Newsletter · · Ashutosh Agarwal

AI Agents Start Spending Real Money and Finance Has No Rules for It Yet - Vertical Spotlight: Fintech - Week of October 6, 2026

Startups and venture newsletter for the week of October 6, 2026 (fintech podcasts published September 29 to October 6). Instinct reports 1 billion dollars of transactions and a 10 billion dollar valuation, Meta's Muse shopping agent was blocked by Amazon, Kapital earned 55 million dollars in six months, and the shows split on who pays when an agent gets a purchase wrong.

Vertical Spotlight: Fintech

Week of October 6, 2026: AI Agents Start Spending Real Money and Finance Has No Rules for It Yet

The Landscape

Seven weeks ago this newsletter covered fintech builders laying the groundwork for AI agents that pay: wallets, stablecoin rails and "accreditation" protocols. This week consumers actually used them, and the podcasts spent most of their time on what happened next.

The trigger was two consumer agents. Instinct is a startup assistant that lives in your text messages. Its founder says it has already handled $1 billion of transactions while still invite-only, and it raised a $1 billion round at a $10 billion valuation. Muse is Meta's shopping agent. It buys things for you through Stripe's Link checkout, and Amazon cut off its access within days. Almost every fintech show touched one or both.

Three threads came out of it:

  • Banks and payment networks were built on the assumption that a human approves every purchase. That assumption is breaking. One infrastructure founder said banks will need "multi-layered agentic firewalls." A fraud CEO said bad actors can now spin up thousands of fake accounts "with one single prompt." A senior Visa security executive said the time it takes attackers to exploit a newly found software flaw is dropping from 18 months to under an hour.
  • Nobody has decided who pays when an agent gets it wrong. The question came up on at least four different podcasts: who absorbs the refund and the dispute when an agent books the wrong hotel or buys something you didn't want. Today the card-network rules give merchants no way to win those disputes.
  • Meanwhile, the unglamorous AI is already cutting costs. A Mexican business bank runs on AI-native underwriting and earned $55 million in the first half of the year. A mortgage automation company says it saves lenders $400–$500 per loan. A Philippine lender cut a three-week credit review to minutes. Nubank has 15 million people using an AI "private banker."

The open question, which we come back to at the end: are consumer agents a genuine shift in how people buy things, or a hyped demo with company-reported numbers? The podcasts were sharply split.

Companies to Know

The agents doing the buying

Instinct: $1B in transactions while still invite-only

Instinct is a personal assistant that lives inside iMessage or WhatsApp, with no separate app. It books travel, makes purchases and handles errands. On My First Million, the hosts relayed the founder's numbers:

  • About $1 billion in transactions so far, roughly 40–50% of it travel.
  • 40% of users hand over their credit card within three weeks, and about 80% of those stick around.
  • Growth of about 10% compounding daily.

One host's reaction to the travel figure: "So imagine you're Booking.com or your Expedia. Sorry, boys. Game's over."

The business model is what really caught the hosts' attention. Instinct says it will be free forever and take a cut of each transaction. Shaan Puri argued that is a stronger position than Google's: an advertiser pays Google just to be shown, and the user still decides. With an agent, the agent decides and the merchant pays. "If you're the agent that makes the buying decision, that's even more powerful than the most powerful business model of the world so far, which has been the Google ad platform." Sam Parr was more cautious. He noted that Google's founders also once said Google wouldn't run ads. (My First Million, "We tested Instinct, MUSE and Grokbot. They're ridiculous.", Sep 30)

On 20VC, Jack Altman confirmed his Benchmark partners led a $1 billion Series C at a $10 billion valuation. Instinct had raised at $2.5 billion, then $5 billion, then $10 billion in about three to four weeks, with a team reported at around 14 people. Altman called consumer agents that "can interact with the entire third party Internet" a shift "on the level of chat and coding." Jason Lemkin set the test: "Will we run Instinct or Muse eight hours a day? If we do, I guarantee it wins." Another panelist was uneasy about the pace: "What worries me is when you have three rounds in three weeks with no material movement in between." (20VC, "Instinct Raises $1B at $10B Valuation | AMD Buys Fei-Fei Li's World Labs for $8.2B…", Oct 1)

On Fintech Insider, Irina, a payments executive and former Grab staffer, said she had been using Instinct for about a week: "It's basically my executive assistant now… So I don't have a separate app anymore." It reads her email, sends her briefs, and texts her husband when her flight lands. (Fintech Insider by 11:FS, "1104. News: Nubank rules out Monzo takeover, Checkout.com opens the books, and Ramp hits Broadway," Oct 5)

