Newsletter · 2026-08-18 · Ashutosh Agarwal

AI Agents Become Fintech's Customers as Stablecoins Become the Rails - Vertical Spotlight: Fintech - Week of August 18, 2026

How AI agents went from fintech's helper to fintech's customer, from agent wallets and stablecoin settlement rails to AI inside bank service, fraud and credit stacks, drawn from operator and founder podcasts for the week of August 18, 2026.

Vertical Spotlight: Fintech

Week of August 18, 2026: AI Agents Become Fintech's Customers as Stablecoins Become the Rails

Covering fintech podcasts from Aug 11–18, 2026

The Landscape

For a decade the story in fintech was "software eats banking." This week the podcasts told a stranger story: the software is now the one holding the wallet.

Across a dozen shows, from Fintech Takes to a16z to a sit-down with Mastercard's CEO, the throughline was the same. AI agents are starting to do the buying, the paying, and even the negotiating on our behalf, and a whole industry is scrambling to build the plumbing that lets a piece of software have a bank account, prove who it is, and settle a bill for a fraction of a penny without a human touching anything.

Two ideas kept colliding all week:

Agents need money, and money needs new rails. A human is fine typing a card number into a checkout page. An agent buying "one second of compute" or "one search request" needs to move fractions of a cent, instantly, globally, thousands of times a day. Card networks and bank wires physically cannot do that. So nearly every builder this week arrived at the same answer, stablecoins (digital dollars that always trade at $1) as the settlement layer for machines. As one founder put it flatly: "No matter if you love it or hate it, that's the only choice you have."
AI is also quietly re-plumbing the boring middle of finance. While the agentic-payments crowd builds the future, AI is already cutting real costs today: 60% of a bank's customer chats answered by a bot, fraud models risk-scoring a trillion dollars of payments, credit models that used to update every few months now updating in days.

The most interesting tension of the week, which we get to at the end: if AI makes every fintech faster and cheaper, does the winner keep the winnings, or does the whole industry just compete its profits down to zero?

Here are the companies people were actually naming.

Companies to Know

Natural, the "bank account" built for AI agents Alex Johnson and Simon Taylor spent the top of their show on Natural, which just raised a $30 million Series A to build what Taylor called an "AI agent orchestration layer" for payments, essentially the infrastructure that lets an agent send and receive money on your behalf. What makes it clever is that the account structure itself acts as the guardrail. Agents pay out of a limited-purpose wallet (built on banking-infrastructure startup Column); withdrawals can only go to a pre-linked account, and there's even a "vault" where an agent can put money in but can never take it back out. Pricing nudges you to keep funds parked in the wallet: transactions from a funded wallet are roughly 50% cheaper than a straight bank transfer. Taylor's framing of the opportunity: Natural could become "a Visa for every other payments rail that touches AI agents and liability." The thing they're really selling, he said, is control, always knowing which agent spent which dollar, and which human is legally behind it. (Fintech Takes, "Not Fintech Investment Advice: Natural, Sky Fusion, Quarters, El Dorado," Aug 12)

Sky Fusion, AI data centers in your basement, financed like solar panels The wildest idea of the week, and Taylor insisted it's fintech at heart. Sky Fusion puts liquid-cooled AI chips inside people's homes, paired with rooftop solar and a battery, and turns the chips' waste heat into hot water. Homeowners get lower energy bills plus payments from companies renting the compute; investors fund the installations the way they fund residential solar today. It starts with 300 homes, expanding to over 1,000. The bet is that "contracted inference capacity" becomes a financeable asset like a solar lease. It fed Taylor's spiciest take of the week: "Fintech is going to be the biggest unlock for AI... financing the AI build-out and creative solutions to do that are going to become a pinch point." (Fintech Takes, Aug 12)

