden on loan officers and eliminate documentation-related rejections. - A Loyola University study of 1,000 real-world loan applications demonstrated that baseline AI models mirrored legacy denial rates—wrongfully rejecting 50% of creditworthy Black applicants—until racial identifiers were stripped from the data inputs.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=4s) partner in crime — Tor Constantino

> [02:44](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=164s) AI is like any other tool, it's only as good as the people behind it. — Pavan Agarwal

> [02:44](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=164s) easier deal — Pavan Agarwal

> [05:48](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=348s) hard reset — Pavan Agarwal

> [09:43](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=583s) model collapse — Tor Constantino

> [09:43](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=583s) intuition — Pavan Agarwal

> [12:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=724s) black box — Pavan Agarwal

> [12:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=724s) doesn't count — Pavan Agarwal

> [12:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=724s) asking the AI why it did something — Pavan Agarwal

## Mentioned Ai4, Tor Constantino, Pavan Agarwal, Forbes, AngelAi, Celligence International, Sun West Mortgage Company, ChatGPT, Gemini, Nature, Consumer Financial Protection Bureau (CFPB), Federal Deposit Insurance Corporation (FDIC), Federal Housing Finance Agency (FHFA), VantageScore, FICO, Loyola University

## Video description Pavan Agarwal - Creator of AngelAi, CEO of Celligence and Sun West Mortgage Tor Constantino - Writer / AI, Quantum Computing, Blockchain Consultant, Forb

How Artificial Intelligence Can Help Eliminate Discrimination in Financial Services

Can artificial intelligence make financial decisions fairer, or will it simply reinforce the biases already embedded in the system? Drawing on Pavan Agarwal’s experience, as president and CEO of Sun West Mortgage Company and founder and CEO of Celligence International, this fireside chat will examine how technology can reduce human inconsistency, expand access and improve the transparency of consequential decisions. The discussion will illustrate how explainable AI, auditable decision-making and meaningful human oversight can help institutions identify and mitigate bias without sacrificing speed or accuracy. Attendees will learn what responsible AI deployment looks like in practice, including the safeguards needed to prevent historical inequities from being replicated at scale. Join us for a candid conversation about how financial institutions can use AI not only to operate more efficiently, but also to serve customers more fairly.

[Source](https://www.youtube.com/watch?v=r2JxaLQfpdU)

中文简报

Sun West 部署 AngelAi 消除房贷核保偏差房贷歧视源于人工核保人员为避免行政疲惫而拒绝少数族裔的复杂档案，这是一种结构性缺陷，而自动化算法核保通过确定性地处理非结构化文档消除了这一问题。 Pavan Agarwal 解释了自 2018 年起在 Sun West Mortgage 内部开发并部署的 AngelAi 如何利用自动决策消除人为偏见，并大幅减少少数族裔和退伍军人借款人因差错导致的拒贷。他阐明，传统核保差异在很大程度上源于审批人员在处理复杂、非传统文件时的疲劳，而非刻意恶意，这使得不知疲倦的 AI 计算周期天然更加公平。为符合监管标准并保护消费者隐私，该平台采用确定性、实时可追溯的审计日志，且完全在私有数据中心内运行，而非依赖第三方 API。讨论最终强调，尽管未调整的模型可能会反映行业历史偏见，但剥离了人口统计特征标记的中立算法架构能够有效消除系统性信贷歧视。 AngelAi 自 2018 年起投入商业化实际运营，已通过自动化核保处理了超过 400 亿美元的交易，服务了逾 20 万个家庭。 (0:04) 该平台历经数十年演进，从 20 世纪 80 年代基于规则的算法和 90 年代的数据库向量化，逐步发展为首个自动化反向抵押贷款引擎和对话式 AI。 (1:06) 对多年银行数据的分析揭示了人工核保中惊人的误拒率：在被拒绝的申请人中，本应获批的包括二分之一的黑人申请人、八分之三的西裔申请人以及三分之二的退伍军人。 (2:44) 核保偏见往往源于认知疲劳而非刻意恶意，人工审批人员会通过拒绝复杂、非传统的档案来减少手工工作量。 (2:44) 与 AI 互动消除了对人为居高临下和评判的恐惧，促使消费者对 AI 借贷平台的信任度超过 70%，而对传统银行机构人员的信任度仅为 19%。 (5:48)

Reading summary

Pavan Agarwal, creator of Sun West Mortgage Company's AngelAi platform, argues that automated artificial intelligence can systematically reduce lending disparities by removing the cognitive friction and judgment inherent in manual underwriting. Drawing on operational data from over $40 billion in transactions, Agarwal contends that disparate denial rates stem largely from human operational fatigue rather than malicious intent. Operational Fatigue as Root Cause of Disparities (2:44) Agarwal asserts that historical lending disparities frequently originate from administrative burdens rather than overt racial bias. An analysis of seven to eight years of commercial bank mortgage data revealed striking decline-to-error ratios: one out of two declined Black applicants, three out of eight declined Hispanic applicants, and two out of three declined military veterans should have been approved. According to Agarwal, minority applicants, gig workers, and small business owners often submit non-traditional income streams and complex documentation. Faced with manual cross-referencing across numerous files, human underwriters face fatigue and prioritize straightforward applications to maintain volume, opting to decline complex files to eliminate administrative effort. Mitigating Judgment to Rebuild Consumer Trust (5:48) Borrowers from historically underserved backgrounds frequently avoid traditional mortgage channels due to fear of face-to-face scrutiny and condescension over past credit events or medical debt. Agarwal cites think-tank research indicating that consumer trust in traditional commercial bankers sits at 19%, compared to over 70% for AI platforms. Conversational AI interfaces evaluate raw financial facts without visible reactions or furrowed brows, creating a neutral environment that lowers the psychological barrier for underbanked populations seeking credit. Technical Governance, Privacy, and Explainability Standards (9:43) To prevent algorithmic degradation, AngelAi prohibits synthetic training data within its primary decision engine, relying on dedicated data teams to curate empirical inputs and counter model collapse. Regulatory frameworks from the CFPB and FDIC require non-opaque decision-making; Agarwal emphasizes that true compliance demands step-by-step deterministic audit logs created in real time rather than post-hoc conversational reprompts. Furthermore, to guard against proprietary data leakage, models operate within private, dedicated data centers rather than external commercial AI APIs. Unstructured Data Integration and Neutral Evaluation (13:56) While the Federal Housing Finance Agency's adoption of VantageScore facilitates credit assessment via unstructured data, processing these records manually intensifies operational fatigue in traditional underwriting environments. Agarwal notes that automated parsing removes this barrier. Highlighting a Loyola University study of 1,000 mortgage applications, Agarwal explains that commercial LLMs initially mirrored systemic biases—wrongfully denying half of qualified Black applicants—until racial identifiers were removed from prompts, demonstrating that AI bias mirrors historical training baselines rather than inherent machine flaws. While automated underwriting can neutralize human administrative fatigue and remove overt demographic markers, maintaining fair lending requires ongoing governance to protect against proprietary data leakage, unexplainable black-box models, and the risk of algorithms ingesting historical industry prejudices.

