Cash flow underwriting, explainable AI, and credit risk analytics are reshaping how lenders approve borrowers and price loans. Tedd Huff, CEO of fintech advisory firm Voalyre and founder of Fintech Confidential, sits down in Las Vegas with Jamie Twiss, CEO of Carrington Labs, and Kasey Kaplan, Carrington Labs' Chief Product and Commercial Officer, to break down how machine learning models trained on bank transaction data are outperforming traditional credit scores. Jamie spent years building credit risk models inside banks before serving as Chief Data Officer at a global top 50 institution. Kasey built his career across payments, program management, and lending. Together they launched Carrington Labs out of the mission-driven lender Beforepay to tackle the blind spots that credit bureaus have ignored for decades.

The core tension is straightforward. Tedd pointed to a LexisNexis report showing that over 50 percent of loan applicants cannot produce a reliable credit score, and he pressed on who gets hurt the most: self-employed workers, gig earners, younger borrowers who avoid credit cards, and anyone new to a country. These groups get filtered out by a system that only measures past engagement with credit products. Lenders lose business. Borrowers lose access. The economy loses capital deployment that could be productive.

Carrington Labs trains its models on billions of lines of bank transaction data, pulling behavioral signals that standard bureau scores miss entirely. The examples are specific: when a cash crunch approaches, does a borrower cut discretionary spending 21 days out, seven days out, or three? When an obligation hits, do they move money around to meet it, or let a direct debit bounce against an empty account? Do they shift from Safeway to Costco when balances drop? These behavioral patterns are, according to Jamie, enormously predictive of willingness and ability to repay.

The natural pushback from compliance teams and risk officers is that AI in lending means black box decisions that regulators will reject. Tedd framed the concern directly: 

Build secure, compliant crypto wallets without touching private keys.

Dfns - Wallets as a service provider offering API-first, multi-chain digital asset infrastructure with security, compliance, key orchestration, and blockchain integration for fintech platforms and custodians

Request your demo now at fintechconfidential.com/dfns

Carrington addresses this head-on by splitting their workflow at what they call a "control point." Upstream of that control point, AI and generative models do heavy lifting during feature generation and model creation. A human reviews every feature and signs off before anything goes into production. Downstream of the control point, everything is deterministic: same input, same output, fully explainable. Kasey shared a telling example; some lenders they've encountered were literally calling the OpenAI API with borrower data and asking ChatGPT whether to approve a loan, then getting different answers at different times. Carrington's architecture avoids that by keeping final decisions within reproducible, auditable machine learning models.

For lenders dealing with high decline rates and rising acquisition costs, the commercial case goes beyond approval percentages. Carrington builds personalized models per lender, per product, and per customer segment, using that lender's own data. A wage advance carries a different risk profile than SMB lending or auto leasing, and the same borrower can look very different across those products. Off-the-shelf models are available for lenders with no existing data, and the system shifts to custom models as performance signals come in. Kasey compared the approach to e-commerce personalization: instead of generalized risk bands, lenders get probability of default percentages for each individual borrower, which maps directly to unit economics and lifetime value.

Tedd also pushed the conversation toward a shift most lenders have not made yet: moving from point-in-time underwriting at origination to ongoing, recursive underwriting throughout the life of a loan. He raised a scenario where rising fuel costs, with gas in Las Vegas ranging from $5.29 to $6.29 a gallon at the time of recording, start compressing discretionary income in ways that a one-time credit check would never catch. Carrington's cashflow servicing capability can detect that kind of early distress, flag when spending patterns change, and help lenders adjust credit limits on products like credit cards. Jamie described a scenario where a lender with 100,000 credit card holders could identify 5,000 showing emerging stress and 30,000 performing well enough to support meaningful limit increases.