Quarters, flipping the Built playbook to fund a down payment A Canadian rewards app aimed at people who fear they'll never own a home. You link your existing cards through open banking and earn 1% back (2% at partner merchants), but the rewards can only be redeemed against housing: a rent deposit, moving costs, a mortgage down payment. Johnson pointed out it's a clean inversion of the well-known startup Built: instead of using rent spend to drive other spending, Quarters uses everyday spending to pull you toward a housing goal. Because the rewards are only spendable with named real-estate and landlord partners (who foot the bill as a marketing cost), they never sit on Quarters' books as an open-ended liability. (Fintech Takes, Aug 12)

El Dorado, the LatAm stablecoin app whose business customers surprised everyone On the surface, another dollar-stablecoin wallet for Latin America, 13 countries, more than a million consumer users. The interesting part is the small-business side, which Taylor said is "exploding" among importers and exporters. Picture a manufacturer in Bolivia paying a supplier in China: the traditional correspondent-banking route is slow and expensive, and small firms sit at the very back of the line for dollars. El Dorado lets them hold and send dollars from a Mercury-style account without forming a U.S. company. The tell that it's working: the founders (all Venezuelan) opened a physical branch, because business customers wanted to look a human in the eye before trusting them with cross-border money. (Fintech Takes, Aug 12)

Stripe, making stablecoins boring, and killing the checkout page A Stripe executive laid out the company's agent strategy on a16z, and the numbers are the story. Stripe now supports users in about 60 countries in regular currency but roughly 150 in stablecoins, the digital-dollar rails simply reach more of the world. Stablecoins are now native inside Stripe Treasury (you hold a stablecoin balance like you'd hold dollars or euros), and Stripe crossed over $2 trillion in volume. It launched a "Link Agent Wallet" so an agent can use your saved Link credentials (Link has 400 million users), and "Stripe Projects," which lets an agent sign your company up for services like Vercel or Browserbase without you ever visiting their websites. The exec's bigger prediction: checkout pages disappear entirely, "should you even need to go to one, or should you just be able to say, buy it on a product display page? I think just checkout pages will go away." And a proof point that stablecoin rails are already real: a remittance company (Felix) built on stables now handles 5–10% of the U.S. to Mexico corridor, the largest remittance route in the world, in just a few years. (The a16z Show, "Stripe's AI Strategy: Build More, Not Less," Aug 17)

Mastercard, volunteering to be the referee for a world of machine payments Mastercard CEO Michael Miebach walked through the company's agent stack. First, AgentPay: a protocol that "accredits" an AI agent as a recognized party in a transaction and tokenizes every data point, so if your card gets debited for two grills instead of one, there's a paper trail to dispute it. Then the striking one, AP4M, "AgentPay for Machines": true machine-to-machine payments where one company's software buys APIs, data, or compute from another's, paying "fractions of a dollar" continuously as it consumes them rather than via monthly invoices. Miebach's pitch is that this dramatically improves working-capital efficiency, you pay as you use, dialing spend up and down in real time. The underlying rails, he conceded, "could be stablecoin"; Mastercard is deliberately agnostic. Its play is to be the trust-and-interoperability layer on top, so a world of competing stablecoins doesn't become "a plate full of spaghettis." To that end Mastercard is closing an acquisition this quarter of what Miebach called "the largest stablecoin platform out there" (the name came through the recording as "PVNPay," unverified). His summary of Mastercard's role: "Who sits in the middle and drives interoperability...? That's the card." (Motley Fool Hidden Gems Investing, "Mastercard CEO: AI Shopping Agents, Machine-to-Machine Payments, and the New Infrastructure of Commerce," Aug 16)

