- Fintech Brainfood
- Posts
- Scaling Human's with Financial AI
Scaling Human's with Financial AI
Plus; SumUp raises $300m, Google does BNPL and Apollo's deconstructed Bank
Hey everyone 👋, welcome to Brainfood, the weekly read to go deeper into Fintech news, events, and analysis. Join the 36,489 others by clicking below, and to the regular readers, thank you. 🙏
Hey Fintech Nerds 👋
While it appears ChatGPT is already winding down and appears to be taking a winter break because it’s nearly the holidays, Fintech is doing the opposite.
We are so back! (kinda).
SumUp (European Square) raised over $300m and is targeting acquisitions for growth. Google is adding BNPL, and now the UK may cap interchange again on cross-border fees.
I loved the Good Read this week by Marc Ruby. I can’t unsee how screwed banks are by capital requirements. Apollo is now originating $40bn of loans to Fintech (and other sectors) and expects this to be $75bn in two years. Where it can’t, it’s helping to syndicate and provide warehouse facilities to others.
Unless regulation changes, the future of banking doesn’t need banks.
Last Guest Rant before I'm back in the saddle, ranting again in the New Year.
Also, I might drop a prediction special if a certain toddler and newborn cooperate over the break 😇
Here's this week's Brainfood in summary
📣 Rant: Guest Rant: Generative AI in Financial Services
💸 4 Fintech Companies:
👀 Things to Know:
📚 Good Read: How Apollo is creating a deconstructed bank
Weekly Rant 📣
Guest Rant: Generative AI in Financial Services
By Ntropy CEO and Founder Naré Vardanyan
It has been a year since the release of ChatGPT. Suddenly, the pace of technology and the trajectory of what is possible has completely changed.
If you were in AI before 2022, you know how different things are today and it is very hard to hide your excitement. Although the recent Gemini demo seems to be staged, most of this magical multimodal functionality is real.
We finally have computers that can reason and generalize across tasks and domains.
You also have to keep in mind that the state we're in will change rapidly, and we need to prepare for a world with 10x more compute.
My day job is running a team of data scientists and AI engineers working with financial data and building intelligence for banks and FI-s. Hence I get to think about the impact of this technology on banking a lot more than your average person.
When it comes to finance, the opportunities are massive, but the requirements are tougher.
The requirements in financial services
Nicely staged demos are obviously not going to get this technology into production. We all operate in a regulated industry with high stakes and a high cost of mistakes. Precision, high recall, and scalability are must-haves in every finance function. Let us take lending as an example.
We need to be precise because if we get a wrong outcome or an inaccurate answer, we can underwrite loans that will result in financial losses, regulatory scrutiny, and possible job loss or prison for business owners.
We need high recall because we do not want to miss out on good applicants and narrow down the revenue opportunity.
We need high scalability because the business of selling money is highly profitable, and the only link that does not scale are operations, i.e., humans needed to deliver financial services,
Traditionally we have had to make tradeoffs between precision and recall, resulting in loose policies and badly performing financial products or serving very restricted and niche audiences with high barriers to entry. As an example, lenders would be less precise (e.g. not accepting FICO scores below 700) to avoid the risk of lending to a segment of customers that were higher risk.
With larger models that can generalize across tasks and data types these tradeoffs are no longer necessary. You are able to have a system that is highly precise but equally good at understanding out-of-distribution data, hence much more inclusive.
We can find the needles in the haystack.
(e.g., Finding good credit below 700 FICO).
We could always do this, but it requires too many humans.
So let's talk scalability.
Human's don't scale well and are expensive.
When we think about banks and profitability, one of the most important metrics too look at is their efficiency ratio.
Humans are the least efficient in the value chain of financial services.
Your marginal costs are a function of how many customers you can successfully serve per newly added employee. The lower this number, the less efficient a bank is and the harder it is to deliver great customer service or economics to end clients.
We saw the first wave of efficiency gains with digitization and mobile banking.
Suddenly, your employee in London could serve people in Somerset, Wales or Newcastle without having to operate branches and be within walking distance of your customers.
With the recent advancements of AI, the operational leverage for banks will be transformed once again..
Today, Nubank is 20 x more efficient than a traditional bank in South America and makes 100k more per employee than any of its incumbent competitors.
With foundation models, we will be adding more zero-s to these numbers and the operational leverage of banks will be transformed.
LLMs and LVMs help the human scale expertise affordably.
Banks have been automating for decades and are the largest users of AI.