Skyflow - Zero trust data privacy vault delivered as an API; collect, secure, and tokenize personal information like card data and payment details with built-in PCI, CCPA, GDPR, and SOC 2 compliance - skyflowsecure.com

Advertisment

Open banking access remains a pressure point, especially in the US where institutions like JP Morgan have moved to charge for data access. Jamie argued that restricting customer data sharing signals a lack of confidence and pushes consumers toward less secure methods like emailing bank statements. Tedd connected that to a broader pattern he has seen across fintech, noting that email has become a workaround when API-based data access gets blocked, essentially a regression to less secure channels. Kasey added nuance about compliance costs for smaller institutions and differences in regulatory frameworks between the US and Australia, where restrictions on derived data are tighter.

Looking ahead five years, Jamie predicted credit will become embedded in financial life, with AI agents seeking and deploying capital on behalf of borrowers while lenders use far richer data sets for more precise decisions. Kasey expects regulation to catch up, friction to drop across applications and access to capital, and consolidation among smaller institutions that cannot fund the shift. Both signaled that non-financial companies closest to customers, like retailers, may eventually step into lending if the decisioning, origination, and servicing tools become accessible enough.

Tedd closed the conversation by distilling three takeaways that cut across the full discussion: lending is not a simple yes or no decision, siloed data can still produce strong outcomes without consortium sharing, and the industry is still in the early innings of what this technology can do. For treasury managers, CFOs, and compliance leaders evaluating AI tools for lending, this conversation lays out the operational line between generative AI and explainable machine learning, the commercial math behind personalized scoring, and the practical steps to move from bureau-dependent underwriting to cashflow-based decisioning. The specifics on behavioral features, model retraining, and lifecycle monitoring make the case that this is an operational shift already producing measurable results for lenders globally.

Credit scores have barely changed since the late 1980s, and over half of loan applicants still cannot get a reliable one. Tedd Huff, CEO of fintech advisory firm Voalyre and founder of Fintech Confidential, sits down with Jamie Twiss, CEO of Carrington Labs, and Kasey Kaplan, Chief Product and Commercial Officer, to unpack how cash flow underwriting using bank transaction data is giving lenders a sharper, fairer way to assess borrowers. They break down why behavioral signals buried in everyday spending habits predict repayment better than bureau scores, how to use AI upstream while keeping lending decisions fully explainable, and where the real margin gains hide across origination, limit sizing, and post-origination servicing. Jamie and Kasey also share where they see AI-powered lending heading over the next five years, including the rise of embedded credit and agent-driven borrowing. If you make, manage, or evaluate lending decisions, this one is worth your time.

Hawk AI - Real-time payment screening, ML transaction monitoring, and dynamic customer risk rating tools designed to fight fraud and financial crime while reducing false positives - gethawkai.com

Advertisement

Younger borrowers watched their parents drown in debt, and now they avoid credit products even when they can afford them. Traditional scoring systems penalize that restraint, leaving responsible earners with weak credit files and fewer lending options.

Carrington Labs proves confidence in its models by running a large portion of its own capital through the same decisioning system clients use. That alignment, lending significantly more than the CEO's personal net worth every single day, gives partner lenders a level of assurance most vendors cannot match.

Within five years, AI agents are expected to seek, compare, and deploy credit on behalf of consumers without a single traditional application form. The shift from manual applications to embedded, agent-driven borrowing could make today's origination process feel as outdated as a fax machine.

Some lenders brag about having thousands of features in their credit models, but many are just the same metric repeated across different time windows. Grocery spend at seven days, 14 days, 30 days, and 60 days adds volume without adding predictive quality, and that brute force approach loses to a handful of well-designed behavioral features.

A quality issue in AI-generated marketing copy is forgivable; a quality issue in a lending decision is not. Financial leaders need to classify every AI use case by its tolerance for error before selecting the architecture behind it.

Bank transaction data can reveal whether a borrower received money from a cousin, whether that transfer was a loan or a gift, and whether it was repaid before or after the next paycheck arrived. That granularity of behavioral insight sits in the data most lenders already have access to but never examine.

