Generative AI, Transformers, and the Future of Underwriting

Welcome to Fintech Brainfood, the weekly deep dive into Fintech news, events, and analysis. You can subscribe by hitting the button below, and you can get in touch by hitting reply to the email (or subscribing then replying)

Subscribe

Weekly Rant ??

Generative AI, Transformers, and the Future of Underwriting

The competitive advantage in credit underwriting is getting ever closer to the truth. Transformers and Generative AI will unlock a new era of productivity in Underwriting by gathering data and using transformers to identify profitable borrowers.

This will result in an inevitable regulatory pushback because lending is a spider's web of regulation, and AI and regulation are already in the spotlight.

That shouldn't stop us from innovating.

But first:

1. The constant push pull of mo' data, mo' regulation
2. A brief history lesson on the quest for perfect data
3. Obvious new AI use cases are data gathering and analysis, but the kicker is using transformers for underwriting
4. There will be negative consequences and regulatory push back
5. The solution is better data governance and engineering as we learned from machine learning
6. In the end its always push pull but transformers will be transformative

If it's Fintech, regulators are watching. If it's AI, regulators are watching. Lending is the high-risk part of finance. These are three highly combustable elements.

Here's how it will play out.

1. The history of credit history mo' data, mo' problems

Better data analysis usually requires more data. The problem with collecting more data is it creates pushback. Is the data collected too much? Is it used fairly?

The simplified version of the history goes something like this.

1. Credit history and the score. FICO and credit bureau took consumer income and repayment history from the market and returned a "score." The higher the score, the more likely someone was to repay.
2. Data analysis and the algorithms. Capital One's "Information Based Systems" in the early 90s were founded on the insight that customers close to a credit score line (e.g., 10 points below 670) might be less risky than those 100 points away.
3. Alternative and cashflow data and the algorithm. Features, like Earned Wage Access (EWA), emerged in the late 2010s when payroll and account cashflow data became more readily available. Lenders now have a much clearer picture of a consumer's current position regardless of their credit history. (We've barely scratched the surface of using this data today, and it has a long way to run)

Over time, we've gotten progressively closer to the full picture. But every time we do, we get pushback. Personal data is private, sensitive, and can be used to discriminate.

For example

• In its oversight role, the CFPB looks to test that any information-based system (algorithm) is not demonstrating bias.
• The CFPB is looking to bring BNPL or Earned Wage Access under its remit and rule-making.

Every time there's a lending innovation, there's a regulatory pushback.

This cycle repeats.

2. The quest for perfect data isn't easy

As Alex Johnson has discussed , there is an information asymmetry between what a consumer has ever done and what the lender can find out.

In a perfect world, a lender would have perfect data.

This could be a nightmare for consumer privacy and fairness.

Because perfect data doesn't obviate a perfect lending decision.

It can often create bad outcomes.

Lenders innovate to get closer to the truth, creating unintended consequences.

• Credit reports once had inaccurate data, and consumers had little recourse - helping the calls for what became the Fair Credit Reporting Act (FCRA)
• Lending decisions weren't always fair and could discriminate based on age, gender, or ethnicity. Today, the Equal Credit Opportunity Act (ECOA) is managed by the CFPB, which has taken a clear stance on ensuring that all lending decisions can be proven (explained) to be fair.

There's a law covering everything from credit cards (CARD) to Mortgage Disclosures (HMDA) and the Gramm-Leach-Bliley Act (GLBA) requiring an explanation of data collection.

Information asymmetry also exists for consumers. Do they know the product they're buying and how it works? (Spoiler alert: no, they don't.) The Truth in Lending Act (TILA) requires the total amount repayable, repayments, fees, and APR % to be clear to create some sense of fairness.

Information asymmetry sucks.

Regulation tries to fix this.

But it becomes incredibly complex.

This constant push-pull over data is coming to a head again.

The opportunity surrounding AI will drive advances in credit scoring and data gathering.