But there is something novel here, especially for financial services. Generative AI (large language models and vision models) are different because, for the first time, we have systems that can deliver both breadth and depth of expertise faster and cheaper than humans.
Before, you had to train a model on a specific task, gather well-labeled examples it could learn from, get feedback, and improve. This was expensive and slow, and the models often failed at anything they had not seen before and anything out of the data distribution.
Moreover, you need specialized machine learning talent to train, optimize, improve, and maintain all these models.
Today, you can have a model to read, write, summarize, reconcile, label, predict, and even write a poem. This model is exceptionally good at generalizing and does not need special examples. Moreover, to deploy it, you do not need specialized ML experts. You need software engineers who are way more common and people who are very good at natural language instruction. This can also be the same person.
Using the right model for the right task.
However, as we all know, this new AI breed has limitations.
Costs can easily spiral when selecting the wrong model. Inference costs, for instance, are too high and are a big shortcoming. When building a small showcase bot for internal use, paying x per token to Open AI is sensible. However, it is impossible to justify processing millions of real-time transactions and paying that level.
The operational reality is that companies are using some OpenAI APIs, with a mix of open source, proprietary, and fine-tuned models to suit their needs. You will have junior models with fewer parameters, specialized data fit for specific tasks, and larger billion+ parameters used to solve increasingly more complex problems or act as oracles to update your local knowledge base.
The GenAI business case
This approach could drastically reduce financial product management costs, upending traditional banking economics and altering profit margins and customer benefits.
To achieve this, we need:
A clear task breakdown and benchmarking, considering the time, computational complexity, cost, and required scale.
The ability to leverage internal data and integrate various language models for specific tasks.
Scalable inference strategies, as inference costs often outweigh training costs for these models. This can be achieved through a strategic blend of large and small models deployed, tailored to the complexity and value of the task.
Artificial Intelligence is fundamentally about optimization. The next groundbreaking development might not be an immediately apparent or new output. What it clearly will do is transform prohibitively expensive processes into very cheap ones.
Such innovations have historically been the catalysts for widespread global prosperity.
Buckle up!
4 Fintech Companies 💸
1. Chalk - The bridge of the Starship Enterprise for ML engineers
Chalk provides everything an ML engineer needs to compute, create, or store features and monitor performance. Chalk can integrate any data source, execute Python as Rust (for performance), and create one source of truth for training vs production. Customers include Ramp, Melio, and Pipe.
🧠 Zero to One financial services ML Ops. Building internal ML models is complex, and off-the-shelf tools (like Apache Spark) need massive configurations to solve financial services use cases. Most Fintech companies will buy in APIs and data from multiple sources but must also build their own ML. When they reach that maturity, they need an ops platform to bring everything together to avoid writing custom code. That's what Chalk solves. It's the code you would have had to write anyway in a way your team wanted to do it. While everyone is excited by LLMs, machine learning is 10 years old, and it's time to make a massive impact. Most Fintech's aren't that mature in ML yet; this helps them get there.
2. Oatfi - Embedded "Stripe Capital for Platforms"
Oatfi provides an embedded working capital solution for platforms that serve businesses. Platforms can enable their business customers to get paid instantly (BNPL for B2B), defer payments up to 90 days, or manage cash flow more efficiently.
🧠 Embedded lending for both sides of the transaction. "BNPL for B2B" is a crowded space, but what makes Oatfi unique is they're providing credit to the buyer and the seller. But, by connecting the invoicing and payment data with the platform and the SMB buyer, they underwrite both parties instantly. 🧠 Bonus thought: This B2B BNPL space is taking more market share from banks and making it available to folks like Apollo (see Good Reads for why that's important).
3. Housr - Student and Landlord billing platform
Housr provides an app for students to pay and manage bills and pays landlords to advertise their property through the service. Landlords get baked-in tenant screening, direct chat with tenants, and a property viewing scheduling service. Students can find properties, split expenses, and pay one bill (which is then automatically paid to suppliers).
🧠 Riches in niches. This UK app with 30,000 users is on track to break even in January and is looking at US expansion. Estimates place the US student population between 15m and 19m people. I don't know what the unit economics look like, but with the right partnerships, it wouldn't take much to make this a healthy business.
4. Brico AI - The regulatory license and charter management AI.
Brico helps Fintech companies acquire, renew, and manage compliance with various regulatory licenses. This includes credit, MTL, and mortgage loan originators in 50 US States.
🧠 Your Head of Compliance just punched the air. Life as a compliance team is filled with spreadsheets, so why shouldn't there be a platform to manage and simplify that. There is known best practice. Half of life is turning spreadsheets into SaaS platforms, after all.