Non-financial businesses sitting closest to the customer and the transaction data stream are evaluating whether they should originate credit themselves. If decisioning, limit setting, and servicing tools become accessible to smaller operators, lending could spread well beyond banks and fintechs.

AI-powered loan origination is projected to hit $20 billion within the next decade, and most of that growth will come from lenders replacing point-in-time bureau pulls with ongoing, data-rich decisioning. The lenders who wait for the market to mature risk losing both margin and market share to those already building the infrastructure.

Carrington Labs built an MCP server that allows credit teams and support staff to query model results for a specific borrower without reading through hundreds of lines of schema. That same server supports agentic workflows, giving AI-powered systems access to contextualized risk scores and limit rationale in real time.

Most lenders treat the approval decision and the terms of the loan, limit, duration, and pricing, as separate processes handled by separate teams. Connecting those two workflows into a single optimized decision is, according to the advice given, the single highest-impact change a lending operation can make right now.

Hawk AI - Real-time payment screening, ML transaction monitoring, and dynamic customer risk rating tools designed to fight fraud and financial crime while reducing false positives - gethawkai.com

Advertisement

Lenders pour money into acquisition and then have no idea where borrowers drop out or why. Map analytics to every single step from account creation to loan acceptance, pair quantitative data with qualitative research at each stage, and you will find the exact friction points bleeding your conversion rate. Pricing too high? Bureau pull anxiety? Open banking consent hesitation? You cannot fix what you are not measuring.

If your core strength is the customer relationship, spending two years and a seven-figure budget building proprietary underwriting tech is the wrong move. Best-of-breed origination, decisioning, and servicing tools already exist and can be assembled faster than any internal team can ship. Redirect that engineering budget toward understanding customer pain points and differentiating the product, because that is where lenders actually win or lose.

You do not need a massive data set on day one to deploy machine learning models for credit risk. Put a small number of loans out using off-the-shelf models, collect performance signal, and rapidly shift to custom models trained on your own portfolio data. The conservative ramp-up approach gets you to precision faster than dumping capital out the door and hoping the data catches up.

Borrowers who get approved for a smaller amount using traditional methods often want more but do not know how to get it. Give them a clear path: sync additional bank accounts through open banking, allow cashflow underwriting to assess their full serviceability, and reward that transparency with higher limits. You increase average balances, the borrower gets better terms, and your model gets richer data to sharpen future decisions.

A credit score tells you what someone did with past loans. Transaction data tells you what they do when money gets tight. Track whether borrowers trade down to cheaper providers when their balance runs low, whether they adjust spending weeks before a crunch or days before, and whether they prioritize obligations over discretionary purchases. Build those behavioral responses into your underwriting criteria and you will catch risk signals that bureau scores completely miss.

Jamie’s LinkedIn: https://www.linkedin.com/in/james-twiss/ 

Kasey’s LinkedIn: https://www.linkedin.com/in/kaseykaplan/

Website: https://www.carringtonlabs.com/ 

Company LinkedIn: https://www.linkedin.com/company/carringtonlabs/ 

YouTube: https://fintechconfidential.com/watch

Podcast: https://fintechconfidential.com/listen

Notifications: https://fintechconfidential.com/access

LinkedIn: https://www.linkedin.com/company/fintechconfidential

X: https://X.com/FTconfidential

Instagram: https://www.instagram.com/fintechconfidential

Facebook: https://www.facebook.com/fintechconfidential

Transform Your Merchant Applications with Under. The Under platform revolutionizes how you handle merchant applications, offering a seamless transition to digital forms. Say goodbye to outdated processes and hello to efficiency. Discover the future of financial applications at https://under.io/ftc 

Advertisement

Jamie Twiss is the CEO of Carrington Labs and Beforepay Group. He is a data scientist and banker who has spent his career at the intersection of data, AI, and financial services. Jamie began his career as a consultant at McKinsey & Company, then moved through a series of banking and financial services roles before becoming Chief Data Officer at a major Australian bank. His focus has always been on the hardest problems in credit risk, turning unstructured data into high-stakes lending decisions. He founded Carrington Labs to bring explainable AI and cash flow underwriting to banks and non-bank lenders globally, building the company's credit risk models on billions of lines of transaction data from its sister lending business, Beforepay. Jamie and his team prove confidence in their models by running significant capital through the same decisioning systems their clients use every day.