ZeroHash, the quiet giant powering everyone else's stablecoins Founder and CEO Edward Woodford runs a nine-year-old company (~220 people) that most people have never heard of but that already powers Morgan Stanley's crypto product, plus Stripe and Gusto. It has close to 10 million end customers and expects to be able to reach about three-quarters of U.S. households through its partners by year-end. The news was a Visa partnership: ZeroHash will use stablecoins for two things, "merchant pre-funding" (so a marketplace doesn't have to park idle cash waiting to pay sellers) and global payouts (paying a freelancer in the Philippines "takes seconds, not days via the SWIFT network"), riding Visa's 12 billion endpoints across 190 countries. Woodford also launched an "agentic finance suite" two weeks earlier for agent-to-agent payments, and argued the metrics the industry watches are outdated: stop counting how big a stablecoin is (assets under custody) and start counting on-chain volume and the 2.5 billion stablecoin-enabled accounts that already exist (anyone with a Revolut, Robinhood, PayPal, or Cash App account). (The Rollup, "zerohash CEO: Wall Street's Biggest Are All In On Tokenization (What This Means)," Aug 11)

Sapium, an "operating system" for AI agents, with a credit score for bots Founder Ilan just raised a $45 million Series A led by Dragonfly, with a deliberately eclectic cap table: Anthropic (AI), Coinbase and VanEck (crypto), Okta Ventures (identity), and Accel. Sapium sells an agent studio, a runtime, and a "router" that sends each task to the cheapest, fastest model that can do it. Two ideas stood out. First, a reputation system, effectively a credit score for agents, because a merchant needs to know a given agent isn't "rogue." Second, the reason all this matters commercially: one Sapium customer's inference bill grew "from zero to $1.2 million after three months," at only thousands of users, because an autonomous agent runs 24/7 and burns tokens with no human deciding each step. As Ilan put it, if you default to frontier models for everything, "you go bankrupt... it's negative margin by hundreds of percent." The cost of intelligence, not the technology, is the real barrier to shipping agents. (Tokenized, "From Shopify to Sapiom: Building the OS for Agentic Commerce," Aug 13)

Deel, payroll company turns payouts into a savings product Deel rolled out its own stablecoin, DLUSD, across 80 markets (Latin America, APAC, the Middle East, Africa) and bolted on an "earn" feature, and the adoption numbers are eye-popping for a financial product: 74% of payout recipients deposited into earn, with 85% still there 30 days later, and about 60% of eligible users actively earning. Under the hood the money flows into a lending vault (curated by a firm called Centora, collateralized by Coinbase's wrapped Bitcoin). This is the "distribution company adds a bank product" pattern: Deel already has the users, so a bit of interest income and float lifts customer lifetime value. The same episode flagged Rain, whose stablecoin-linked cards (clients include RedoPay and DollarApp) are expanding from plain cards into loyalty, credit, and buy-now-pay-later. A useful note of caution from Luca, who sits on the board of one of Rain's largest clients: many of these "earn" vaults "yield less than risk-free," and a vault curator that's actively picking assets "is a fund manager... should be regulated as a fund manager," an unresolved gray area before these products get much bigger. (Tokenized, "Deel Goes Live With Stablecoin Earn in 80 Markets," Aug 17)

Nubank, the clearest look at what AI actually does inside a bank On a deep-dive with Vanshap Capital, the guest gave hard numbers on how AI is reshaping the $70 billion Brazilian neobank: more than 60% of customer inquiries are now handled by AI on the first pass; engineering saw a 12x efficiency gain in hours and a 20x cost saving on implementation (per a Nubank blog); and its credit models, which used to be refreshed every few months, now update in days, a direct edge in lending, because faster model updates mean sharper risk decisions. Founder David Velez still owns about 20% and just joined OpenAI's board, which the guest read as a signal of how central AI has become to the company's plan. (Yet Another Value Podcast, "$NU: is Nubank Capital One in 1994 or Capital One in 2006? | Vanshap Capital," Aug 17)