3. The newer AI types present new opportunities

Large Language Models (LLMs) excel at data gathering and entry. Casca* acts as an agent in the loan origination process, gathering customer KYC information, handling emails, and entering that information into existing back-end systems.

Coris and Baselayer help screen for merchant risk by pulling in website data, NAICS codes, and KYB information. (Today, many more complete platforms like Sardine* do the same. )

The obvious use cases involve:

• Taking complex, unstructured data.
• Summarizing it.
• Reformatting it.
• Making it fit some esoteric legacy system or form.

Generative AI can also handle customer communication. Klarna announced that its customer support AI can now handle the same volume of customer queries as 700 human agents. They also said customer satisfaction was higher; the average query completion time went from 11 minutes to 2. Oh, and it speaks 35+ languages.

I'll take it if Generative AI is a better solution to waiting in line for 30 minutes for a call center. Chatbots have been awful for a long time, but the best ones are starting to get genuinely good. I've been the biggest cynic of chatbots for over a decade. But when applied to a limited set of use cases and well packaged, they're transformative for the customer experience.

Generative AI is helping make crucial decisions. Axle has AI agents that help review sanctions alerts, generate suspicious activity reports (SARs), transaction alerts, and onboarding documentation. These alerts are generated whenever the existing automation flags something as strange for a human to look at.

Generative AI gets the first look and can perform much of an investigation's heavy lifting and manual work. If that sounds cutting edge, consider that these features are becoming standard, too. Whether built by in-house teams or your risk platform, it's incredible how quickly this becomes standard.

Transformers can also be used for data analysis. The Nubank acquisition of Hyperplane is fascinating. They train foundation models on large internal data sets for use cases like Underwriting, collections, and marketing.

Why does this matter? Cast your mind back to Alpha Go. In this breakthrough moment, a learning model (a residual convolutional neural network or CNN) performed a move in the game that no human could have imagined. But it worked. Humans found the movie beautiful or even breathtaking. The model saw something new that it was never trained on.

Today, it is generally accepted that transformers (the type of model powering Generative AI models like Llama, Claude, and ChatGPT) are significantly more efficient than CNN's (a machine learning model class used by some underwriters).

Therefore, applying transformers to problems like Underwriting could help you identify profitable pockets of risk that other lenders did not see.

Imagine you have a giant data set like Nubank's 100 million customers and a data infrastructure capable of running models against it. Any bank, Fintech company, or BNPL provider with millions of customers could use transformers to identify profitable pockets of risk.

To me, this is the rub: We confuse the conversation about AGI, AI risks, and large language models with transformers themselves. That means there’s a massive opportunity to use them for more efficient underwriting (and other use cases).

This will be the next wave of "Generative AI" - Generative Underwriting. (Which, to be technically accurate, is a transformer-based foundational model)

PS. It won't work for most legacy incumbents because they likely can't gather the data, cleanse it, and format it well enough to run this type of model successfully. Gotta get the basics right first. Better use of AI requires better data engineering. Focus on that first.

4. There will be negative consequences

But your best bet is to innovate and take the fine.

Don't get me wrong. I'd never advocate for intentionally harming consumers. The opposite: we should bend over backward to bake in protections from day 0.

However, if you're going to innovate in financial services, your work doesn't neatly fit into an existing regulatory category.

This is heightened with AI, which is topical. Whether its hyperbole about species risk, AI summits, executive orders, or anything else, the need to be seen to be regulating AI is pretty strong, regardless of the material risk.

There will be material risks.

We already see that fraudsters are using Generative AI to scale the quality of their fake IDs, phishing, and even deep fakes that pass liveness checks.

There will also be unintended consequences of using AI for customer communications, sales, collections etc.

• Will Generative AI consistently gather the correct data or will it overstep a privacy boundary?
• Will a Generative AI agent ensure adequate disclosures subject to TILA?
• How do we know Generative AI decision support agents will make the right calls?
• How do we know humans reviewing "AI Agent" decisions will not become over reliant?
• Will a transformer model be consistently fair under ECOA?

