Listen on Apple Podcasts or Spotify

Everyone already seems to know everything about AI and LLMs. I think I missed the part where the basic questions were asked and answered, and given AI will almost certainly be a core part of all our lives, that felt like a gap worth closing.

Luckily, my good friend Dimitri Masin, CEO and Co-founder of Gradient Labs (which builds customer operation agents for financial services), knows all about it, and was willing to answer my “stupid” questions. So, I finally understand why the frontier LLMs cost so much to build, how the training actually happens, and how to build actually useful agents with them.

Training LLMs basically requires running billions of small maths problems over and over again. Imagine a spreadsheet with 130 BILLION rows, where each row represents a parameter or weight that determines how the model responds. Training means continually running over those 130 billion parameters and slightly adjusting each weighting until they’re as accurate as possible.

That training requires a lot of GPUs (Graphics Processing Units), which are electronic circuits designed to rapidly process data in parallel. You might need 50,000 GPUs to train an LLM, and each one might cost $35k. That’s nearly $2bn spent, already, before you’ve accounted for the electricity those GPUs require.

Once you’ve trained an LLM and have customers using it, those weights are fixed, but every time you or I ask Claude or ChatGPT a question, the model has to apply those weights to our input, calculating the most probable next word, one at a time, until a full response is built. So even after training, more GPUs and more electricity are needed, constantly, just to answer questions.

So. Much. Maths. And suddenly, the costs make sense!

I’ve also wondered why some companies release their LLMs as “open source”. The reasons vary. Meta, for instance, has strategic reasons for releasing Llama freely, even having spent billions training it. But here’s the practical upshot: when you use an open source model, someone still has to provide the compute. Whether you’re running it locally on your own machine, or accessing it through a cloud provider, the GPUs and electricity don’t disappear - they just move.

What’s the AI question you’ve been too embarrassed to ask? I’d love to hear it! My DMs are open.

Listen on Apple Podcasts, Spotify, or watch on YouTube

Banks have used machine learning for years. Often found within marketing and fraud teams, these models typically take in structured data and output a probability by way of a score. The machine learning models have traditionally been deterministic, which means that if you ask the same question 10 times, you’ll get the same answer 10 times.

In the new world of generative AI, this all changes.

I recently spoke to Kareem Saleh, CEO of Fairplay. Fairplay is a technology that tests, tunes and validates models for banks, and so he was the perfect person to ask about the risks - and mitigators - of generative AI within financial services. Here’s what we talked about.

Subscribe now

Generative AI and Agents are Not the Same

When thinking about risks, Kareem recommends separating out generative AI from agents.

For generative AI the key risks and mitigants are:

1. Non-determinism: The same input can produce different outputs. A bank that approves everyone above a credit score of 700 can’t have the model denying eligible applicants 20% of the time.

   1. Mitigant: Layer hallucination guardrails on top of the primary model (ideally a second, different base LLM checking the outputs). Back this up with human-led audits sampling ~10% of outputs.

2. Hallucination: The model gives confident but incorrect responses. It is eager to please and sometimes prioritizes that over accuracy.

   1. Mitigant: Layer hallucination guardrails on top of the primary model (ideally a second, different base LLM checking the outputs). Back this up with human-led audits sampling ~10% of outputs.

3. Lack of explainability: Deep neural networks make it virtually impossible to trace why a decision was reached.

   1. Mitigant: Move to continuous system-level validation (not just point-in-time model validation), and apply conceptual soundness testing to ensure the system is appropriate for its application.

For agents, the key risks and mitigants are:

1. Runaway behaviour: Agents taking actions beyond their intended scope.

   1. Mitigant: Narrowly scope the agent’s authority. Assign very specific, limited tasks to each agent, with clear handoff boundaries between them. Break work into small enough pieces that the autonomy to create problems is severely limited.

2. Misaligned objectives: The agent optimises for the wrong goal, e.g. closing cases quickly by denying them all.

   1. Mitigant: Run population-level analyses to detect systematic bias by geography, race, age, or gender. Test whether certain groups are passing through processes at higher or lower rates than expected.

3. Data leakage: Sensitive or regulated information leaking out, especially where there are no clear limits on what the agent can and cannot access.