Kasey Kaplan is the Chief Product and Commercial Officer of Carrington Labs and was recently appointed Deputy CEO of Beforepay Group. He is a product, strategy, and growth executive with over 15 years of experience across payments, program management, and fintech lending. Kasey's entry into fintech started with a personal problem, trying to link credit card transactions to automatic check-ins during grad school, which led him into the mechanics of payments and cards. He went on to found a prepaid debit card program, served as Chief Operating Officer at Urban FT, and held product leadership roles before joining Beforepay and Carrington Labs. At Carrington Labs, Kasey leads commercial execution across credit risk scoring, cash flow underwriting, and loan limit solutions. He is responsible for driving growth across Beforepay's Pay Advance, Personal Loans, and Carrington Labs business lines.

Carrington Labs is the AI and enterprise software division of ASX-listed Beforepay Group (ASX: B4P), headquartered in Sydney, Australia. The company builds explainable AI-powered credit risk scoring, loan limit recommendations, and cash flow underwriting solutions for banks and non-bank lenders worldwide. Carrington Labs grew out of the lending operations of Beforepay, which has issued more than 4 million loans in Australia using the same technology. The platform uses bank transaction data and advanced behavioral features to assess creditworthiness beyond traditional bureau scores, with models custom-trained on each lender's own data. Products include credit risk models, cash flow scoring, limit sizing, line management, cashflow servicing for early distress detection, and an MCP server for agentic and support team workflows. Carrington Labs supports lenders globally, with clients spanning wage advances, consumer lending, SMB lending, and auto leasing, including a partnership with Flexcar announced in late 2025.

Tedd Huff is CEO of Voalyre, a fintech advisory firm, and founder of Fintech Confidential. Over the past 25+ years, he has contributed to fintech startups as an Advisory Board Member, Co-Founder, and Chief Experience Officer, providing strategic and tactical direction for global companies. His expertise focuses on growth while delivering process improvements and user experience-driven value to simplify the complexity of payments. As host and executive producer of Fintech Confidential, Tedd brings entertaining and informative content focused on fintech industry insights, market trends, and stories from fintech leaders, thinkers, and doers. He is a recognized thought leader and U.S. Army veteran known for making complex financial technology approachable and engaging through his conversational storytelling style and deep understanding of global payments, cross-border transactions, and payment localization.

Fintech Confidential is produced by DD3 Media and hosted by Tedd Huff, CEO of fintech advisory firm Voalyre. Established in 2022, the show brings you the people, tech, and companies that change how you pay and get paid. Each episode features in-depth conversations with the operators, founders, and executives shaping payments, banking, and financial services. Subscribe on YouTube, Spotify, Apple Podcasts, or wherever you listen. For behind-the-scenes coverage and exclusive content, visit fintechconfidential.com.

The Newsletter Platform Built for Growth

When we started the newsletter, there were SO many choices. But until now, there hasn’t been a publishing tool built to help us grow our publications as quickly and sustainably as possible!

beehiiv was founded by some of the earliest employees of the Morning Brew, and they know what it takes to grow a newsletter from zero to millions.

It is an all-in-one publishing suite that comes with built-in growth tools, customization, and best-in-class analytics that actually move the needle - all in an easy-to-use interface.

We are excited to engage with you through — responsive audience polls (find out what you want to hear about most), custom referral programs( get rewarded for referring people to the Fintech Confidential newsletter), SEO-optimized webpages (make it easy to find the content you are looking for), and so much more.

If you have or are considering to starting a newsletter, there’s no better place to get started and no better time than now.

Try beehiiv absolutely free with no credit card required.