These are valid questions. Regulators have started and will continue to ask them.

All of these questions exist for previous generations of machine learning and any other form of data-driven analysis.

However, in recent years, we've also seen regulators' default posture be to enforce now and learn later. Therefore, we can expect a huge push back.

But that should never stop us from innovating.

Most of the time, companies like Revolut, Robinhood, and Coinbase, who've seen incredible pressure from regulators, end up just fine. If anything, they thrive.

There's no way you can pre-calculate every possible regulatory pushback if you're doing something novel. Instead, work from first principles, understand the intent of the existing rules, and design sensible controls.

5. Better AI needs better data governance

Data governance becomes critical.

You can't avoid potential regulatory pushback; it's inevitable. However, you can apply data engineering to build a sustainable business.

Show you're workings out. Show how you got to the answers. Show your process. This is obvious to data pros in Fintech, but if you're getting into embedded finance or haven't worked in Fintech before, this stuff will be a lifesaver.

• Track every bit of data: From its source through transformations to use. If you know what data was used, you can discuss how it was used.
• Make models explainable: Ensure your AI models can be explained in human terms, especially for credit decisions. Regulators will be especially mindful of ECOA here. But is the model working as you've advertised (and disclosed subject to TILA)?
• Test for Bias: Regularly test your models for unfair bias. Beyond protected classes – look for unexpected correlations that could lead to unfair outcomes. Again, this is crucial for ECOA.
• Monitor for Model Drift: Set up systems to monitor model performance and data quality in real-time. Drift happens, and you need to catch it fast.
• ?? Humans in the Loop: Establish a group to review new AI applications and major changes. Include legal, compliance, and even consumer advocates. (I hate myself for saying this, but
• If it moves, document it: If you document it, you can explain it during an examiniation. (Generative AI can help you here)
• Compliance is everyone's job: Ensure everyone, from data scientists to customer service reps, understands the importance of TILA and ECOA's goals.
• Don't wait for the exam: Share your governance practices, seek feedback, and demonstrate your commitment to responsible AI use. (Pro tip: This is the kind of thing regulators love and will save you a ton of pain later)

If this sounds like a lot for a 10-person startup, remember that generative AI helps document and audit Generative AI. There's an AI tool for just about everything in AI these days.

All of this work helps you come correct to the regulator when they do push back.

An example of how not to do it: I was talking to a former senior at Neobank yesterday who said,

6. Ultimately, it's always push-pull, but the most efficient algo wins.

The story of lending is a constant dance between innovation and regulation. Each breakthrough in Underwriting is a mixed blessing:

• We get closer to fair, accessible lending.
• However, we also risk overstepping privacy boundaries or inadvertently creating new forms of discrimination.

This push and pull isn't new, but AI is accelerating the cycle and raising the stakes. Transformers represent a step change in model efficiency, not just for LLMs but for other use cases too.

Transformers and Generative AI are already being applied to revolutionize lending (by companies like Nubank). There's plenty to like here:

• Streamline data gathering and analysis
• Improve customer experiences
• Identify creditworthy borrowers that traditional models miss

With great power comes great responsibility (and regulatory scrutiny).

Here's the thing: We can't let fear of regulation stop innovation.

Lending, at its best, is a lifeline. It creates opportunities, builds businesses, and helps people weather financial storms. Despite their challenges, innovations like Earned Wage Access and BNPL have expanded financial access for many. They're a massive net positive for the economy, consumers, and society.

The key is to innovate responsibly:

1. It's never too late to get data governance right
2. The earlier you bake in fairness and transparency the faster you'll unlock sustainable lending.
3. Regulators love a chat. Please show them your homework.

Yes, there will be pushback.

Yes, there will be fines.

However, the companies that get lending data governance rights will emerge stronger, with better products and more trust from consumers and regulators alike.