   1. Mitigant: Define clear action boundaries for agents and implement guardrails that restrict what data and systems an agent can touch.

4. Adversarial attacks / prompt injection: Manipulation of agents through malicious prompts or inputs.

   1. Mitigant: Adversarial testing. Actively try to induce the model into failure before deployment, and continue testing for new threat vectors after launch.

And then there are additional risks when running a multi-agent system:

1. Coordination failures and cascading errors: Agents working at cross-purposes, duplicating actions, or one agent’s small error compounding downstream.

2. Algorithmic collusion / emergent misbehavior: Individual agents each staying within policy, but collectively producing an outcome that would be impermissible.

   1. Mitigant for both: A bank-level “control plane” across all agentic systems, with alerts that flag degradation or harmful outcomes. Best practice also includes maintaining a parallel non-AI fallback system that can be switched on if the AI fails.

What struck me most about this conversation was how everything in banking seems to come back to oversight and testing. I used to think that banking was having a license + transaction reconciliation + compliance. Now I think it might be having a license + transaction reconciliation + oversight.

What do you think banking is? And how can banks use AI safely?

Kareem and I also covered:

• How to test AI for hallucinations and runaway behaviors

• How banks can evidence to the regulators that they are using AI safely

• Why banks may end up having to have 2 systems - their AI native system, and a secondary non-AI system as a failsafe

• What keeps Kareem up at night

If you work in financial services, you should probably listen to this episode.

Listen on Apple Podcasts, Spotify, or watch on YouTube

Finding product market fit is the most important thing any startup can do. But, even though it’s a VC’s first question to any founder, and every accelerator hails it as the key milestone, and CEOs talk about it constantly, no one has actually defined it in truly quantitative, measurable terms.

Until now.

Mark Roberge, Founding Partner at Stage 2 Capital, has done just that.

Subscribe now

According to Mark, product market fit is defined as generating customer success consistently, which is measured through long-term customer retention.

Mark acknowledges, though, that this is a lagging indicator, and so he defines a Leading Indicator of Retention - effectively an early warning signal of the likelihood of achieving customer retention and therefore product market fit.

That Leading Indicator of Retention is a percentage of customers that achieve some event, every specified time period. For example, 70% of customers process five or more transactions every week.

Finding the Right Leading Indicator of Retention for Long Sales Cycles

So far, already one of the most useful insights I’ve heard a VC provide. But, where I got even more interested was when we talked about how to apply this to startups selling to banks, or any other companies that have really long sales cycles.

When a sales cycle might be 6+ months, and implementation might be another 6+ months, and each contract might be 2-3 years, looking at the Leading Indicator of Retention above won’t provide early enough signals.

So, Mark suggests you use a “Set-up Leading Indicator of Retention”. Choose a time period that feels reasonable, and measure the percentage of customers that are set up within that time period.

Once you’ve gotten good at that, move on to an “Engagement Leading Indicator of Retention”. Choose a time period that feels reasonable, and measure how often they engage with your product.

By using these leading indicators, you can get a really good sense of product market fit, even with long sales cycles, long implementations, and long term contracts.

Give them a go and see how they work!

Other Topics

This was only one of the many topics that Mark and I dove into on the latest episode of In Control. The full episode is absolutely jam packed with additional great insights, such as:

• What Board Decks and Meetings should look like for pre-seed and seed stage companies

• What selling to banks (and other companies) with long sales cycles means for raising money from VCs

• Why serial entrepreneurs are usually the ones that chase capital-intensive new businesses (like SpaceX)

• How to optimize pricing and contract term length at the earliest stages of a startups life

• How AI is changing sales

• Whether market timing actually matters

• What, other than scaling too early or too late, can kill a startup

If you are building an early stage startup, and especially selling to banks, I really encourage you to listen to the whole thing. I wish I had, 6 years ago!

Listen on Apple Podcasts, Spotify, or watch on YouTube

On this week’s episode I had the pleasure of speaking to Peter Hazlehurst, the CEO and co founder of Synctera, the company that just acquired my company, Cable.

Peter has a fascinating story, having grown up in Australia and teaching himself to code as part of his “gap year” job working for the government, before traveling to Silicon Valley and landing a job building a core banking system. Many tech jobs later, including stints at Google and Uber, he is now building Synctera, which could in many ways could also be described as a core banking system.