精炼阅读摘要

金融科技高管Pavan Agarwal认为，传统信贷审批中的歧视多源于人工审核面对非标复杂材料时的疲劳与规避，而非刻意偏见；通过全流程自动化、确定性追踪与隔离种族标识的人工智能决策系统，金融机构能够在消除审核偏见的同时重塑消费者的制度信任。人工审核负荷与历史信贷偏差的根源 (2:44) Agarwal分析了商业银行7至8年的政府房贷数据，指出传统人工审核存在显著的误拒偏差：在被拒批的申请人中，黑人借款人每2人中有1人本应获批，退伍军人每3人中有2人本应获批，拉美裔借款人每8人中有3人本应获批。Agarwal认为，这种偏差通常不是源于核保员主观的恶意歧视，而是因为少数族裔、零工从业者及小微企业主往往提交非传统的复杂财务材料，需要耗费巨大的精力去交叉比对。在业务量的压力下，人工核保员往往优先处理简单文件，倾向于直接拒绝复杂案例以降低行政负担。相比之下，AI在处理1份文件与跨1万份文件进行交叉验证时保持完全相同的客观标准与算力消耗。针对传统红线政策（Redlining）的调查也显示，机构规避少数族裔社区的根源常在于无力核验复杂的非标收入；如果要求审核员同时具备超越注册会计师（CPA）的税务分析能力并不现实，而集成专业会计规则的AI系统可以直接化解这一瓶颈。Agarwal援引洛约拉大学（Loyola University）对1000份真实房贷申请的研究指出，通用大模型在默认状态下会复刻历史偏见（同样导致约50%合格黑人借款人被拒），但只要在输入端剥离种族标识，决策结果即可实现公平，证明技术偏差主要源于历史训练数据而非算法本身。消除人际评判与重塑弱势群体的制度信任 (5:48) 传统信贷场景中，边缘化群体往往因惧怕银行人员的负面评判而放弃申请房贷，甚至被迫转向发薪日贷款等掠夺性金融工具。调研表明，背负医疗债务或催收记录的申请人在面对贷款专员的面部表情和肢体语言时会产生强烈的心理压力。Agarwal指出，交互式AI能够提供情感上的“硬重置”，机器仅评估客观事实而不带任何冷嘲或皱眉动作。这种交互差异直接体现在公众信任度上。Agarwal引用智库研究数据称，消费者对传统银行家的信任度仅为19%，低于公众对国会的支持率；而在使用具备可预测性的对话式AI时，消费者信任度可超过70%。通过消除人工审查过程中的人际审判压力，合规、透明的AI系统能够降低弱势群体进入主流金融体系的心理门槛。实时可追溯性与规避“黑箱”监管合规 (12:04) 针对自动化决策引发的黑箱质疑，美国消费者金融保护局（CFPB）与联邦存款保险公司（FDIC）发布监管指引，明确允许银行在核心决策中使用AI，但严禁不可解释的黑箱操作。Agarwal强调，合规的可解释性不能依赖事后重新提示（即回过头询问AI为何做出某项决策），而必须在决策生成的同时记录确定性的逐步追踪日志。每一笔审批或拒绝都必须生成连贯的决策日志，以同时满足消费者解释诉求与监管审计标准。在底层架构上，为防止数据漂移和“模型崩溃”（即模型在自身生成的合成数据上递归训练导致性能退化），其研发团队坚决禁止在核心决策引擎中使用合成数据或AI生成的数据，由专门的数据工程团队每日维护与核实原始经验数据。此外，团队正致力于研发类似人类直觉的架构，使模型在面对现实世界中杂乱的差异和不确定性时，能够在不产生幻觉或不陷入历史偏见的前提下解决数据冲突。自研闭环架构与非标数据下的隐私防护 (13:56) 在信贷评估标准的演进中，联邦住房金融局（FHFA）已批准引入非结构化数据的VantageScore模型，使没有传统FICO高分的借款人也能评估信用。然而，在传统银行体系中，引入非结构化数据会加剧人工处理疲劳，进而推高随意拒贷的概率，唯有通过端到端全自动评估才能消除这种人为摩擦。针对非标数据的欺诈隐患，Agarwal表示通过自动化系统与政府验证数据源交叉比对，实际欺诈率在业内保持在极低水平。在基础设施与数据安全方面，Agarwal对依赖第三方通用大模型API调用的银行模式提出警示，指出一旦敏感借款人数据离开专用数据中心，机构将丧失对私有记录是否被用于训练第三方模型的控制权。Agarwal介绍，其系统从底层即依托内部5亿美元资金在Sun West Mortgage内部孵化并部署于私有数据中心，避免了传统风险投资追逐短期回报的弊端，在保障数据隐私的同时实现了信贷模块在生产环境中的合规验证与直接落地，自2018年全面投产以来已处理超400亿美元交易并服务超20万个家庭。该对谈的核心观点是，信贷审批中的歧视主要是人工处理高复杂度非标材料时的产能逃避，而非单纯的道德偏见；以私有数据中心、实时日志审计和剥离种族标识为底座的专用AI能够纠正这一结构性缺陷。然而，这一路径仍依赖于金融机构对底层数据工程的持续投入与严格的私有化部署，行业在广泛整合替代数据时如何平衡算力成本与数据主权仍有待观察。