We can't copy + paste transformers and Gen AI into lending and call it Generative Underwriting (as catchy and tempting as that may seem)

We can use AI to create a more efficient and equitable financial system. AI can identify pockets of creditworthy consumers that other models miss and give them a chance.

That's not just good business. It's just plain good.

And that? That's genuinely exciting.

ST.

4 Fintech Companies ??

1. Aviva Credito - The Kiosk for Loans (MX)

Aviva powers credit cards or loans up to $20,000, which users can start online and complete at an in-person kiosk. Users visit a kiosk, validate their IME (SSN equivalent) number, and receive their cash within 24 hours. The service is aimed at consumers "invisible to banks and Fintech companies."

?? Kiosks are the new branches? Digital channels like mobile have a high risk of fraud, and it's hard to tell if a user really is who they say they are. The historic control for this was visiting a branch. Weirdly enough, it's so effective as a fraud control because it takes a lot of audacity to commit fraud in person. The kiosk is an interesting proxy for this. 10 years ago, I'd have said kiosks were crazy. But now it's every fast food place and soon, every branch? It's a great idea. Let's see if they can make the credit model work.

2. Hyperbots - AI agents for CFOs

Hyperbots builds agents that perform accounts payable, expense processing, and accounts receivable. The bots automate manual tasks like assigning a purchase order number and entering data into an ERP. They can handle exceptions and process refunds if they see evidence of a delivery failure.

?? These all feel like Zapier automations for legacy businesses. If your business doesn't already have a platform like Ramp, Airbase, Brex, etc., your finance team still lives in a manual world. Most of these capabilities will eventually be added to the workflow owners in the modern CFO stack. However, building bots for legacy tools is an opportunity like robotic process automation (RPA) x 100. It might not be very interesting compared to AGI, but there's much opportunity.

3. Mintos - Alternative Investments for Europe

Mintos is a digital wealth platform that offers high-yield cash, fractional bonds, and loans as investment opportunities. It is available for consumers and businesses in Europe, Israel, Canada, Japan and several key LATAM markets like Chile, Brazil and Argentina. The service is MIFID registered (so it likely provides access to EU-based assets).

?? I would legitimately use this if it were available in the UK. It's clean and well-packaged, and the mix of business / personal accounts is ideal for those of us who operate companies in the global workforce. Services like Wise often offer stocks, but this wider array of investment opportunities feels like Treasury for the solopreneur.

4. Reiterate - Finance ops workflow SaaS

Reiterate helps finance teams automate their workflows (like accounts payable or receivable) and reconciliations and streamline month-end. It targets more complex contract and pricing structures and aims to ensure accuracy.

?? This is a crowded category, but I love the depth and detail here. One example is calculating the fees a payment processor is charging vs. what they should charge, given their performance, which can be incredibly challenging. There are complex things in this world, and then there are the volume discount payment volumes. Earlier in my career I remember the most complex spreadsheets in the company were always the pricing ones.

Things to know ??

1. Affirm stock jumps 16% on massive earnings beat .

Affirm's revenue is up 48% year over year, and Gross Merchandise Volume (GMV) is up 31% in the period. They now have 300,000 merchants and 18.6 million active customers. Their goal is to hit GAAP profitability by mid-2025, and with rates falling, the cost of funds for Affirm should fall, too, making them more profitable.

?? Affirm's partnerships are paying off. Affirm was the first BNPL provider for Apple and has done well with Amazon and Shopify. Add to this a core business that continues to add merchants, consumers, and repeat customers, and this growth story is far from over.

?? Affirm's product expansion is also going well. They now have 1.2 million active Affirm card users. They're launching their Affirm Money account and BNPL for B2B, and while these aren't big yet, they represent new markets and revenue potential.

?? They plan to launch in the UK next year, one of Klarna's most profitable markets. Affirm is on strong footing in its home market, but Klarna is nipping at its heels. There's something poetic about Affirm going back the other way. Market expansion is hard, as Klarna and Adyen showed, but it can also pay off in time (again, as both showed).

