The AI Problem: Why Finance Can't Have Nice Things (Yet)

Hey Fintech Nerds 👋

Stripe had a HUGE year, passing $1.4trn TPV, acquiring Bridge for $1.1bn and Billing passed $500m ARR. ICYMI, I had a conversation with Stripe President John Collison here, and there’s more analysis below.

Brex is reported to have passed $500m ARR, and if it’s heading toward profit, it’s likely also contemplating an IPO.

Meanwhile Bank of America’s CEO is stablecoin curious. Making the clearest statements yet to Bloomberg that “if regulatory clarity allows’ they intend to play there.

You need a stablecoin strategy.

Here's this week's Brainfood in summary

📣 Rant: The AI Problem: Why Big Finance Can't Have Nice Things (Yet)

💸 4 Fintech Companies:

1. Caribou - Intra-company finance, tax & legal as a service

2. Limited - The Borderless account for businesses

3. Accrue - The dollar wallet for Africa (on stablecoins)

4. Revving - Cashflow financing for UK businesses

👀 Things to Know:

1. Stripes annual letter - My reflections

2. Bank of America CEO says it "will go into the Stablecoin business" when regulatory clarity allows

3. Brex’anssance? Brex passes $500m ARR and signs Anthropic & Robinhood.

📚 Good Read: The Future of Small Business Lending

Weekly Rant 📣

The AI Problem: Why Big Finance Can't Have Nice Things (Yet)

AI is Genuinely Magic.

Sometimes, when I play with AI, I feel like a caveman who discovered a mobile phone. Yes, there’s endless hype and its’s easy to get a little cynical, but I promise you, if you feel that way, you’re also just not playing with this stuff nearly enough.

We're not far off being able to vibe-code new realities. (Vibe coding is the name given to writing in natural language and having AI develop apps or products from just that chat)

If you haven't yet used Replit, Bolt, or v0 to create a little app, you should do so immediately. The first thing I did when Grok 3 and Sonnet 3.7 dropped was ask them to make a little Pac-Man game. And it worked. This is a feeling of magic and an unlock in human potential.

Here’s an example of the type of UI you can rapidly create with the new models:

Here’s a couple of finance examples (the second one is immense).

This chap builds Stripe integration through conversation. No coding, just chat. The engineer 

• Talks to their code through their dev environment 

• Builds subscription logic through conversation

• The AI understands Stripe's entire API ecosystem at an expert level (thanks to the Stripe MCP server)

That's a glimpse of what's possible.

That same magic is waiting in all financial systems.

Now imagine this same approach applied to BSA/AML compliance, fraud detection, or credit underwriting. "Show me unusual transaction patterns for customers with recent address changes." "Identify applicants with strong payment history but thin credit files." 

(PS. We do his at Sardine today, as I’m sure do many innovative companies)

But.

Most institutions, banks, and companies can't have this new reality because they don't know how to make it comply.

Compliance and explainability is why we can't have nice things.

So for today’s Rant: 

1. Compliance and Explainability is why we can’t have nice things

2. We really need AI to solve the hard problems in finance

3. Finance is already automated and using models, new AI is extending that

4. We’ve entered a more permissive era for AI use generally

5. The compliance reality is that its never easy to lend or manage risk

6. Outcomes matter more than explanations (explanations still matter)

7. Here’s how to use AI and Comply

8. Because the promise is enormous

1. The Compliance and Explainability Paradox

Imagine you apply for a mortgage and get rejected. The loan officer looks at you sympathetically and says, "Your debt-to-income ratio is concerning." You nod, disappointed but accept this vague explanation.

Now imagine getting rejected by an AI system. Suddenly, everyone demands mathematical precision from algorithms that could actually provide more transparency than human brains.

This is banking's AI explainability paradox - we accept human decision-makers who can't articulate their neural processes, yet we hold AI to an impossible standard.

If we're going to solve the massive issues in our financial system, from financing terrorism to financial exclusion, we need AI. Not to make banks more efficient, but to fundamentally remake the financial system.