Detailed notes — zh

## Background 本视频由 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上。内容为《福布斯》人工智能专栏作家 Tor Constantino 与 Sun West Mortgage 首席执行官兼 Celligence 创始人 Pavan Agarwal 展开的炉边对谈，深入探讨金融机构如何借助可解释、可审计且具备监督机制的 AI 技术消除历史偏见，在兼顾效率与准确性的同时实现更公平的信贷决策。 ## TL;DR Pavan Agarwal 结合其投入 5 亿美元自有资金研发并实际促成超 400 亿美元交易的 AngelAi 系统，剖析了传统银行体系在信贷审批中对少数族裔和退伍军人的高误拒率真相。他指出，传统核保偏见往往源于人类审核员面对繁琐非标材料时的精力疲劳与避难心理，而 AI 凭借纯粹的算力与客观性不仅能高效处理非结构化数据，还能消除借款人面对人工审判时的羞耻感。此外，对谈强调了金融级 AI 必须摒弃事后追问的“黑盒”模式，通过生成实时可追溯的推理审计日志与私有化数据闭环，从底层防御模型崩塌与数据泄露风险。 ## Atmosphere & style 整场对话洋溢着务实、严谨且极具同理心的专业基调，主持人的敏锐追问与嘉宾数十年的信贷实战积淀形成了极具张力的思想碰撞。Agarwal 的叙述既有详实的数据与专利技术支撑，又流露出对普通借款人尊严的深切关怀，彻底打破了外界对金融与算法冰冷疏离的固有印象。交流节奏紧凑平实、直击行业弊病，为观众呈现了一场兼具技术深度与人文温度的金融科技启示录。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=4s) Sun West 研发的 AngelAi 早在 2018 年即投入生产，至今已累计处理超 400 亿美元交易并服务逾 20 万个家庭，从技术底层推动消除金融信贷准入门槛。 - [02:44](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=164s) 数据分析表明传统大银行存在严重的系统性误拒：非裔申请人每被拒贷 2 人就有 1 人本应获批（误拒率 50%），西裔误拒率为 37.5%，退伍军人误拒率更是高达 66.7%。 - [04:20](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=260s) 信贷偏见的本质往往是人工核保的“避繁就简”惰性：少数族裔的财务材料通常更繁琐非标，核保员因疲劳和心智负荷倾向于直接拒贷，而 AI 能不带情绪地完成全部计算时钟周期与客观交叉验证。 - [05:48](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=348s) 机器交互消除了面对面信贷审批带来的羞耻感与被审判感：消费者对银行家的信任度仅有 19%，而对 AI 的信任度超 70%，无偏见的中立界面成为降低准入门槛的“硬重启”。 - [08:05](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=485s) AngelAi 依托 5 亿美元自有资金并在成熟银行内部历经数十年研发闭环，依靠真实核心业务数据滚动反哺与模块化即时变现，走通了传统急功近利的风险投资无法实现的落地路径。 - [09:43](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=583s) 防范“模型崩塌”的关键在于维护底层训练数据纯度：坚决杜绝 AI 合成数据自我反哺，并结合专有算法让 AI 具备类似人类的直觉，以有效应对现实世界的数据变异与不确定性。 - [12:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=724s) 信贷决策的 AI 可解释性必须是“实时可追溯”而非“事后提问”：根据 CFPB 等监管要求，系统必须在决策执行时同步生成每一步推理的端到端审计日志，彻底拒绝黑盒化运作。 - [13:56](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=836s) 新型信用评分机制（如 VantageScore）引入非结构化数据后更需端到端 AI 处理，且金融机构必须在自建高安全私有数据中心运行系统，避免将敏感数据流向第三方模型造成泄漏与二次训练污染。 - [16:51](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=1011s) AI 解析非结构化数据的能力有助于破除“红线歧视”：许多银行拒贷并非出于恶意，而是基层人员缺乏注册会计师（CPA）级别的查账能力，而具备专业财务知识的 AI 能精准穿透零工和小微业主的真实偿付能力。 - [19:35](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=1175s) 行业基准测试表明，主流大模型的决策偏差源自金融业长久以来的历史惯性数据；只要在算法中明确识别并剔除种族等敏感变量，AI 就能摆脱历史偏见、带来更加公平的决策。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=4s) 破局金融歧视：AngelAi 的诞生与四十年探索对谈在一场关于金融与科技的对话中拉开序幕。主持人 Tor Constantino 介绍自己是《福布斯》（Forbes）的人工智能专栏作家，并引出嘉宾——Sun West Mortgage 首席执行官兼 Celligence 创始人 Pavan Agarwal。Agarwal 自嘲且自豪地回顾了自己在金融服务领域长达 45 年的从业资历。早在 2018 年，他就带领团队将自主研发的 AngelAi 投入全面生产运营。谈及此，他笑着调侃道："我知道大家肯定会想，2018 年那可是在 ChatGPT 问世之前。"然而正是从那时起，该系统已累计完成了超过 400 亿美元的交易规模，通过自动化的 AI 决策与银行业务执行，帮助了逾 20 万个家庭跨越金融门槛，直接从技术底层遏制歧视。 Agarwal 步入金融科技领域的契机颇具传奇色彩。作为家族企业的第二代传人，他在 11 岁时就开始坐在 PC 机前，为公司的金融业务编写代码解决深奥的数学难题。随着对信贷业务理解的加深，他着手构思能够完全自动化贷款决策的算法雏形。然而在 20 世纪 80 年代，算力条件的极端受限使得早期的 AI 算法根本无法投入实际生产。真正的技术突破发生在 90 年代，团队成功将数据库和审批决策流程进行了“向量化”（vectorized）。到了 2000 年代，他们推出了业内首个针对反向抵押贷款（reverse mortgage）的 AI 自动化核保系统。最终在 2018 年，所有底层模块整合成了一套端到端的完整解决方案。借款人无需再面对冰冷机械的表格，而是与具备同理心的 AI 展开对话，倾诉自身的财务现状、生活目标与家庭梦想，而 AI 则会在后台通过精准的金融工程计算，为用户量身定制最契合的金融方案。 - **上线时间**：AngelAi 于 2018 年全面投入商业生产环境，早于主流大语言模型热潮。 - **业务体量**：累计促成超 400 亿美元交易，惠及超过 200,000 个家庭。 - **技术演进**：80 年代受制于算力瓶颈；90 年代实现数据库与决策机制向量化；2000 年代推出反向抵押贷款自动化核保。 --- ### [02:44](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=164s) 揭示系统性不公：真实数据背后的被拒群体 Tor Constantino 提出了当下业界的普遍担忧：现有的人工智能多依赖历史数据进行训练，而历史数据往往凝结了过去数十年在信贷与房屋产权获取上的不平等现象；金融机构究竟该如何防止算法在极高速度和超大规模下复制甚至加剧这些历史不公？Agarwal 强调，AI 和其他任何工具一样，“它的成色完全取决于背后掌控它的人类”。作为第一代移民，Agarwal 亲眼目睹了太多家庭在传统银行体系中遭遇的不公对待，因此 fair lending（公平借贷）与平等获取机会从一开始就被刻进了企业的愿景。为了说明传统银行体系中偏见的严重程度，Agarwal 援引了团队对过去七至八年间官方及行业公开数据的深入分析。分析表明，传统银行的信贷拒绝充斥着惊人的误判： ```mermaid flowchart TD accTitle: 传统大银行拒绝借款人误判统计 subgraph Stats["大银行信贷审批误判率统计"] direction TB A["非裔申请人 （Black Borrowers）"] -->|每拒绝 2 人| A1["本应批准 1 人 （误拒率 50%）"] B["西裔申请人 （Hispanic Borrowers）"] -->|每拒绝 8 人| B1["本应批准 3 人 （误拒率 37.5%）"] C["退伍军人 （Veterans）"] -->|每拒绝 3 人| C1["本应批准 2 人 （误拒率 66.7%）"] end ``` 面对如此悬殊的失衡，Agarwal 指出，这些边缘化群体的痛点正是团队研发 AI 的基石。AngelAi 的训练依托于企业数十年服务这些弱势社区所积累的第一手真实解决经验，而不是盲目继承大银行充满歧视性结果的历史审批记录。 --- ### [04:20](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=260s) 偏见的潜意识路径：为什么繁琐的审批催生了拒绝谈及金融机构内部偏见的本质时，Agarwal 打破了外界的刻板印象。他指出，现实中几乎没有哪位核保员或信贷经理会在看材料时公然心想：“这个人是非裔或西裔，所以我直接拒掉。”真实的偏见往往披着“避繁就简”的人性外衣。受历史和现实经济处境影响，边缘化社群或少数族裔借款人的财务证明材料往往更加繁琐非标。虽然他们同样具备还款能力，对拥有属于自己的家园抱有同样的渴望，但核保这类案件所需的精力和心智负荷远高于标准借款人。面对堆积如山的复杂材料，核保员在潜意识中倾向于选择更轻松的单子，为了“让麻烦快速消失”，最简单的途径就是直接出具一纸拒信。 ```mermaid flowchart LR accTitle: 人类核保与 AI 处理复杂贷款申请的路径对比 subgraph Human["人工核保路径"] H1["复杂非标材料"] --> H2["耗费脑力/疲劳"] H2 --> H3["趋易避难心理"] H3 --> H4["直接拒贷出局"] end subgraph Engine["AI 自动化路径"] A1["复杂/多来源材料"] --> A2["时钟周期运算"] A2 --> A3["全量数据交叉验证"] A3 --> A4["客观精准决策"] end ``` 这种因繁琐导致的隐性排斥在传统借贷中屡见不鲜。而 AI 则从根本上终结了这一人性惰性：算力不具备厌烦情绪，无论输入的是 1 份文件还是 10,000 份文件，面对多么错综复杂的交叉索引，AI 都会一丝不苟地走完每一个计算时钟周期，给出严格基于财务实质的客观结论。 --- ### [05:48](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=348s) 消除面对面的羞耻感：当 AI 的信任度超过银行家 Tor Constantino 指出，人类决策往往不可避免地带来不一致性、主观性以及无意识偏见。在探讨贷款流程中哪些环节最能通过 AI 获得实质性改善时，Agarwal 将焦点放在了借款人的心理门槛与交互前端。长期以来，处于劣势地位的借款人宁愿承受高昂成本依赖发薪日贷款（payday loans），也不愿步入主流银行机构。调研显示，他们最普遍的心声是：“我觉得银行家看不起我”、“我觉得自己得不到公平对待”、“我害怕被审判”。当申请人不得不当面向信贷员解释因突发重大疾病、家庭变故导致的医疗欠款或催收记录时，对方脸上哪怕一丝微妙的表情变化，都会造成巨大的心理压力与屈辱感。 AI 的介入彻底切断了这种因人际互动产生的情绪壁垒。面对机器，借款人无需再揣摩对方的面部表情，可以放下防御、毫无顾虑地坦陈真实的财务窘境。Agarwal 援引一家智库的调研数据展示了这一强烈反差： - 普通消费者对银行家的信任度仅有 **19%**（该比例甚至低于民众对国会的信任度）； - 消费者对 AI 的信任度则高达到 **70% 以上**。在 Agarwal 看来，向大众提供一个公正、可靠、能给出精准答复的机器交互界面，是消除金融服务准入门槛和心理偏见的最有效“硬重启”（hard reset）。 --- ### [08:05](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=485s) 5亿美元自有资金与内部打磨：为何风投资本无法复制 AngelAi Tor 提及一个关键背景：AngelAi 并非脱胎于独立的科技创业公司，而是在一家成熟的持牌按揭银行内部孕育而成。Agarwal 直言，如果走传统的风险投资（VC）孵化模式，这类产品根本不可能诞生。风投资本通常缺乏长期主义，往往在两至三年内就耗尽耐心要求回报，而这套金融 AI 体系的打磨耗费了二十余年。更致命的是，初创公司往往无法触及真正的银行核心业务和底层全量借贷数据。Agarwal 表示，许多大型企业坐拥价值连城的数据资产，却对自身的“金矿”茫然不知。在研发过程中，Agarwal 累计投入了 5 亿美元的自有资金，从未依赖任何外部融资。这种“在银行内部造 AI”的架构带来了无可比拟的落地闭环： 1. **真实数据反哺**：算法直接在母机构的核心数据流与复杂业务场景中进行数据工程治理与模型训练。 2. **模块化即时变现**：团队每研发出一个新功能模块，立即部署到银行自身的实际审批业务中产生效能回报。 3. **步步迭代闭环**：通过长达数十年的滚动推进，各独立模块逐步贯通，最终在 2018 年实现了全流程端到端的自主化决策。 --- ### [09:43](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=583s) 对抗“模型崩塌”与数据不确定性：让算法拥有直觉针对 Tor 提出的核心质询——“如何证明 AI 是真正让决策更公平，而不仅仅是以更高的连贯性输出基于错误假设的决定”，Agarwal 结合 Constantino 近期发表的技术专栏，深入探讨了模型漂移（model drift）与模型崩塌（model collapse）的防治机制。 Agarwal 指出，确保算法公正的前提是绝对保证数据的纯度与真实性。如果系统陷入“让 AI 生成的数据反向喂养 AI”的死循环，模型必然发生灾难性劣化。在 AngelAi 的研发管线中，团队配置了庞大的数据工程师团队进行全天候监控，坚决杜绝任何由合成数据（AI-generated data）污染训练集的可能。此外，Agarwal 团队在可预测分析与高精度决策领域布局了多项核心专利，并在近期开源了一篇最新学术论文，攻克 AI 在“现实世界数据不确定性”（real-world data uncertainty）下的推理难题： - **人类认知与 AI 的差距**：人类在见过一两只猫或狗之后，就能凭直觉迅速认出下一只动物；而传统 AI 必须吞吐数百万张图片才能建立基本识别。 - **构建机器直觉**：团队的最新算法正在赋予 AI 类似人类的“直觉”，使其能够从现实世界的噪声与数据变异中准确辨识出核心差异，做出高度稳健的判断。 Tor 随后补充了自己 2024 年 8 月撰写专栏的背景：该文章引述了一篇发表在《自然》（Nature）杂志上的重磅研究，证实一旦 AI 模型连续接受自身输出数据训练超过 5 个世代（five levels），就会彻底退化并输出完全不可理喻的胡话（gibberish）。对于金融借贷这类涉及大额资金与法律法规的封闭严密体系，这一崩塌风险尤为致命。 ### [12:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=724s) AI 决策的可解释性与监管合规要求在抵押贷款业务中，自动化决策直接决定了一个人是否有资格购买住房。主持人 Tor Constantino 提出，对于被拒贷的申请人而言，可解释性（Explainability）应当呈现出何种形态——申请人需要得到怎样的具体反馈，而不仅仅是收到一个冰冷的“拒绝”通知。 Pavan Agarwal 指出，美国消费者金融保护局（CFPB）和联邦存款保险公司（FDIC）在大约两年前就对此出台了明确的监管指引：银行完全获准利用 AI 来运行内部系统并做出信贷决策，但核心前提是**绝对不能将 AI 当作“黑盒”（Black Box）来运作**。监管机构要求的可解释性并非在事后向大模型重新发问（Reprompting）“你刚才为什么做出这个决定”，因为事后追问并不具备可信的解释力。真正的合规要求系统具备全流程的“可追溯性”（Traceability）——即在 AI 做出决策的同时，实时记录并生成每一步推理与计算的审计日志（Log）。 Pavan 强调，这正是其自研平台 AngelAi 的底层设计核心。对 Sun West 而言，系统托付着高达 400 亿美元的资金规模，“即使抛开监管要求不谈，赌上的是我们自己的真金白银，我们必须百分之百清楚系统决策背后的原因”。因此，系统在架构底层就内嵌了可审计机制，无论是批准还是拒贷，平台都能针对任何一笔申请，向消费者和监管机构提供详尽的端到端决策步骤拆解。 ```mermaid flowchart TD accTitle: 可解释性合规决策流程对比 subgraph S1["合规的实时可追溯架构 (AngelAi)"] A["贷款申请数据输入"] --> B["AI 实时决策执行"] B --> C["同步生成逐步推理审计日志 (Step-by-Step Log)"] C --> D["面向用户/监管输出透明解释报告"] end subgraph S2["不合规的黑盒模式 (Black Box)"] E["数据输入"] --> F["黑盒 AI 给出拒贷/批准结果"] F -.->|事后反向提示 (Reprompt)| G["'你为什么拒贷？' 产生不可靠解释"] end ``` --- ### [13:56](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=836s) 非结构化数据、新型信用评分与隐私安全边界针对传统信用评分（如 FICO）无法充分覆盖弱势群体的问题，Tor 询问了引入非结构化数据或替代数据（Alternative Data）在扩大信贷准入方面的作用，以及这是否会带来新的歧视或隐私泄露风险。 Pavan 解释称，美国联邦住房金融局（FHFA）近期批准了名为 **VantageScore** 的新型评分机制。与传统 FICO 评分不同，VantageScore 能够纳入非结构化数据来综合评估信用。然而，这给传统借贷机构带来了巨大的技术挑战：贷款机构必须能够接收、解析并准确理解这些复杂的非结构化数据，再将其报送给信用局（如 TransUnion）以生成准确评分。