Muse (Meta): paid through Stripe, blocked by Amazon

Meta's (META) agent gives each user a sandboxed cloud computer and pays through Link by Stripe. Link covers lost or damaged items and offers free returns. Ecommerce consultant Rick Watson explained the Amazon fight. Amazon (AMZN) says Meta never asked permission, and that agents "should operate in the open and accept a retailer's decision." Watson thinks that's fair, but awkward for Amazon. Amazon runs its own "Buy For Me" agent that lists products from brands' websites unless a brand emails to opt out. "Agents should identify themselves and stores should get a say… But it's also an awkward argument for Amazon to make while running an opt-out program of its own." (The Watson Weekly, "Anthropic Has $518B in Compute Bills and $4.6B in Revenue," Oct 5)

On My First Million, the hosts gave the deeper reason Amazon blocked Muse: disintermediation, meaning a middleman that sits between Amazon and its customers. People default to Amazon or DoorDash because comparing prices is tedious. A tireless agent will happily go to the brand's own site "and get it for $4 cheaper."

Robinhood: 150,000 people already letting agents near their brokerage account

At Robinhood's (HOOD) Hood Summit, CEO Vlad Tenev said its first agent feature was clunky and still caught on. Users had to take an outside AI coding agent, such as Claude Code or OpenAI's Codex, and "plug it into their Robinhood account. So they had to kind of stitch it together." Even so, "we had about over 150,000 people using our Agentic offerings," and Robinhood is now building the feature into the app. Dan Nathan's read was that the value isn't agents trading on autopilot but "idea generation… risk management. I think that's something that pros have been doing for a little bit now." (CNBC's Fast Money, "Micron Reports Results… And Sights & Sounds From Robinhood's Hood Summit 9/30/26," Sep 30)

The plumbing for agents and stablecoins

Hyperlayer: firewalls for a bank's non-human customers

On Breaking Banks, Brett King interviewed a Hyperlayer executive with a long banking career about a problem most banks haven't thought about. Today a card approval is simple: right customer, approved merchant, enough money. An agent told to "buy me some Nikes" may arrive with no specific merchant, a list of acceptable stores, an exact product and a condition like "I want my shoes tomorrow." Every one of those is a rule the bank has to check, in milliseconds. "The number of sub authorizations that a bank is going to have to process… are going to be dramatically higher in number and volume… and far more complex." His answer: "The banks are going to need essentially what we call multi-layered agentic firewalls."

King raised the identity gap: "We don't have… digital identity layers capable of identifying the origin of all these… agents coming in. For now, we're making the assumption that every agent comes from a human at… some point in the chain."

The guest's advice for banks, Nubank and Revolut included, was to keep new agent features out of the core banking system "no matter how modern and sexy you think it is." Instead, add a layer on top. He was also candid about AI's limits. Hyperlayer writes all its code with AI agents and needs "far fewer lawyers," but "AI did not conceive of the question… it makes a lot of mistakes." (Breaking Banks, "Emerging Infrastructure for Next Gen AI-based Financial Tech," Oct 1)

Checkout.com: 70% of merchants are preparing, but consumers aren't ready yet

Rory from Checkout.com gave a hard number: more than 70% of its merchants are exploring agent-driven shopping or payments. But more than 60% of consumers it surveys still worry about refunds and "what happens when something goes wrong." His view is that we're in an era of "agentic discovery", where agents help you find things, and agent payments come later. He predicted businesses will let agents buy from each other first, for things like ad slots and search fees, and that this will build the trust systems consumers then rely on. Irina disagreed on air, calling agents "one of the biggest transformations… since… the internet in the 90s." They admitted they'd argued about it over dinner. (Fintech Insider by 11:FS, Oct 5)

Stripe and OpenUSD: stablecoins are going on cards

A stablecoin is a digital token designed to always be worth one dollar. Stripe now plans to offer stablecoin-funded cards in 100+ countries by year-end. Customers already include Kraken and Ramp. Ramp uses the cards to take its corporate card into new countries "using stablecoins instead of rebuilding banking and payments connections market by market." About $1.2 billion of stablecoins were spent through cards last month, triple a year ago and "still just a fraction of global card payments." Stripe's crypto lead Henri Stern: "By and far, the bulk of our work happens with stablecoins." (Bitcoin And, "IM Blech! | Bitcoin News," Oct 5)