So here's the question keeping financial leaders awake: 

2. We REALLY Need AI to help Solve the Massive Problems in Finance

Fraud is growing faster than the finance industry or the global economy. Mckinsey, Juniper, and the Federal Reserve put the global fraud growth rates at between 17 and 20%. Meanwhile, global GDP is growing closer to 2.4%, and the banking industry closer to 6%. 

In aggregate, the bad guys are winning.

LLMs, deepfakes, and adopting digital onboarding meant the bad guys could open accounts and scams from anywhere. These are sophisticated, targeted criminal gangs. Money laundering too, continues to be an enormous problem. 

The recent $3 bn fine by TD Bank showed that we have not solved this issue. The crypto exchange Bybit was hacked, and Lazarus (North Korea's hacking arm) stole $1.4bn. AI fixes this, but only if we can explain it. 

The bad guys don’t have to explain their AI.

The challengers aren’t under as much pressure to explain their AI as the banks are.

Fixing this explainability paradox can unlock massive growth if we reverse engineer the problem from first principles.

One of my first principles is finance is moving faster. Year by year, day by day, it gets a little faster. This isn't a new story. Finance has always been about finding ways to automate, scale, and move faster. AI is just the next chapter in that journey.

3. Finance is becoming Automated - AI is just an extension of that

Mo’ AI, Mo’ Problems.

The overarching first principle of change in finance is that over time, everything is becoming faster, global and more automated. You can see it throughout history.

• The ATM is an Automatic Teller Machine. The name is a giveaway, its job was to be the bank teller and give you cash, 24/7. 

• ACH stands for the Automated Clearing House. Instead of manual check clearing, why not automate that with digital files (American’s have a handy name of labelling things after their function)

• The credit and debit card avoids the need for cash entirely, and works 24/7, and in most parts of the world.

• Online and mobile banking are “branchless,” and work 24/7 and in most cases instantly. 

Two of those four examples literally use the word automated. Following this logic why wouldn’t we automate underwriting, fraud prevention, customer service? The reality is we already have to differing degrees. Partly due to customer preference for human interaction, but partly due to different risk appetites.

We’ve seen a similar level of increasing sophistication over time in how institutions apply AI.

The explainability gap widens as model power increases. So does the compliance overhead.

And here's the cruel irony: The models that could do the most good – detecting the most fraud, approving the most deserving borrowers, spotting the most money laundering – are exactly the ones we're most hesitant to use because they're newer and often organizations don’t know how to explain them (or work with the smaller supplier).

There’s an old saying, the future is here it just isn’t evenly distributed. Due to consumer and institutional risk appetites being different. 

• Are you an early adopter who’s curious to try out a Waymo ride, or are you a little freaked out every time you see a self driving car?

• Are you a CRO who’s trying out AI Agent tooling to screen false positives in KYC or Sanctions, or are you worried about explainability and bias?

Finance has always automated. The question is how quickly we'll figure out the right guardrails.

4. You’re probably allowed to use AI now but should you?

The shift in regulatory tone towards AI in the past 6 months is palpable. A change of US administration has created a much higher focus on the benefits of AI. In the cital of regulation itself, the EU, there’s even an increasing push to balance regulation with competitiveness.

And yet there’s a catch.

Our threshold for machines is much higher than it is for humans. Self driving cars save lives, but if they get into a road accident its headline news. We’re fine for humans to fail, but we have a really hard time with it from machines.

I heard a story yesterday that tickled me. In the early days of machine learning a search engine’s internal product teams got stuck in endless debates about if they should use ML, because they couldn’t explain how it got to the outcomes.

So there are two things here.

1. We, humans, whether as consumers or staff feel a bit weird about AI

2. Regulation may have stepped back, but there’s a lot of law financial institutions have to comply with that’s still standing (and a lot of states, agencies and courts that could still enforce those laws).

AI promises to make our lives better, unlock economic opportunity and could even catch more bad guys.  