在传统人工审核流程中，非结构化数据越多，涉及的人员链条就越长，审核员极易因繁琐的数据处理而产生挫败感，最终草率拒贷。因此，数据越是非结构化，就越需要端到端的全自动 AI 来接管，消除人工干预的不确定性。在隐私与数据安全方面，Pavan 明确指出了当前行业的潜在风险： - **核心合规底线**：金融机构依法必须将消费者敏感数据保留在自身的高安全级别数据中心内。 - **第三方调用隐患**：Pavan 对部分银行将客户数据直接发送给外部第三方商业 AI/聊天引擎获取结果的做法提出了严厉质疑。一旦数据流出机构专属的数据中心，机构根本无法掌控第三方是否会将这些私密金融数据用于其基础模型的二次训练。 - **私有化安全架构**：AngelAi 坚持将所有模型与数据闭环部署在自有的受保护服务器和数据中心内部，完全规避外部数据泄露与训练污染风险。 --- ### [16:51](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=1011s) 欺诈防范与消除“红线歧视” 在谈到非结构化数据可能引发的欺诈与歧视问题时，Pavan 进行了深入剖析： - **欺诈风险（Fraud Risk）**：尽管非结构化材料（如散户自制凭证）理论上更容易被借款人、房产中介或贷款经纪人伪造，但在现代风控体系下欺诈概率极低。金融机构能够通过大量的官方及政府权威数据源进行交叉比对与实时验证，行业现有的反欺诈能力已经非常成熟。 - **消除红线歧视（Redlining）**：Pavan 认为，AI 高速处理非结构化数据的能力不仅不会加重偏见，反而能从根本上减少歧视。许多大型银行曾因“红线歧视”（即蓄意拒绝向低收入、少数族裔或服务匮乏社区提供贷款）而遭受监管重罚。银行往往并非出于恶意主观排斥，而是因为这些社区包含大量零工经济从业者（Gig Workers）或小微企业主，其收入来源虽合法但难以用标准化格式证明。 - **超越人类 CPA 的能力边界**：传统模式下，银行要求基层信贷审核员解析极其繁杂的非结构化财务文档，这无异于要求普通技术人员具备“超越注册会计师（CPA）级别”的专业能力，将员工置于无法胜任的境地。当审核员因处理不了而拒贷时，机构便落入了歧视指控。相比之下，经过数十万专业会计师与专家知识训练的 AI 能够迅速穿透非结构化数据的复杂表象，客观核算真实偿付能力。 --- ### [19:35](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=1175s) 行业历史数据偏见与未来五年前景当被问及未来五年内需要何种证据才能证明 AI 真正减少了金融歧视时，Pavan 引用了来自洛约拉大学（Loyola University）学者进行的一项突破性实证研究： - **基准测试**：该研究选取了 1,000 笔真实的抵押贷款案例，将其输入主流大语言模型（如 ChatGPT、Gemini 等）进行审核测试。 - **揭示行业偏见**：模型最初的输出重现了现实中的严重偏差——在每两位被拒绝的非裔借款人中，实际上就有一位完全符合批准标准却被误拒。 - **剔除种族变量**：研究人员随后在输入数据中完全移除了“种族”（Race）这一变量，将相同的数据再次输入同一批模型，结果迅速趋于公平与均衡。 Pavan 总结道，这表明各大基础模型本身表现出的偏差，本质上是因为它们学习了借贷行业过去存在偏见的历史存量数据。“责任不在于 AI 引擎本身，而在于底层数据的源头——即整个金融行业长期以来的历史操作惯性。”通过清晰识别并剔除带有偏见的历史特征变量，AI 恰恰能够打破行业惯性，建立更加公平公正的金融服务新范式。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=4s) 我知道大家肯定会想，2018 年那可是在 ChatGPT 问世之前。 — Pavan Agarwal > [02:44](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=164s) 它的成色完全取决于背后掌控它的人类 — Pavan Agarwal > [04:20](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=260s) 这个人是非裔或西裔，所以我直接拒掉。 — Pavan Agarwal > [05:48](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=348s) 我觉得银行家看不起我 — Pavan Agarwal > [05:48](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=348s) 我觉得自己得不到公平对待 — Pavan Agarwal > [05:48](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=348s) 我害怕被审判 — Pavan Agarwal > [12:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=724s) 绝对不能将 AI 当作“黑盒”（Black Box）来运作 — Pavan Agarwal > [12:04](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=724s) 即使抛开监管要求不谈，赌上的是我们自己的真金白银，我们必须百分之百清楚系统决策背后的原因 — Pavan Agarwal > [19:35](https://www.youtube.com/watch?v=r2JxaLQfpdU&t=1175s) 责任不在于 AI 引擎本身，而在于底层数据的源头——即整个金融行业长期以来的历史操作惯性。 — Pavan Agarwal ## Mentioned Pavan Agarwal, Tor Constantino, AngelAi, Celligence, Sun West Mortgage Company, Forbes, ChatGPT, Gemini, Nature, CFPB (Consumer Financial Protection Bureau), FDIC (Federal Deposit Insurance Corporation), FHFA (Federal Housing Finance Agency), FICO, VantageScore, TransUnion, Loyola University ## Video description Pavan Agarwal - Creator of AngelAi, CEO of Celligence and Sun West Mortgage Tor Constantino - Writer / AI, Quantum Computing, Blockchain Consultant, Forb How Artificial Intelligence Can Help Eliminate Discrimination in Financial Services Can artificial intelligence make financial decisions fairer, or will it simply reinforce the biases already embedded in the system? Drawing on Pavan Agarwal’s experience, as president and CEO of Sun West Mortgage Company and founder and CEO of Celligence International, this fireside chat will examine how technology can reduce human inconsistency, expand access and improve the transparency of consequential decisions. The discussion will illustrate how explainable AI, auditable decision-making and meaningful human oversight can help institutions identify and mitigate bias without sacrificing speed or accuracy. Attendees will learn what responsible AI deployment looks like in practice, including the safeguards needed to prevent historical inequities from being replicated at scale. Join us for a candid conversation about how financial institutions can use AI not only to operate more efficiently, but also to serve customers more fairly. [Source](https://www.youtube.com/watch?v=r2JxaLQfpdU)