One of the themes of our conversation were the lessons that we can teach our children. My favorite of these anecdotes begins with music.

Make Big Fucking Mistakes!

Music has always been a hugely important part of Peter’s life, and an interaction with the conductor of the Sydney Symphony Orchestra when he was 14 years old gave Peter a life philosophy that has stuck with him since.

Playing second piccolo as part of a school work experience week, the notorious conductor threw his lectern aside and shouted at Peter to stand up and play louder. And, to “make big fucking mistakes”!

The lesson that Peter took away was to create and do and take risks. As Peter put it, “it’s better to do that than to be perfect and nobody knows you exist.”

Creating, doing and taking risks are all lessons that I hope to teach my children, and I love how oftentimes the most important life lessons come from moments that seem to have nothing to do with where we end up, but at the same time are absolutely the reason we ended up where we are.

Laziness as Power

So many great entrepreneurs have these stories of teaching themselves to code to find shortcuts to do something else entirely, with no plan on becoming developers. Bill Gates was the first to coin this “lazy person” idea, when he said, “I choose a lazy person to do a hard job. Because a lazy person will find an easy way to do it.”

Peter’s version of that was during his first job, where he was typing up handwritten quarantine reports in Australia in the early 90s, and he messed up because he was racing to take a girl out for lunch. Instead of typing up the report again, which is what 99% of us would choose to do, he taught himself enough code to fix it, and in his words, “unlock my ability to go for lunch whenever I wanted”.

It can feel hard to allow or encourage our children, or our teams or colleagues, to find their own way through problems, instead defaulting to how things are usually done. Peter reminded me that the most successful people find their own way, and often those ways are far better.

On Acquiring Cable

Aside from the obvious reasons for acquiring Cable - revenue, customers, team - Peter talked in detail about his belief system and how that led to the acquisition of Cable.

In Banking as a Service, it’s easy to say that compliance is not your problem. “We are just the technology layer!” But Peter believes that if you build something, you have a responsibility to make sure that, as far as possible, people can’t do bad stuff with it. In the world of payments, that means taking responsibility for compliance, as well as the money movement and ledgering.

Doing the right thing, even at the expense of growth, is how he likes to build his business.

We also talked about the secret to making businesses, and especially B2B businesses, successful, which is the relationships built over many years.

Those two factors together - the shared belief system of doing the right thing, and the strength of relationships - are what enabled Synctera and Cable to come together.

Other Topics

As well as the above, we talked about:

• Lessons on managing teams learnt from Google

• The similarities of the dot come bubble and today’s AI hype cycle

• How startup life influences parenting styles

Listen along to hear more from Peter, and for a chance to reflect on your own early life lessons. I hope you enjoy the conversation!

Listen on Apple Podcasts, Spotify, or watch on YouTube

My main takeaway from speaking to Renata Caine, who leads Embedded Finance for Green Dot, is that embedded finance, or BaaS, is super hard. And it’s difficult to make money.

I went into this conversation thinking that standing up a BaaS program was mostly about choosing 1 great partner and making sensible decisions around compliance. Now I know how wrong I was!

• A bank getting into BaaS is a bit like a contractor building a house. They may do everything - the plumbing, the electrics, the painting - or they may outsource all those things to others. Deciding what skills you have as a bank, and therefore what you can offer, what you can build internally, and what you need to contract, is step 1.

• We got into detail about the pros and cons of For the Benefit Of (FBO) v Ultimate Beneficial Owner (UBO) account types. An FBO account is a single pooled account at the bank on behalf of a brand or a fintech, whilst a UBO structure means every end consumer (or business) is an actual customer of the bank. There are distinct pros and cons to both set ups, and again, the bank should work out where its skills lie to decide which route to take.

• Making money is harder than it seems! You might make revenue from interchange, interest and/or program management fees, but the size of those buckets varies hugely depending on the type of brand or fintech you work with. Diversification is the key, and scale is needed to make any meaningful revenue. This isn’t an area you can “try out”. You either go hard, or go home.