5. The Compliance Reality - Its never easy

Laws don't exist to keep us from innovating. They exist to prevent harm and create the society we want. And last I checked, financial exclusion, money laundering, and fraud ARE harms we want to prevent.

Lots of laws have very good intentions at the abstract level

• Equal Credit Opportunity Act (ECOA) prohibits discrimination on the grounds of demographic data. This is a good intent!

• Bank Secrecy Act (BSA) requires financial institutions to file reports on suspicious activity, and know their customer so criminals can’t launder money. This is a good intent! (PS, yes I know AMLA and Patriot Act up levelled this)

• Electronic Funds Transfer Act (ETFA). Provides a liability framework for wire, credit and debit card transactions so consumers are refunded in the event of fraud. This is a good intent! (I’ve used three whole exclamation marks)

Regulations like ECOA, BSA and ETFA weren't created to stifle innovation. They were created to prevent harm. Yet there's a paradox - sometimes demographic patterns correlate with risk. We want systems that create fairness, not entrench systemic inequalities. So a pure, cold statistical model has a meaningful risk of entrenching those inequalities. 

That concern is valid and prompts the question:

If your AI can't explain itself, how do you know it isn't causing harm?"

This is why we push for explainability.

The most dangerous AI isn't the black box that works mysteriously well. It's the glass box that works mysteriously poorly. 

6. Outcomes matter more than Explanations

There I said it. 

What if an AI lending model with limited explainability:

• Reduces defaults by 30%

• Approves 25% more applicants from underserved communities

• Cuts false positives by 60%

This isn't hypothetical. I see this every day in my $day_job. 

The dirty secret of compliance? It's never been about understanding every gear in the machine. It's about managing risk and delivering better outcomes.

Regulators know this. SR 11-7 guidance (the boogeyman of AI use) doesn't demand perfect transparency - it requires "effective challenge" and focuses on materiality and risk.

I used to describe compliance officers as nice folks that scare easily. Their job is to worry and think about edge cases. And frankly, this stuff is hard, it is very complex. But the path forward isn't a perfect explanation - it's demonstrably better outcomes with sufficient guardrails.

7. How to Use AI and Comply

We’re on the cusp of an AI revolution and if we’re going to unlock the economic benefit we need financial institutions, and, well, the rest of us, to get out of our own way.  Most CROs would welcome AI tools that improve KPIs across their departments. The barrier isn't philosophy but practicality - how to satisfy model validation requirements.

For some, this is the basics of applying AI, but for most people with a day job in finance, its the kind of thing you’d never have to think about. Now I think we all have to get good at model validation and financial services compliance. Every one of us. Yes even you (VC, engineer, CEO or whomever you are).

Here’s an oversimplified flowchart with some of the basics.

1. Define Success Criteria: This is make-or-break. Before you deploy any AI, answer this: What exactly does "better" look like? 30% fewer false positives? 25% more approvals for qualified borrowers? Double the fraud detection? No improvement metric = no way to prove success.

2. Shadow Mode Implementation: Run AI alongside humans, compare decisions humans make vs the AI, and compare those against success criteria.

3. Feature Attribution: Know which variables are driving decisions through methodologies like LIME and SHAP, giving you visibility into the decision forest without having to inspect every tree.

4. Adversarial Testing: Push your models to breaking point with synthetic edge cases that rarely appear in training data but could emerge in production, because the most revealing test of AI isn't how it handles the expected but how gracefully it fails at the unexpected.

5. Outcome-Based Validation: Tracking real-world results across demographic groups, default rates, and economic outcomes. Make this a consistent feedback loop thats periodic and for every major model change.

6. Human Oversight: Keep humans as final decision-makers during transition. Gradually increasing automation as confidence grows through demonstrated performance.

Over time, we’ll learn that the human oversight is less and less needed. Just as Tesla insists on keeping your hands on the wheel as you drive, we’re probably not quite there yet where the AI Agents can run a financial institution by themselves.