Source quotes

[0:04] Pavan Agarwal: 我知道大家肯定会想，2018 年那可是在 ChatGPT 问世之前。 [2:44] Pavan Agarwal: 它的成色完全取决于背后掌控它的人类 [4:20] Pavan Agarwal: 这个人是非裔或西裔，所以我直接拒掉。 [5:48] Pavan Agarwal: 我觉得银行家看不起我 [5:48] Pavan Agarwal: 我觉得自己得不到公平对待 [5:48] Pavan Agarwal: 我害怕被审判 [12:04] Pavan Agarwal: 绝对不能将 AI 当作“黑盒”（Black Box）来运作 [12:04] Pavan Agarwal: 即使抛开监管要求不谈，赌上的是我们自己的真金白银，我们必须百分之百清楚系统决策背后的原因 [19:35] Pavan Agarwal: 责任不在于 AI 引擎本身，而在于底层数据的源头——即整个金融行业长期以来的历史操作惯性。

Source quotes

[0:04] Tor Constantino: partner in crime [2:44] Pavan Agarwal: AI is like any other tool, it's only as good as the people behind it. [2:44] Pavan Agarwal: easier deal [5:48] Pavan Agarwal: hard reset [9:43] Tor Constantino: model collapse [9:43] Pavan Agarwal: intuition [12:04] Pavan Agarwal: black box [12:04] Pavan Agarwal: doesn't count [12:04] Pavan Agarwal: asking the AI why it did something