Listen on Apple Podcasts, Spotify, or watch on YouTube

Early on in my Monzo days I spent a lot of time trawling through sanctions alerts trying to figure out if the people signing up for accounts were sanctioned individuals. It’s one of those operationally exhausting, pretty thankless tasks that every bank has to do. I worked so many sanctions alerts that I even started talking about them in my sleep!

So I thought I knew a lot about sanctions. But it turns out I knew barely anything!

I recently had the privilege of sitting down with Peter Piatetsky, former intelligence analyst in Afghanistan, former Treasury Department official focused on sanctions and terrorist financing, and now CEO and co-founder of Castellum.AI, on the latest episode of In Control, and it blew my mind.

Sanctions are intended to be a political tool used to try to exert pressure on countries, or groups of people, or individuals, in a non-violent way, but I had questions about how effective they really are.

Peter walked me through one of the clearest examples of when sanctions worked exactly as intended. During Trump’s first administration, Turkey detained an American pastor. The State Department tried diplomacy and got nowhere. Then someone suggested sanctions, but instead of targeting low-level officials (which is how it had traditionally been done), they went straight for Turkish cabinet members. After the third cabinet member was sanctioned, the pastor was released. The whole thing took less than a week. That single episode changed how policymakers thought about the tool.

We also got into some genuinely wild territory. Did you know the United Nations has its own credit union? It banks UN employees, including North Korean government officials, because if you’re going to let a country participate in the UN, those diplomats need to be able to buy food and pay rent somewhere. It’s one of the highest-risk financial institutions in the world. I had no idea it even existed!

Peter also explained how sanctions expanded from targeting people and companies to sanctioning actual physical objects, vessels carrying Iranian oil, and eventually aircraft. He also has one of the coolest flexes I’ve heard in a while - he was the person at the Treasury who figured out how to sanction planes, after discovering that Iranian civilian aircraft flying fighters to Syria were then returning to serve passenger routes to European cities.

Give it a listen. I think you’ll come away understanding a system that touches every bank and fintech on the planet, but that most people in finance have never really examined closely.

Listen on Apple Podcasts, Spotify, or watch on YouTube

When I joined Monzo back in 2016, one of my first tasks was enriching merchant data. Every morning I’d log into our BizOps tool and be confronted with transactions showing gibberish like “EC23STAR123”, and my job was to figure out that our customers should see “Starbucks” as the merchant. I’d add the name, upload the logo, and move on to the next one.

I assumed that in the decade since, banks everywhere had solved this problem. It turns out, they have not.

In this week’s episode of In Control with Natasha Vernier, I sat down with Oban MacTavish, CEO and co-founder of Spade, to understand why, and what his company is doing about it.

Spade also just announced a $40 million Series B led by Oak HC/FT, with participation from Andreessen Horowitz, Flourish, Experian, and others.

When a merchant sets up a point of sale terminal, the descriptor field that’s supposed to tell your bank where you spent money is only about 40-50 characters long. That means that merchants often summarize or shorten things. For e-commerce transactions, the “city” field often contains a website URL or phone number instead of an actual location. And the merchant category codes that banks rely on? There’s no category for “marketing software,” so businesses just pick whatever looks closest. Your bank is then making fraud decisions, rewards calculations, and compliance judgments on top of this.

What makes Spade’s approach different is they flipped the problem on its head.

Instead of trying to clean up bad transaction data (which is what everyone else tried, including Mastercard and JP Morgan), they built a massive body of merchant data covering 99.9% of payment-accepting businesses in the US and Canada, and turned enrichment into a search problem. Send them a messy transaction, and they match it against what they already know.

The conversation gets really interesting when we dig into what banks can actually do once their data is good. Oban walks through how better merchant data unlocks personalized rewards (not just generic cash back), smarter fraud detection, and - critically - makes a bank’s data ready for AI. That last point matters more than most people realize right now.

We also talked about what broke during Spade’s nearly 500% year-over-year growth, why Oban is deliberately saying no to obvious opportunities like KYB, and why he thinks we’re living through a genuine renaissance in banking technology.

Give it a listen. I think you’ll come away thinking differently about what your bank actually knows about you.

Listen on Apple Podcasts, Spotify, or watch on YouTube

There are the legitimate predictions - how will interest rates move? What will the outcome of certain elections be?