On Tokenized, Simon Taylor and guests discussed OpenUSD, a stablecoin jointly backed by Stripe, Visa (V), Mastercard (MA), Coinbase and Shopify. It has about $477 million in circulation, no fees to create or redeem it, and redeems one-for-one for dollars. Banks liked that redemption feature most when Taylor met them at the industry's big Sibos conference: "A dollar is a dollar. It should work like that." Visa's head of crypto, Kai Sheffield, was blunt about where demand isn't: "We still have not seen any meaningful demand for direct consumer-to-merchant stablecoin payments." The real uses are stablecoin-linked cards and business payments. As Sheffield put it, "what used to be a wire is now a $100,000 stablecoin payment." Taylor's advice to banks for the agent era: "Give the agent a reason to choose you. You've got to make your bank worth choosing." (Tokenized, "Banks Are Going Onchain on Weekends," Oct 5)

Cashi: a stablecoin card that found different users than it expected

Esther Wong helped launch Crypto.com's card and is now one month into a public beta of Cashi, a stablecoin spending app and cashback card. She expected freelancers and digital nomads. "It's actually not really the creators yet. We are actually seeing users that likely seem to be expats." In Venezuela, where merchants see up to a 17% gap between official and street dollar rates, people clearly want to hold dollars. But spending them by card is hard, because shops have QR codes instead of card terminals. On cost: as card volume grows, Visa and Mastercard rebates can cut unit costs "by up to 80%." On agents, she was deliberately cautious. She expects "agent-assisted payments first because of the trust that humans need to give to the agent," and fully autonomous agents "maybe two, three years" out. (On The Brink with Castle Island, "Esther Wong (Cashi) on Making Stablecoins Spendable," Oct 5)

AI that's already earning its keep

Kapital: the AI-native business bank making $55M in six months

The most substantial operator interview of the week. René Saul, CEO of Kapital, Latin America's largest business-focused fintech, announced a $125 million Series C extension (equity plus debt) that doubles the company's valuation, now over $1 billion. His numbers:

  • $700M+ revenue run rate
  • $55M+ net profit in the first half of 2026
  • Loans and payments up more than 300%
  • Deposits up 234% to over $4 billion
  • 350,000+ customers, of which 250,000 are businesses
  • 3,500 employees

How the AI works, in plain terms: in Latin America, almost every business invoice is issued electronically with a digital stamp the tax authority can track. That gave Kapital a huge stream of data. "We don't need to connect to an ERP… We just ask their tax ID and we can understand everything about the company." The company uses that to underwrite small suppliers and give them credit, then sells banking to the big companies those suppliers sell to.

On payments, Kapital claims 10% of the Mexico–U.S. cross-border business payments market, and says it moves money abroad "in less than a minute," around the clock, against the usual one to three days. Saul argued early that fintechs must become regulated banks, and some VCs pushed back. Kapital bought Banco Autofin Mexico to get a license. "Fintechs eventually will become banks… So now we can say that they were wrong and we were right." He expects each region to consolidate to "1 or 2 players." (The Reboot Chronicles with Dean DeBiase, "Latin America's Fintech Unicorn: Rene Saul, CEO of Kapital," Sep 30)

Nubank: 15 million people with a "private banker in your pocket"

Ethan Eismann, Nubank's (NU) Chief Design Officer (ex-Airbnb), said a Brazilian group of around 15 million customers is using the bank's AI Private Banker. "We're not just slapping on a chat bot." It focuses on goals, spending habits, and finding and cancelling unwanted subscriptions. The bigger ambition is personalization. Nubank today runs one app for 140 million customers, and Eismann wants "every single… customer [to have] a product that's slightly unique." If a specific loan makes sense for you at a given moment, "it should be unavoidable." That line will make some consumer advocates nervous. Agent-style features are coming to Nubank's new U.S. app. (Tearsheet Podcast, "Nubank's Ethan Eismann: 'You have to love your customers fanatically for them to love you fanatically,'" Sep 29)

Gateless: automated mortgage underwriting for $400–$500 less per loan

Mike Brown of Gateless, which automates mortgage underwriting, said its financial model shows lenders save $400–$500 per loan on average. The savings come from fewer manual hours, more refinance applications that make it to closing, and faster funding. The host framed it as cutting up to 10 days off a loan. When an application comes in, Gateless calculates income and assets and writes them straight back into the lender's system, so "a loan officer… can approve a borrower at like literally at night." Lenders pay only if the loan closes. It covers about 80% of the market today (standard conforming and FHA loans), with self-employed borrowers next and other loan types in 2027.