Original transcript

[0:04] As our hostess said, I'm Tor Constantino. I have an AI column at Forbes that I write about and [0:08] um I'd like to turn over to my uh partner in crime here so that I can hear his [0:12] uh his exemplary background and introduction. Please Pavon. Yeah. Hi everybody. I'm Pavan [0:19] Agaral and my background is basically 45 years in financial services. Um and in that time period I [0:27] built Angel AI went into full production in 2018 and and I know you all like 2018 that's [0:34] before chat GPT [snorts] um yeah but 2018 went to full production since then we have transacted [0:41] uh 40 billion plus and over 200,000 uh families have been helped by AI and ending discrimination [0:49] with automated AI decisions and uh banking execution. Amazing. Um, can you tell us a [0:57] little bit about uh I mean when we talked before uh last week you were saying how you [1:02] got into the business if you could just provide a little bit of context for the for the audience [1:06] because I think it's fascinating. Okay. Um so I'm actually second generation in this business. So [1:12] um I literally started when I was like 11 years old and I started writing um solving um esoteric [1:20] um math problems uh on a PC for the uh financial services and working inside uh our company and [1:29] from there I started writing algorithms to fully automate um and automate the decision process of [1:36] of lending and um that's that meant like early AI I algorithms that I worked on was like in in in [1:43] the 80s the computing power wasn't sufficient to actually put them in production. Um first major [1:49] breakthrough that we had on the AI track was in the was in the '9s when we um when we vectorized [1:56] actually the database when we vectorized the the uh the decisioning process. Um, fast forward, [2:04] um, in in the 2000s, we we we created the first AI automated underwriting for something called a [2:11] reverse mortgage. Um, and I'm almost there. I'm a few years away. I'd be able to qualify for a [2:17] reverse mortgage. And, uh, then in 2018, the whole thing um, came together as a complete end to end. [2:26] You have a an empathetic conversation with an AI. It it chat, you chat with it. You tell it [2:32] your financial information. You you tell it what your goals and aspirations and and dreams are and [2:37] the AI will structure and engineer a financial solution for you. Uh fascinating model that you [2:44] designed and built there. Unfortunately, a lot of the AI models that are out there right now, [2:49] they learn from historical data and that data may reflect decades of unequal access [2:55] uh to credit home ownership. How can financial institutions prevent an algorithm from reproducing [3:03] those inequities at scale and speed? Yeah, that's a great question because AI is like any [3:09] other tool, it's only as good as the people behind it, right? It's it's the hands in which the tools [3:14] uh controlled by. So, uh, you know, the the the mission and vision of our company has always been [3:20] fair lending and equal access and and being first generation immigrants, we saw firsthand of of so [3:26] many families that were not properly treated by the banking system. So for example, um there's the [3:35] real world data, government data, uh we analyzed the past seven, eight years, which is that for [3:41] every two black borrowers that are declined by the big banks, one should have been approved. Okay? [3:47] And every uh eight Hispanics that are declined, three should have been approved. Okay? And and the [3:52] worst one is that for every three veterans that are declined, two should have been approved. Okay? [3:57] Okay. So, there is this this underlying problem in the banking system um that we've always fought [4:03] against and our our AI is built on data um from our experience and what we've done everything [4:12] that we've done to solve the problems for um these underserved communities. Uh so when we talk about [4:20] um bias in financial services, what are forms of bias in in your experience that are most [4:27] consequential and where do they typically enter in that decision-making process? You know the [4:33] the bias is actually when when you look at when you've been inside a bank and you see the process, [4:38] it isn't like some underwriter or some banker saying, "Oh, this person is black or this [4:42] person is Hispanic, so I'm going to decline." It doesn't actually work that way. Uh what happens is [4:47] that certain um um minority communities, certain impacted communities have financial documentation [4:57] which requires more work, right? Doesn't mean that they're equally qualified, equally, [5:02] they love their home just as much as the next guy. They deserve it just as much as the next guy. But [5:07] the paperwork is more complicated. And human beings being, you know, sorry, in our nature, [5:12] we'd rather work on the easier deal, quote unquote, what's easier for us, the underwriter, [5:16] right? The underwriter than the the deal that would take more brain power. So So the tendency [5:21] is to say, okay, let me just make this problem go away by declining it. And and that's what happens [5:27] in in the industry. So um so it's more of a a circumstance thing. So when you apply AI to it, [5:34] uh the AI doesn't care. It's going to it's going to just run through this clock cycle. is going to [5:39] crunch the numbers and come up with an answer. It doesn't care if you give it one document or [5:42] 10,000 documents or how many cross links it has to create. It it just it just does it. One of [5:48] the things that you and I discussed before was uh the need for hiring the right people [5:53] who have an ethical basis and a and a bias towards integrity. Um human decision makers can introduce [6:00] inconsistency. They can introduce subjectivity, sometimes unconscious bias. So, which parts of the [6:07] mortgage process are most likely to become fairer when the judgment is replaced or supported by AI [6:15] in your expertise and opinion? Be back. Um, yeah, that's a great question and and if you [6:20] if you survey consumers and talk to especially if you survey impacted communities and say why [6:27] aren't you involved in the financial process? Why aren't you thinking about uh purchasing a home or [6:32] why aren't you thinking about, you know, moving away from payday loans into mainstream banking? [6:37] And the answer is almost always the same, which is which is I feel um poorly treated by my banker, [6:46] right? I I don't feel I don't I don't think I'm going to get a fair shot. Okay. Um I I feel I'm [6:52] going to get judged. So you can eliminate all of that. So that's where the the the the real [6:57] discrimination starts and and it's and in in some to some degree since it's in the mindset of of [7:03] you know past history and it creates a certain mindset. So you have to hit a hard reset and the [7:07] hard reset is um that these consumers if you give them a technology that they can talk to and and [7:15] um that's automatically the judgment is removed because there's no human they don't have to look [7:19] at the expression in the face of a banker and and worry about what is he thinking right and what [7:24] what what faces he making is as I'm explaining my my you know uh medical problems and collections [7:31] and things like that because of life circumstances and so much easier to talk to an AI about it. Um [7:37] there was a study done by a by a think tank that actually said that they found is that [7:43] uh across the board consumers trust in bankers is at 19% only like that's like lower than [7:49] Congress and consumers trust in AI is over 70%. Okay. So the the the solution is to have a fair [7:58] um and reliable machine that consumers can talk to and get accurate answers back. the the fact that [8:05] you built um an AI within an established related mortgage uh industry rather than beginning with a [8:16] standalone tech company. What did that experience teach you uh about the gap between the promise of [8:22] unbiased AI and the realities of Yeah, that's a great question. And I mean I can tell you that a [8:28] product like this could never have been built um in a standard startup venture funded process cuz [8:34] it's taken you know if you like I said the history goes back over 20 years uh and venture runs out of [8:39] patience in a couple of couple of years um and so it's a huge advantage by by building it inside a [8:45] financial institution because we get all the data and so many um large corporations today [8:51] are sitting their data is worth more than the company but they don't even realize it right and [8:55] and something we figured figured out long time ago and we focused on the data engineering and and the [9:02] data science and and the training. So when we're able to deliver a solution that meets their goals [9:10] of of fair lending and meets the goals of the bank, you know, a stepbystep manner. So every, you [9:18] know, over the years, you know, I I've put in $500 million of of my own money into building this. And [9:24] so we haven't had to take any outside funds and we're able to get returns immediately because we [9:29] build a section we immediately apply it within the within our bank and we get immediate returns on [9:36] it and and and you build a section by section by 2018 is complete. It's working end to end. [9:43] How do you determine whether an AI system is genuinely making lending decisions fairer rather [9:51] than simply producing more I guess consistent decisions based on flawed assumptions? How do [9:58] you how do you measure that? Yeah, that's a great question and it goes back to the um article you [10:03] recently wrote about model drift. Um and there's um a a ton of science around this. Um, of course, [10:11] the most important thing and as you explained in your article is about data integrity and quality. [10:16] You you you know, if you have AI feeding itself, AI generated data feeding itself, [10:21] right? Then you're going to have a big problem. So, one of the things is is we we don't feed our [10:25] model with any AI generated data. We have a huge number of data uh engineers on the