The bank-side plumbing, stablecoins reach small-town America Two banking shows showed the same trend arriving at community banks. The Bank of North Dakota (a state-owned bank) partnered with Fiserv on a "Rough Rider coin" aimed at the 85% of U.S. banks and credit unions with under $5 billion in assets, with concrete uses: bank-to-bank loan payoffs (killing the Friday-to-Monday delay where a borrower pays extra interest waiting on a courier), line-of-credit advances of $5,000–$10,000 available within an hour, and participation-loan reconciliation done by smart contract in under 36 hours. The same episode covered Visa's $2.4 billion acquisition of fraud firm BioCatch (roughly $185 million in recurring revenue), which a Visa executive tied directly to AI: account takeovers and scams "cost the global economy over $1 trillion annually, and AI is enabling these attacks at an unprecedented scale." (What's Going On In Banking, "Kinetic Deposits, the Fraud Stack, and the Inheritance That Won't Arrive," Aug 15; The Community Bank Podcast, "The Basics of Stablecoin, Tokenized Deposits, and What it Means for Your Bank," Aug 17)

The fraud fighters, where "old" AI is quietly winning A useful counterweight to the agent hype: two shows featured companies using unglamorous machine learning to catch financial crime at scale. Kona AI, run by Vince Walden, has processed over $1 trillion (approaching $2 trillion) of payment data, risk-scoring every invoice across hundreds of tests to flag likely bribes before they're paid. His line was that machine learning "is a form of AI; it's not the hot, sexy, generative AI, but that's been around for 20 years." And at SAS, Stu Bradley said the firm regulates 300 billion transactions and saves about $500 million a year in bank fraud, with some state tax agencies saving $500–900 million annually, while noting roughly 5% of all payments involve fraud, waste, or abuse and that chasing money after it's paid ("pay and chase") recovers it only in the single digits. (FCPA Compliance Report, "Data Analytics in Compliance: Lessons from Scoular," Aug 17; Federal Tech Podcast, "Ep. 341 The Future of Government Fraud Prevention with AI," Aug 13)

Frec, brokerage 2.0 for the newly wealthy Founder Mo Al-Adham (ex-Twitter) is building a brokerage around "tax alpha," using direct indexing to harvest losses that defer your tax bill. His target: the ~10 million U.S. households with $1–30 million in investable assets, who control 40% of all investable wealth and grew by 750,000 new millionaires last year. The surprising data point for founders chasing "sophisticated" customers: they aren't fee-sensitive. Frec's fastest-growing product, a long-short direct index available nowhere else without an advisor, charges 2.7%, and churn is near zero (Frec lost only ~1% of assets in Q2). (Founders in Arms, "Building Brokerage 2.0: Direct Indexing and Tax Alpha with Mo Al Adham," Aug 14)

One Debate: Does AI make fintechs richer, or eventually broke?

The sharpest unresolved question of the week came on the Nubank episode, and it applies to every fintech building with AI.

The setup: AI is clearly making Nubank faster and cheaper right now, 60% of service automated, engineering 12x more efficient, credit models updating in days. So the near-term case is obvious. The host's worry was the long term.

His argument, roughly: a financial product is about the most commoditized thing there is. "If I'm looking for a credit card loan, you offer me 19%, you offer me 18%, cool, I go with the 18%." If AI lets everyone run leaner, and, worse, if consumers get their own AI agents that constantly shop their deposits into the highest-yield account and refinance every loan automatically, then the advantage lasts "six months, a year, two, three years" before rivals catch up. After that, in his words, "the profit pool is going to zero," and Nubank's prized 30% return on equity gets competed away.

The guest's rebuttal was equally clean, and it's the bull case for AI-native finance generally: in a commodity business, the lowest-cost operator is the last one standing. Nubank is branchless and already the most efficient bank in Brazil, so it can afford to offer the best rate and still make money, meaning AI-driven price competition hurts the bloated incumbents first. And he leaned on human inertia: even with an AI whispering "move your money," people keep cash at the big banks that pay them almost nothing, because they want security, FDIC insurance, and decent service. If you have the low cost and the service, he argued, "I don't see why someone would go somewhere else."

Neither side won. And that's exactly why it's the debate to sit with, because the same logic hangs over every name in this issue. If AI is a temporary edge, this is a land grab and speed is everything. If it's a permanent structural advantage for the leanest operator, the winners here could compound for a decade. The podcasts this week were confident about the plumbing. They were honestly split on who gets rich from it.