There are the ridiculous predictions - will Mark Zuckerberg get divorced? Will the US government reveal the existence of aliens?

And then there is sports betting. Predictions about the outcomes of sporting events… which I would just call sports betting… make up 90% of Kalshi’s volume.

As I discussed with Alex Johnson, founder of Fintech Takes, prediction markets are, by law, supposed to be for accurately forecasting the probability of events that have some economic impact.

But here’s where we are - allowing some very, very well funded startups to evade state by state gambling regulation to bring sports betting to the entirety of the United States.

No wonder the largest betting agencies are losing their minds!

Right now, financial nihilism feels very real. For most people I know, and for most younger people, the path to being able to buy a house feels like a gamble. The path to building a solid financial future feels simply impossible.

And so, understandably, people are trying to get rich quick. I think that has always been a big part of the crypto boom. And I can’t shake the feeling that it is behind the astronomical rise of prediction markets as well.

Now, some of the most popular fintech apps in the US are presenting prediction markets as an investment opportunity, alongside things like investing in stocks.

Is this a sensible move for the financial wellbeing of future generations?

Alex and I had a pretty fascinating conversation that covered all of this, as well as:

• What prediction markets are and how they work

• The regulatory landscape they exist within, and that which they’ve managed to avoid

• Some pretty suspect behaviors linked to politics and predictions on these markets

What do you think? Are prediction markets good or bad for financial health? Are they a truly useful tool to predict the outcome of events that have an economic impact?

Listen on Apple Podcasts, Spotify, or watch on YouTube

When I moved from the UK to the US, I was amazed (and scared!) by how everyone uses credit cards for everything.

In the UK, I grew up using debit cards. Tap, money leaves your account, done. You can’t overspend. But in America? Credit cards are everywhere and used for everything. And debt is real! The average American is carrying $6,000+ in credit card debt.

But of course the credit card industry doesn’t care about that number. They’re focused on something else entirely: the $3.5 trillion in annual spend moving across their networks.

That’s more than the entire GDP of the United Kingdom.

Susan Ehrlich came on In Control this week to teach me all about the finance of credit cards. And she would know, having built credit card divisions at Citi, Washington Mutual, Sears, and Amazon.

Banks make money from credit cards in three different ways:

1. Interchange fees. The 2–3% your merchant pays when you swipe.

2. Interest income. Banks earn 16–18% APR on credit card portfolios.

3. Fees. Annual fees, late fees, foreign transaction fees. You name it, they’ve tried it!

Of the three, interest income is the biggest. Banks earn 2–4% return on assets from credit cards, and it’s their most profitable product, ahead of mortgages, auto loans, everything.

And so what would the proposed 10% interest rate cap do to the credit card industry? In Susan’s opinion, it would mean that credit cards would revert to what they were when they launched: a product for wealthy, low-risk consumers only.

Here’s why: banks can’t make money at 10% on high-risk portfolios. They would have to strip rewards, tighten approval criteria, deny access to anyone without a super high credit score. The people the cap claims to protect - paycheck-to-paycheck Americans, people building credit for the first time - suddenly have no access to credit cards at all.

Susan told me something that stuck with me: ‘Price constraints don’t eliminate risk. They just reallocate it.’

The rate cap wouldn’t eliminate the risk of lending to subprime borrowers. It just pushes that risk into the informal economy.

Here’s what else we covered:

• The first credit card was Diners Club in 1950, started because a businessman forgot his wallet at a business lunch in Chicago

• Delta earns $7 billion a year from its loyalty program, which is a huge 20% of total revenue. During COVID it was basically their only lifeline

• Susan launched the Amazon 5% cashback card over a decade ago and explains the hidden economics of who really pays for those rewards

• Whether credit cards have pushed past doing social good into causing real harm, and Susan’s honest take from 30+ years in the industry

Have credit cards pushed past doing social good? Susan thinks on balance they still do more good than harm. I’m not sure I agree. What do you think?

Listen on Apple Podcasts, Spotify, or watch on YouTube

Name. Image. Likeness. How student-athletes are able to receive compensation for endorsements, social media, and appearances. 

Or, how student-athletes are paid to play?

On the latest episode of In Control, Richard Scioli of Analog taught me how NIL works, including how creative the compensation can become. 