project every [10:32] single day making sure that the data is pure. Um, that's that's number one. Number two is um [10:38] we have numerous patents on um um predictable and precise analytics uh and decisioning from an AI. [10:48] And the most recent paper we just published and we open sourced it um is on dealing an AI model [10:56] that can deal with real world data uncertainty, right? Because when you have real world data like [11:02] if a human being can look at one or two cats and one or two dogs and can pick out the next cat or [11:08] dog right away. Whereas a whereas an AI you have to look at millions of pictures of cats, millions [11:14] of pictures of dogs, right? And that's real world data uncertainty. So we're getting closer, [11:19] right? We're building intuition into the AI. So, we're getting closer to the way a human being can [11:25] decide between two uh two distinct sets or or make a decision between a variation in the real data. [11:32] uh to to expand a little bit on your point about uh model drift. There is a the article [11:38] that I wrote was in uh like August of 24 and it was based on a nature study that talked [11:44] about this concept of model collapse where AI is trained on its own outputs. And what happens [11:52] is that after five levels of that, it just turns back gibberish. And they were able to codify and [11:59] document that. And so within a closed system and financial lending and mortgage approvals, [12:04] that is a pretty closed system. So it's so that is a real threat and a real risk. Um in mortgage [12:10] lending, an automated decision can determine when someone is able to purchase a home. So what should [12:18] explainability look like for the applicant whose loan is denied? Um not just for the regular or [12:24] data scientist reviewing the system. I mean, what what what kind of feedback should they get other [12:28] than just a a no? Yeah. So, the CFPB, the uh Consumer Finance Protection Bureau and the FDIC [12:36] um issued a guidance on this about two years ago. Okay. And they what they said was a bank is [12:43] permitted to use AI to run everything in the bank and to make all the decisions, but the caveat is [12:48] that you can't treat the AI as a black box, right? So, so it's about as the AI makes a decision, [12:55] you have to have traceability, right? Um, so as it makes a decision, the steps it takes to make [13:00] the decision, you need to be able to it needs to be able to generate that log and and and not not [13:07] reprompt it back, right? That that doesn't count. You can't just reprompt and say, why did he make [13:11] that decision? But as it makes a decision, you need you need to be able to trace. You need to be [13:15] able to know why it made the decision. Okay? And that's one of the the core fundamental principles [13:20] of angel AI because as we're trusting you know $40 billion into it for forget the regulators [13:27] it's our money at stake we want to know why it made the decision. So it was as we designed it [13:31] u that was built into its core. So so any decision it makes uh we have the ability to generate the [13:39] detail step by step exactly how it came to that conclusion. So anybody if a consumer want to know [13:44] why did he decline me or why did he approve me we can we can get that information and and deliver to [13:49] anybody. [snorts] So, as part of the uh the the application process, we didn't talk about this, [13:56] so I'm going to it's kind might be a surprise to you, but um what role should alternative [14:00] data or unstructured data play in expanding access expanding access to credit for people who may be [14:08] poorly served by traditional credit scores, let's say, and where could alternative data create new [14:14] forms of discrimination or privacy risk? That's a doublefolded question. Sorry. Well, that's a great [14:21] question and actually um the industry recently u FHFA um which is the master regulator of housing [14:30] um housing finance um just approved something called a Vantage score. Okay. And and it's [14:39] different than the FICO score because the Vantage score takes this unstructured data [14:44] um and uses that into calculating a new type of credit score. Okay. Um so there's there's the the [14:52] challenge then is that lenders have to be able to accept the unstructured data right understand it [14:58] interpret it and then give it uh give it to the to TransUnion so that an accurate advantage score [15:05] can be generated and more unstructured normally a traditional bank more more unstructured data you [15:12] have more people you have and then more people who can get frustrated with the deal and just decline [15:17] it right Um so all the more reason the more unstructured data you have the more AI you need to [15:24] to automate the process. So everything that we do um we we do it so that it doesn't require a human [15:32] touch. Okay. So it'll it'll go through end to end unstructured structured data. Uh it'll it'll go [15:39] through it'll figure it out and and generate the right decisions. Does does unstructured [15:44] data or alternate forms of data introduce more risk of fraud and discrimination um or privacy [15:51] risk within the within that system? Um I I'll I'll tackle privacy risk first. I think any any banking [16:00] platform there's always that uh challenge right uh under the regulations we have to keep the data and [16:08] uh within our secure data center. Um, I question I question a bank that is sending their consumer [16:17] data to an outside um thirdparty chat engine uh AI AI engine and and getting answers back if if it [16:26] leaves your data center and you don't know which part of the data that they may have used to train [16:31] their models. Interesting. Right. And if so I I question those strategies, right? Um and I haven't [16:37] dug into those contracts specifically. We have we don't need to because we keep our entire model [16:43] um secure within our our our uh our secure servers our our secure data center. Um the there was other [16:51] two parts of your question. Oh, it it was it just had to deal with uh discrimination and fraud. [16:56] Yeah. uh fraud, unstructured data. Obviously, yes, greater chance of fraud because the the the [17:02] uh realtor, the the the borrower, even the loan officer, the the mortgage broker could, you know, [17:08] create something. It's unstructured and and you know, it could be completely fabricated. Um you [17:15] know, however, um the the probability of fraud is pretty low um these days because there's so many [17:22] government data sources that you can cross check against. Um and the and you know the industry's [17:28] gotten really good at detecting uh the the fraud. That's just in general. So not terribly concerned [17:33] about about fraud. Um and the last part was it was just discrimination. Discrimination. [17:39] I think I think having a system that can process unstructured data fast and immediately, right? Um [17:46] and and it actually reduces the discrimination. Okay. because the the the you know when you when [17:52] you when you hear these headlines and I think just a year ago a couple of major banks were were um [17:58] slapped for redlinining. Okay. Um and the reason explain that redlinining. Oh, sure. Redlinining is [18:04] that that that a bank will purposely purposely not lend to an underserved area. Interesting. Okay. [18:10] A low income or underserved area or minority area. Okay. And the reason they they they do [18:15] that is because um the nature of like there's a lot of gig workers or or uh small businesses or [18:22] you know um you know difficult to document income you know not that it can't be documented but just [18:29] difficult to document income. So they just don't want to do the work right and and and all of those [18:33] things are are unstructured documents unstructured data. And so the more unstructured it is, the more [18:40] automation you need because you you can't, you know, the other problem with unstructured data [18:45] and human beings reviewing it is you're asking um a a technician to be beyond CPA level expert, [18:58] right? So you're asking human beings to do stuff that they simply don't have the training or the [19:03] education um or the intellect to do. And it's just not fair. you're putting people in a very [19:08] difficult spot. Then you then turn around and blaming them saying you discriminated. That's [19:12] just not fair, right? Why why ask them to do something that they that is just so far beyond [19:17] even the best CPAs would have challenge would have a challenging time dealing with. Okay. So, [19:23] but but an AI which is trained with the knowledge of you know hundreds of thousands of CPAs and [19:29] experts um it can uh it can cut right through it all. So in our last minute here, uh looking [19:35] ahead five years, what evidence would convince you that AI has meaningfully reduced financial bias? [19:43] Uh great question. So uh there was a study done by a researcher at Lelay University um and he hit the [19:50] nail on the head. So he took um a thousand real uh real world loans, mortgage loans and he ran it [19:57] through um the major um uh chat bots uh you know whether Gemini or TAGPD or whatever. He ran it [20:04] through ASIS and he and his data showed the same thing that our data showed which is um that you [20:12] know out of every two black borrowers are declined one should have been approved. Okay. And then he [20:18] ran the same exact data through the same engines and he removed race as one of the inputs, right? [20:25] And he got he got equalized results. Interesting, right? So what that means is that all these [20:32] engines are trained on on bad historical data. So so there is a a problem across the industry, [20:39] right? Because you know it's you don't blame the engine, blame the the source of the data and [20:43] the source of the data is the industry. what the industry's actual practices have been. Understood. [20:47] Well, that's going to do it for our time here. I'd like to thank our expert uh insight from Pavan [20:52] Argawal as well as for your participation and your attention. Thank you so much. Thank you everybody.