Things to know 👀
The return of the mega-rounds in Fintech? SumUp, the "Square of Europe" that provides payments to 4m small merchants in Europe, has raised $307m to continue its growth. Today it offers invoicing, loyalty, and card readers and is active in 36 markets. It claims 30% YoY revenue growth and is "positive on an EBITDA basis." The press stipulates the funding may be used for M&A.
🧠 There is too much revenue in Fintech for growth funding to vanish. Over the past 12 months, rumors of Fintech's demise have been overstated. With Adyen returning to growth and winning back investor confidence, we may have passed "peak fear" for Fintech.
🧠 Has SumUp reached the end of its organic growth? However, this feels like a sign of a market maturing rather than an amazing success story. M&A is
🧠 Europe has been a tough market for payments businesses. Each European country has different payment networks, banks and regulatory norms required to operate. Notably, large incumbents like Worldline are struggling for growth.
🧠 Payments are a tale of two types of companies: organic growth vs M&A-led. Adyen and its competitors are growing in Europe without M&A, whereas Wordline and its largest competitor Nexi may look to merge to solve their growth challenges.
🧠 M&A-led growth *IS* a race to the bottom. M&A is a classic example of things that work on the spreadsheet but not in practice. Companies quickly grow revenue and improve their cost basis by firing "duplicative" staff. But now they have two companies and tech platforms where one used to exist.
🧠 Platform R&D is a competitive strength in payments. Once a company goes the M&A route, they can't offer the R&D advantage the might have enjoyed with a single platform. I fear that SumUp is being pushed to become just another payments company and that its most dynamic years are behind it.
Google is partnering with Zip and Affirm to pilot BNPL on Google Pay at checkout in the US. Users who press the Google Pay button will be offered Zip or Affirm as a BNPL option to convert more users. UK users can now also store their tax ID (National Insurance number, a bit like a social security number) in the Google Wallet along with other passes like the drivers licence.
🧠 Google Pay is massively under performing. In the US, Apple has a dominant market position at checkout in-store and online. In the eCommerce world, the BNPL buttons and Shopify Pay (Shop Pay) are much more attractive to retailers. Google Pay is just sort of there. Copying Apple a couple of years later on cheaper devices.
🧠 Google cannot get its sh*t together on payments but has the best data moat. In theory, with all of its advertising and commerce data, Google should have the best payment and loyalty experience ever. Instead, it’s Klarna and Affirm who are leading the way here.
🧠 The UK is a great place to win digital identity in a wallet. Driver's licenses and tax IDs are both done at the national level. Unlike in the US, where Apple has to state by state with DMVs, the UK has one Driver Vehicle Licencing Agency (DVLA). While the strategy is still copy+paste whatever Apple did three years ago (yawn), the focus on the UK is smart. The UK is an amazing petri dish for Fintech and identity.
Good Reads 📚
Apollo now originates loans across 16 asset classes, including mortgages, auto, and assets like aircraft or fast food chains. These platforms will originate $45 billion this year, and $75bn within two years. Non bank lenders rarely last the test of time; however, in 13 years of operating asset-backed lending, Apollo has recorded losses of just 0.01%. Asset-backed loans are typically "overcollateralised."
It originates and funds loans directly; it will also distribute some and help fund others who wish to become originators. In effect, Apollo is doing everything across the value chain a bank would, but in a unique way. Apollo argues that the way its insurance business gives it "float" is lower risk and cost than how banks use deposits to fund their lending operations (as Warren Buffet and Berkshire Hathaway first demonstrated).
🧠 Competing as a bank is incredibly hard. Apollo doesn't have regulators pushing up the cost of funding with capital requirements. The FDIC insurance doesn't seem like a good deal when you look at the SVB aftermath. As a bank CEO, sure some depositors got made whole, but the bank was pawned off in an emergency sale. It's not a great thing for the CEO's resume.
🧠 Will regulators ever view "deconstructed banks" like banks? I imagine Bank CEOs hope they will. Deposits can be an expensive way to fund balance sheet operations.
🧠 Why don't banks copy this model? Some are. I'm not surprised to see JP Morgan now look to start its own private credit business. Expect others to follow.
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 :)
Disclosures: (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 * (3) Any companies mentioned in Rants are top of mind and used for illustrative purposes only. (4) I'm not an expert at everything you read here. Some of it is me thinking out loud and learning as I go; please don't take it as gospel—strong opinions, weakly held.