Unlimited BBQ for offensive linesmen? Done. 

Decoldest Crawford as the face of… you got it… an air conditioning company? Done. 

Richard got into it all to break down the business of college sports. We covered:

• The three buckets of NIL (Name, Image, Likeness) and how they work

• Why college athletes setting up LLCs is now standard practice

• How 10 people regulate 363 Division 1 schools and 175,000 athletes

• The private equity deals reshaping athletic departments

Listen on Apple Podcasts, Spotify, or watch on YouTube

Whenever my mum comes to the US and we go to a restaurant, she never lets her credit card out of her sight - insisting on following the waiter all the way to the register.

They always look baffled, but she might actually be onto something! In-person fraud rates in the US are much higher than in Europe.

Why is card fraud so prevalent?
We share our 16-digit card numbers everywhere - is that safe?
How do merchants actually protect this data?

To dig into all of this, I spoke to Colin Luce, CEO at Basis Theory, and it turns out there's a quiet power struggle happening.

Visa and MasterCard want merchants to rely on their tokens and delete the original card data.

Merchants don't trust what comes next.

Once you give up that data, you can't switch processors, you can't negotiate, you take whatever pricing the networks decide.

We also discussed in the podcast:

- Why agentic commerce isn't really a payments revolution
- Why MasterCard are getting rid of the PAN
- Whether consumers care about card fraud
- How your Stripe token is useless with Adyen

Listen on Apple Podcasts, Spotify, or watch on YouTube

When I was a kid, I used to deposit my Christmas money at the local building society. I had a little blue book with a running balance that I had to take in, and they would print a new line every time I deposited or withdrew any money. I haven’t touched that account in 15 years. I couldn’t even tell you the name of the building society. But I have a small amount of money in there that I can’t access.

I also have about £100 of Bitcoin locked in a Coinbase account I can’t access because I moved countries and have changed my phone number.

So what happens to the money sitting in old accounts that we can’t access?

It gets escheated.

Escheatment is a legal process where unclaimed financial assets are transferred to the state government after a period of inactivity. Usually three years. The state takes custody of your money, and in theory, you can claim it back.

In practice, two thirds of escheated funds are never claimed. They sit with the state. The state earns interest on them. And that interest doesn’t go back to you.

I recently spoke to Allen Osgood, CEO of Eisen, who walked me through how this actually works. What he described is a system that was designed to protect consumers but has quietly become a revenue source for state governments.

US states are currently holding more than $70 billion in unclaimed property. In Delaware, this is the third largest source of state revenue, behind income tax and property tax. States are shortening the dormancy periods to get access to more money, faster. Illinois has already done it. Others are following.

Allen told me a wild story about a man named Walter, who bought $6,000 of Amazon stock in the 1990s and did exactly what the investment advice says: set it and forget it. He went to retire decades later and found a zero balance. His shares had been liquidated when they were worth $8,000 - not the hundreds of thousands he expected. Everyone involved had followed the rules. Walter still lost his retirement.

The advice we’ve all been given - invest early, hold long, let compound interest work - doesn’t account for escheatment. HODL is a terrible strategy when your assets can be seized after three years of inactivity.

Here’s what else we covered:

• California just turned on crypto escheatment rules - Coinbase has already had to send $270 million to Wyoming

• Escheatment doesn’t just affect banks and fintechs - Uber driver payouts, corporate vendor checks, and even storage unit contents can be seized

• States hire contingency-fee auditors who earn 10–12% of everything they recover - creating an entire industry around seizing people’s assets

• Florida has 150,000 companies filing unclaimed property reports, but only six auditors

• Eisen’s business model: help banks comply with the rules, but more importantly, help them reactivate customers before their money gets seized in the first place

Go to missingmoney.com and search your name. One in seven Americans has money sitting with a state government. And the only thing you need to do to stop your accounts from being escheated is log in once every three years.

Listen on Apple Podcasts, Spotify, or watch on YouTube

Consumer losses from fraud jumped from $3.4B to $12.5B between 2020-2024. That's a 270% increase! Should we be worried about what lies ahead?

Back when I was Head of Financial Crime at Monzo, I watched fraudsters pump tens of thousands of pounds of petrol using cards with just £100 on them. We had their names, addresses, photos - everything! But the police didn't know what to do with it.