?? BNPL, as a consumer wedge, is slowly becoming a two-horse race. You can't underestimate PayPal, Afterpay, Zip, and Sezzle, who are all in the conversation, but I wouldn't define any of these using BNPL as the way they drive product expansion.

2. Klarna revenue up 27% and adj profit up to $66m.

Klarna's revenue and profitability jumped, driven by strong growth in its US business, where revenue is up 38% YoY. They added 68,000 new merchants and millions of consumers, and the company says it is "partner of choice" at 1 in 4 merchants, 5 years after market entry. This all comes as the company plans to go public next year for a reported $20bn IPO.

?? Klarna is becoming top of mind BNPL player. Affirm and Afterpay had a head start, but Klarna is out-executing them. Afterpay has gone quiet after its acquisition.

?? Affirm is still bigger and growing at a mighty clip, with revenue up 48% YoY and adjusted income at $600m (up 23%). But something about Klarna’s PR and marketing makes them feel bigger. I think this is important because we’re heading to a world where the checkout button wars will be won or lost on consumer mind-share.

?? My favorite stat: The average revenue per employee over the trailing twelve months increased by 73%. Klarna is making this an AI-efficiency story, but it is equally cost discipline and revenue growth. That takes discipline and execution.

?? Klarna has a way of dominating mindshare. Every month or so there's some press release about how they're using AI, considering an IPO, and even weird board drama makes news . Brands that generate clicks, also generate investor interest. Just last week, they announced a sort of savings deposit, a sort of cashback product. It's working.

?? Klarna has a "take rate" of 2.54% (up 21bps). A lot of this comes from the value add they offer merchants in conversion. But they're also charging consumers for a premium subscription product, where the shopping app offers discounts and cashback

?? Consumers love this thing. Their NPS is 75, and they have 31 million monthly active users. They have huge brand permission to cross-sell to that user base. I expect a lot more product expansion beyond savings and cashback, and I'm curious to see if their credit card can gain market share from traditional players like Capital One and Chase.

?? Can they compete with Shop Pay, Apple Pay and Fastlane?

The future of payments is wallets at checkout and on devices. They'll compete on convenience, cashback, and bake-in BNPL. Affirm has decided to partner with Apple Pay, but Klarna has a deep merchant relationship and app usage that could compete. They'll still default to partnerships over competition. Accessing the 100s of millions of Apple Pay and Shop Pay users is a business case that is too good to ignore.

Good Reads ??

1. Unravelling who can use CashApp Pay

CashApp Pay is a checkout button, a little like Shop or Apple Pay, but its unclear which merchants can and do use it, and worse, what Block’s economics are in the process. Samir breaks it down as follows.

1. Cash App users buy from Square Sellers via Cash App Pay. This results in an incremental yield of ~15bps (what they’d typically pay the payment network
2. Cash App users buy from Stripe/Adyen merchants via Cash App Pay. No incremental revenue, but likely higher conversion.
3. Cash App user buys from select third-party merchants via Cash App Pay. Samir again estimates no incremental interchange revenue but likely higher conversion.

That's all, folks. ??

Remember, if you're enjoying this content, please do tell all your fintech friends to check it out and hit the subscribe button :)

(1) All content and views expressed here are the authors' personal opinions and do not reflect the views of any of their employers or employees.

(2) All companies or assets mentioned by the author in which the author has a personal and/or financial interest are denoted with a *. None of the above constitutes investment advice, and you should seek independent advice before making any investment decisions.

(3) Any companies mentioned are top of mind and used for illustrative purposes only.

(4) A team of researchers has not rigorously fact-checked this. Please don't take it as gospel—strong opinions weakly held

(5) Citations may be missing, and I've done my best to cite, but I will always aim to update and correct the live version where possible. If I cited you and got the referencing wrong, please reach out