What is fascinating though, is that AI is doing a better job of doing AI oversight than most humans. Over time we’re going to move up several layers of abstraction, and be explaining how the AI oversight worked and what it found. (Although that’s a rant for another day).

We're at an inflection point. The gap between what AI can do and what we're letting it do in finance is widening daily. Every day financial institutions hesitate, they're leaving economic value on the table, letting criminals find new attack vectors, and watching competitors (especially those with less regulatory burden) race ahead. The technology is ready. The regulatory space exists. What's missing is the operational courage to move forward.

Let's zoom out from compliance for a moment and remember why this matters. Beyond the regulatory checkboxes and model validation lies the real prize – a financial system that actually works better for everyone.

8. The Promise is Enormous

When teams can vibe code new realities, productivity is unlocked.

We need explainable, compliant AI that scales to get to that reality.

And that starts with a focus on outcomes.

Consumers don't care about model validation documentation. They care about getting approved when they should and being protected when they shouldn't.

The financial system has always operated with imperfect understanding. What's different now is our opportunity to build something better—AI systems that are more fair, more efficient, and more transparent than the human systems they augment. Not because they explain every calculation, but because they deliver demonstrably better outcomes for everyone.

If we want the AI Agent that can help us buy invest, borrow and save as consumers, and we want AI Agents that can underwrite the loans and catch the bad guys we need one thing.

The financial world operates on trust, but we've confused trust with perfect understanding. We trust human loan officers who can't explain their neural processes. We trust markets whose complexity exceeds any single person's comprehension. We trust systems with known biases because they're familiar.

AI doesn't need to be perfectly explainable - it needs to be explainable enough to build justified trust. Enough to show it's not perpetuating harmful biases. Enough to demonstrate it delivers better outcomes than what came before.

Knowing when to trust AI and when to challenge it isn't just the most important skill you'll develop this year - it's the difference between a financial system that works for a privileged few and one that works for everyone.

The explainability paradox isn't a regulatory problem. It's a human one. 

And solving it starts with asking better questions.

ST.

4 Fintech Companies 💸

1. Caribou - Intra-company finance, tax & legal as a service

Caribou helps companies that are expanding internationally, have multiple entities, or might have a separate regulated entity (like a Fintech company) manage payments, taxes, and compliance. Finance teams can treat every company and jurisdiction like its own, as the platform that then provides all the paperwork and documents a local jurisdiction will need for filling.

🧠 The Office of the CFO had one killer app left. Multi-entity is incredibly hard to pull off, but every company reaches a stage where it makes sense. Consider a company like Revolut, who is now in Europe, the UK, Mexico, and countless Middle Eastern countries. Can you imagine the paperwork?

Side note, I think there's a use case for personal holdco's here. Think about the number of freelancers, creators, and investors with various shares and interests they could manage through Caribou.

2. Limited - The Borderless account for businesses

Limited helps companies and founders store and manage multiple currencies and transfer them across borders using stablecoins. All funds are typically held as EURC or USDC, with no fees for transfers or holding balances. They offer physical and virtual cards with spend controls for $11.99 and $195 monthly.

🧠 This works because the "account" is really a self-custodial wallet. This is a go-to market built on stablecoins, credit cards, and the global B2B entrepreneur market. That's not a bad thing, it's just a question of whether they can execute.

3. Accrue - The dollar wallet for Africa (on stablecoins)

Accrue lets users store, get paid, or pay in dollars. It also offers up to 4.5% yield pa on dollars held in the wallet. The service comes with a debit card that allows users to pay for global services like Netflix, Amazon, Apple, Spotify, and the Playstation Store.

🧠 Stablecoins are BaaS for non-US countries. Think about the amount of things you pay for online with a card-like subscription. Now imagine not having that. Now, imagine someone comes along with a Neobank that opens that up and protects you from currency debasement. These products are a no-brainer for their customers. The next big global Neobank could come from this type of proposition.

4. Revving - Cashflow financing for UK businesses

Revving helps publishers and ad tech businesses manage the cash flow gap to getting paid (often over 180 days). Revving provides instant access to sales revenues by collecting data from ad networks like DoubleClick, Amazon, or the App Store.