And that was before AI changed everything.

Now Yuliya Kazakevich, who's led fraud teams at Charles Schwab, Apple, Adyen, Cash App, and Lithic, is watching fraud evolve in ways that genuinely concern her.

And if someone with her experience is worried, we should all be paying attention!

It was great to welcome Yuliya on the pod to discuss all of this. Here’s what we covered:

• Whether AI companies should be liable for enabling fraudsters

• Why JP Morgan is opening more branches

• How AI is scaling fraud to terrifying new levels

• The shift from unauthorised to authorised fraud

• How law enforcement is overwhelmed and under-resourced to tackle fraud

Fraud is evolving faster than we can keep up - and we need to start paying attention.

Listen on Apple Podcasts, Spotify, or watch on YouTube

Half of all US banks could be gone in the next decade. That's Joe Mancini's prediction - and he would know. He's been through three bank M&A deals, most recently as COO at BankProv when it was acquired by Needham Bank.

I've been talking about bank M&A a lot lately because it affects everyone, not just bankers. We all need a bank to function in the world - to get paid, rent an apartment, receive welfare etc. So what happens when banks start disappearing?

Joe and I got into ALL of it on this episode:

- Why $10.1 billion is the worst size for a bank to be
- How Needham Bank integrated BankProv over a single weekend
- The embedded finance future and who's going to win
- Whether branches will exist in 15 years
- How banks might use your data to offer you loans before you ask
- Why the talent crisis is accelerating everything

If you want to learn more about the future of banking, check out this episode.

Listen on Apple Podcasts, Spotify, or watch on YouTube

I hadn't written a check (or cheque if you’re British/ Canadian) since I was 12… and then I moved to America.

Two years in, I use a modern fintech bank but I still can't send money to anyone without using a third-party app. Why do I need Venmo to split a dinner bill?

The payments system here in the US is NUTS! So I asked Will Messina, CEO of GrailPay, to help me understand it better.

We cover: 

- Why businesses still mail checks like it's 1987
- How Venmo actually makes money (it's not from moving money)
- Why banks charge $20 for a wire transfer
- What stablecoins will actually be used for.

One thing that blew my mind: Delta makes $2 billion a quarter from payments alone - they're basically a payments company that happens to fly planes.

So if you’re as baffled as I was by how banking works in the US, this one’s for you.

Listen on Apple Podcasts, Spotify, or watch on YouTube

When I was at Monzo, we didn’t buy an external core banking system - we just built our own and it ended up supporting 12 million customers.

So when I started working with US banks, I was baffled. Why is there a never ending list of systems like Silver Lake, DNA, and Signature?

As it turns out, they almost all belong to the big three providers: FIS, Fiserv, and Jack Henry.

To shed light on this, I spoke to Stacy Bishop, former sales lead at Jack Henry, to find out why banks don't just build their own.

We covered:

- Why choosing a big three provider is the safe career move for bank executives, even if they hate the service.
- The reality of deconversion fees - where banks have to pay to get their own data.
- How legacy cores stay dominant by handling the double letter US regulations so banks don’t have to.
- How banks are finally innovating by building alongside their legacy systems.
- The inside scoop of how core providers secure long term contracts with banks.
- Why the next 5–10 years will look nothing like the last 20.

If you’ve ever wondered why US banking tech feels stuck in 2005, this is the episode for you. 

Listen on Apple Podcasts, Spotify, or watch on YouTube

Starting a bank seems like one of the hardest things you could possibly do. David Jarvis, CEO of Griffin, did exactly that and walks me through the WHY and HOW of building a bank from scratch.

Getting a banking license is no joke! I thought I understood this from my Monzo days, but turns out I'd only scratched the surface.

In this episode, we get into:

- How the licensing process actually works - from the first phone call to the Bank of England through 4.5 years of iterative feedback
- Why every bank has a £1M monthly burn rate 
- The reality of raising £50M before you have a single customer or £ of revenue
- What actually happens when you send a payment (there's probably another bank involved)
- Why Revolut, with 70M customers, is still fighting for a banking license

David also shared about the less glamorous bits of startup founder life, but you’ll have to listen to find out!