His warning to lenders weighing newer AI vendors: ask whether "they have six people that are former underwriters that have worked at the GSEs looking at loans during the week." The GSEs are Fannie Mae and Freddie Mac. "Or are you going to be… the beta tester and the auditor for them?" On a fully automated "one-click mortgage": "Unregulated, you could make a decision today. We could introduce agentic AI into our tech today… The reality is there's regulation." New state rules, he warned, "are naming decision engines, they're naming calculators." (Chrisman Commentary, "10.1.26 Fourth Quarter Changes; Gateless' Mike Brown on Automated Decisioning; Fed Chatter Abounds," Oct 1)

Elio Mortgage: loan officers and engineers in the same room

Elio founder Oren Michaely (ex-Microsoft and private equity) announced a $5 million pre-seed led by Motive Partners and Social Leverage. The company has 10 engineers and about 40 loan officers, and its thesis is organizational: "you need loan officers and engineers to sit in the same room." Loan officers co-design each feature, then present it to the whole company and "own" it, "because everyone feels like they're the product managers all of a sudden."

The concrete result: a loan officer answers 10–20 key questions on a phone call. The borrower instantly gets a list of documents to upload. The standard loan application then fills itself in from those documents, "without [the loan officer] touching it." Host Dustin Owen's blunt take on what this means for pay: loan officers will need to close five times as many loans to earn the same money, "the good news is you'll be able to because of the technology." (The Loan Officer Podcast, "Organize Around What Technology Cannot Do: Elio Mortgage's AI-First Approach | Ep. 667," Oct 1)

GIA: three-week credit reviews done in minutes, in the Philippines

Zach Marks runs GIA, a small-business lender in the Philippines. Its "AI credit engine" reads "handwritten invoices, maybe a receipt from the local BIR" (the Philippine tax bureau) and turns that mess into structured data for credit decisions. Bank analysts do that work by hand and "maybe it takes them three weeks to work on a file. And we take that process down to a matter of minutes." He credits it for a "really low" rate of bad loans; he gave no figure. The new move is licensing the engine to other banks, and using those partnerships to expand across Southeast Asia. (Startup Hustle, "Building the Ramp of the Philippines: Fintech for the Underbanked," Oct 1)

The defenders

SEON: fraud now moves at machine speed

Tamas Kadar co-founded SEON, a fraud-prevention platform whose clients include Revolut, Plaid, Nubank and Bilt. He gave a clear explanation of what AI changed: "Fraud used to leave humans' seams… a journey made too fast to be real. AI has truly removed that constraint, and now identity creation happens at machine speed." Building a mass fake-account operation "used to take real labor… Now, it's possible with one single prompt."

His favorite catch was a fraud ring of more than 5,000 accounts that all showed an 82% battery level. It traced back to a Chinese "device farm" running real, disassembled phones. His defense strategy: "Faking one signal is easy… But faking every dimension at once is still hard." He also said the web "is not prepared for the fluid of agentic sessions," and that businesses will soon need to sort traffic into good bots, bad bots, good agents and bad agents. His sharpest line was about incentives. Companies that treat new sign-ups as their key metric "are sort of like allowing a level of fraud… just to make those numbers as good as possible for shareholders." (Fintech Business Podcast, "Fighting Fraud in the Age of AI, with SEON's Tamas Kadar," Sep 30)

Visa: open-sourcing its AI bug-hunter

A senior Visa security executive described the company's scale: "17, 18 trillion dollars every year… a billion plus transactions" a day. Visa was one of about 50 companies given early access to Anthropic's Mythos Preview model through a program called Glasswing, which finds software vulnerabilities before attackers do. Within a week a tiger team built a nine-step "harness" around it to find, fix and verify bugs. Visa has now released that harness free to everyone, because "99.9% of our customers, our partners, and most companies on the planet don't have access to mythos."

The striking number: of all the flaws the AI found, only 0.03% were actually exploitable, thanks to Visa's layered defenses. Without those defenses, he estimated, it "would probably be at 12%." The warning: the time it takes attackers to exploit a newly found flaw has fallen from about 18 months eight years ago to under an hour by year-end. "You can no longer operate at human speed." He added that an Anthropic contact expects freely available open models to match Mythos within 18 months: "which means it's open season." (AI Security Podcast, "How Visa Secures Trillions Using AI Agents & Open-Source Harnesses," Oct 2)

Wealth management's AI moment

Claude for Financial Advisors, and the compliance headache it creates

Anthropic's new Claude for Financial Advisors connects directly to custody and portfolio systems such as Schwab and Orion. On Do Business. Do Life., host Brad Johnson's compliance lead (Shannon McGinnis) and tech lead (Quin Kilgore) walked through what an advisory firm must do before switching it on:

  • Treat every connector as its own vendor and ask "what doors does this MCP open?" (MCP is the plug-in standard that links an AI to other software.) Can it read data only, or change it too?
  • Pay for the enterprise tier for the audit logs. "Just because it's archiving doesn't mean you're covered… you're expected to supervise the logs."
  • Watch token costs, the usage-based fees for running the AI. Firms could see "$10,000, $20,000, $100,000 spend… in a month." As one put it, "don't take the Lamborghini to go get groceries."