🧠 This is a clever niche data + lending play and the timing could be really good. Historically, nonbank lenders rarely do well in Europe because we have a weak credit supply ecosystem. They either become a bank or stay small. The time is now for that to change. There's a giant opportunity for specialist lending in a growing UK and Europe, but we need a private credit ecosystem that is excited by the opportunity. Asset managers like DWS and EQT could be a real unlock for European growth.

Things to know 👀

1. Stripes annual letter - My reflections

Stripe passed $1.4trn in TPV, Acquired Bridge, released and AI Agent SDK, and counts 80% of the cloud 100 and 78% of the AI 50 as clients. They’re indexed to growth companies.

🧠 A good year for TPV. A 40% uplift in TPV was driven by the sheer scale of AI companies hitting revenue milestones like $10m and $100m at a record-breaking pace.

🧠 Will they IPO soon? I doubt it. John said the business of finance is “lumpy,” and they value being able to invest over a decade-long cycle. If the private markets can fund their ambitions and their profitable, there’s no real reason to have one eye on the exit.

🧠 Why the big focus on fraud? Because it’s what keeps merchants, marketplaces and banks up at night. Stripe is a bellwether for what the industry is talking about, and they’re talking about fraud.

🧠 Why is Stripe doing SMB terminals? A huge segment for Stripe is vertical SaaS, so their customers customer is an SMB. Its an interesting distribution play. They’re also seeing when a sector is served by vertical SaaS, it unlocks growth for that entire sector.

2. Bank of America CEO says it "will go into the Stablecoin business" when regulatory clarity allows

The CEO said stablecoins would become like a Money Market Fund or other assets Bank of America currently offers its corporate and cross-border clients if regulatory clarity came.

🧠 BofA is a major cross-border, LATAM to US bank. Their private banking and corporate banking market coverage here is significant. They’re often a primary choice for immigrant populations and remittances.

🧠 Big banks moving into Stablecoins would dramatically increase the liquidity and user base. With 120m consumers, they’d become the ultimate on and off-ramp in both jurisdictions. Essentially they’d use stables as the alternative to SWIFT.

🧠 The corporate customers would love instant, 24×7 liquidity across borders. Would be a huge win.

3. Brex’anssance? Brex passes $500m ARR and signs Anthropic & Robinhood.

Brex expects annual net revenue to pass $500m in 2025 cited by Bloomberg and The Information. Brex says they have seen 3x YoY revenue growth and expects to be profitable. Previously they had said that would be the trigger to go public.

🧠 This is a great comeback story. When Brex founded the BaaS infrastructure, many of their peers relied on didn’t exist.

🧠 What was a handicap is now an advantage. I think of Brex as the “Adyen” (if Ramp is Stripe). Adyen was always arguably more suited to enterprises (or more focused there, at least). Brex has ended up there because its infrastructure allowed it to go global sooner, with better unit economics.

🧠 The “not for small companies” was a message they were coming back from. That came at a time when insiders reported revenue was stepping back and competition was at an all-time high.

🧠 The “one release” quarterly strategy is very counter-intuitive for the bankers watching from afar. While Ramp talks up their shipping speed, Brex is doing quarterly, big drops, driven from the top by Pedro.

Good Reads 📚

1. The Future of Small Business Lending

Alex Johnson breaks down small business operating systems like Shopify, Square, and Quickbooks, which cover front, middle, and back offices. Lending tends to get embedded because it's convenient for the user, and the operating systems have a data advantage traditional lenders miss.

Square's merchant cash advance (MCA) is the star product of embedded lending, but it doesn't solve the slew of jobs to be done small businesses have. Embedded Lending is evolving to show up at the point of need, receivables, payables, and working capital.

🧠 Every one of Alex's deep dive essays are Fintech Canon. This is no exception. His mental model for a small business's various functions, data, and lending products is exceptional. No notes.

Tweets of the week 🕊

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