Kilgore's risk point: Schwab and Orion have decades of tested security in this industry. "Schwab doesn't also sell shoes." Claude does everything, which makes it an attractive door for hackers. (Do Business. Do Life., "188: Shannon McGinnis & Quin Kilgore – Is Your Firm Ready for Claude for Financial Advisors?", Sep 30)

Morgan Stanley's (MS) research team supplied the big-picture numbers on Thoughts on the Market:

  • About half of a financial advisor's non-client time (meeting prep, notes, onboarding) could be automated, raising advisor capacity by 30–40%.
  • Leading firms could eventually gain up to 15 points of operating margin, though much of that will be competed away.
  • The key line for founders: "Saving someone, you know, 10 minutes on a task doesn't necessarily show up in the P&L. You need to start removing entire steps from workflows."
  • They don't see AI replacing advisors. Instead, "there could be a bull market for advice." (Thoughts on the Market, "How AI and Tokenization Could Reshape Wealth Management," Oct 1)

One Debate: Will agents really do the buying, and who pays when they get it wrong?

The week split into two camps, and both had real evidence.

The believers point to Instinct's numbers. They include $1 billion in transactions, 40% of users adding a card within three weeks and 80% of them sticking. Joseph Chalom, former head of digital assets at BlackRock, went further on The Rollup. He said agents will touch about $4 trillion in financial fees, things like idle cash in low-yield accounts and unharvested tax losses, and "much of that is going to go to zero." His reasoning: people miss their money goals because they don't pay attention, and "when you have an AI agent… [with] unlimited attention, they will do these things for you 24-7." Commissions on stock trades took decades to fall to zero. "This is going to happen in hyperspeed… quarters or maybe a year." (The Rollup, "Joseph Chalom: Ethereum Is The Toll Road To Everything (Larry Fink's Words)," Sep 29) On Onramp, the hosts added the scary version, a warning from Apollo's chief economist that agents could trigger a bank run as they move cash "out of something yielding… 0.1% [into] something that's yielding three, five, or more." (Onramp Bitcoin Media, "This Chart Exposed The AI Trade and Loves Bitcoin," Sep 29)

The skeptics had evidence too.

  • Chit Chat Stocks walked through an investor's test of Muse in Florence. All five hotels it suggested were Booking.com links. After 14 minutes of checking direct prices, "really more like one was cheaper. One was wash and three were more expensive." Co-host Brett Schafer's advice was to ignore company-reported numbers: "they're going to be selective… I would be curious what the retailers say." Ryan Henderson: "They're not booking hotels through these random AI agents yet. I'll put the yet there." (Chit Chat Stocks, "FICO and Nike Tumble; Burry Accelerates AI Short Bet; Compounders Out of Favor," Oct 2)
  • Ben Carman, a developer building payment tools for agents, said agent payments today are "kind of almost zero" and the whole thing is still a "theory of like if this is going to work or not." (Stephan Livera Podcast, "Lightning in x402 and Agent Payments | Ben Carman SLP779," Oct 3)

Then there's liability, which neither side has solved. A chargeback is when a card company reverses a purchase after a customer disputes it. Fraud expert Karisse Hendrick reported that Visa is in talks with AI companies about rules for agent purchases. For now, "there are no provisions for merchants to be able to win chargebacks when an agent is involved." Her view: "If the AI platform makes a mistake, I think the AI platform should be responsible for the chargeback. However, with current Visa and MasterCard rules… that's not the case and it won't be." (Fraudology Podcast, "AI Enabled Fraud: When Growth Comes Before Guardrails," Oct 1) SEON's Kadar said Amazon blocked Muse partly over exactly this fear, that a shopper will say "it wasn't me who prompted the agent to check out." Cashi's Wong raised the same question from the merchant's side: "Is the agent responsible? Who is going to pay for it?" And there were already horror stories: a Muse-booked flight that gave its traveler "a random middle name and I couldn't get on the flight."

The most useful framing came from Jack Altman on 20VC: "Does it make it from where we are right now to you can autonomously trust these things to run your life? And if it can, it's big. And if it can't, it's not."

For founders, the gap between those two outcomes is where the opportunity is. Some of what the agents need doesn't exist yet: identity checks for agents, bank-side "firewalls," dispute rules and fraud scoring for agent traffic. Whoever builds those gets paid either way. Whether consumers ever hand their whole wallet to an agent is still open. The podcasts this week were confident the agents are here, and not at all sure who carries the risk when they make mistakes.