🧠 Meet the ghost card

AI brings zero new capabilities to fraud detection.

But it does change how much of our platform you can actually use. No human has time to read every doc, learn every data schema, or wire up every feature.

Agents do. They have infinite capacity for that work. Sardine’s AI Agents open up our entire platform, so everyone is a power user. The payoff is fraud prevention at machine speed, tighter compliance controls, more approved transactions, and support for more payment rails.

Everyone’s asking what new fraud detection capabilities AI brings to the table. But the better question: is your fraud team working at its full potential?

Unlock Agentic Fraud Ops with Sardine.

Here's this week's Brainfood in summary

👀 Things to Know:

💸 4 Companies:

Content Corner:

1. Meta's Muse hit 10m US downloads in 40 days, a pace only 11 apps have matched since 2015

Sensor Tower puts Muse at 10.2m US downloads between its Sept 8 launch and Oct 17, with 2.5m US daily active users on Oct 16, up 15% week-on-week. ChatGPT did 3.7m US downloads in its first 40 days and took 123 days to reach 10m (Grok took 211 days, Claude 673).

🧠 Consumer adoption is here. Only 11 other apps have reached 10m US downloads in 40 days since Sensor Tower started counting in 2015; over 80% were mobile games, and the last was Threads in July 2023. ChatGPT's US downloads fell 25% over the last 30 days while Muse held number one on both app stores.

🧠 The business model is commerce. Mark Zuckerberg stated on the Meta earnings call that shopping is the most common use case for Muse, and that the average Muse transaction size is $200.

Read that last point again.

🧠 Meta is spending big to get consumer adoption. Muse took 39% of Meta's own advertising budget this month, up from 23% in September, and climbed from the 73rd to the 21st biggest US brand by ad impressions. That's the rational play when you own Facebook, Instagram and WhatsApp and already sit in the top ten US advertisers.

🧠 Muse is also sticky with existing users. 30% of people who installed in September were still using Muse after 7 days and 23% after 14, against 19% and 17% for ChatGPT, Grok and Claude. Day-one retention was 73%, so the ads pull in plenty of tourists, but Muse still keeps more of them than any of the three.

🧠The form factor of an app-based chat that's free is absolutely winning over the $20/m subscription. The vast majority of ChatGPT's best features are not used by 90% of its users. Muse gives away 100m tokens a week, charges only the power users, and does the task for you.

🧠 The payments industry is scrambling to react to Muse. Agentic commerce had the protocols, the wallets, and the insurance and was waiting on demand, and 2.5m daily users is the demand. I expect security, identity, and liability for agent payments to get solved within a year. I think all of the banks know they can’t block agents forever, but they can afford to wait.

🧠 There’s a lot to learn here if you’re building your own customer agent. Muse Wonder is a frontier model trained specifically for the agent. Meta shipped Wonder and Wonder Full last week, and Full scores 56 on the Artificial Analysis Intelligence Index, two points behind Claude Opus 5.5 and ahead of GPT-6 Astra. Fourteen months of rebuilt pretraining and RL under Alexandr Wang went into a model built for the tasks Muse does.

🧠 Anthropic also released its Muse competitor Duet this week. It’s a research preview and “waitlist” for pro and max subscribers, meaning not free like Muse. I wonder if that will harm adoption?

🧠 And you guessed it, Stripe has integrated day 0. The link wallet can hold your cards and manage checkout, and your small business can manage refunds, etc via the main Stripe MCP.

2. Ramp applies to the OCC for a national trust bank charter and bought Vesto

Ramp filed on 16 October to establish Ramp National Trust Bank, with powers for digital asset custody and fiduciary investment management. It will take no deposits and make no loans. The filing came one day after Ramp confirmed its $1.85B raise at a $60B valuation, with annualized revenue past $2B and 85,000 customers. Then on Tuesday Ramp bought Vesto, a treasury platform that pulls a company's cash from 12,000 banks into one view, with direct links to JPMorgan Access, BofA CashPro, CitiDirect and Wells Fargo Vantage.

🧠 Ramp is becoming a stablecoin fiduciary not a “bank.” This is about stablecoins. Ramp runs stablecoin accounts, stablecoin bill pay, and stablecoin-backed cards on Stripe's Bridge and Privy. A charter lets Ramp hold those balances itself.

🧠 Fiduciary investment management brings in much more treasury capability. Ramp holds more than $5B in customer checking and investment balances, with the investment side custodied at Apex Clearing today. The charter brings that in-house and puts Ramp in the running for tokenized money market funds too.

🧠 The Vesto deal makes them a CFO cash management play. Ramp says its average customer runs more than three bank accounts. Vesto shows the CFO where the cash sits, and stablecoins & tokenized money market funds then become a direct way to manage that on platform.

🧠 See this much more like Circle and Bridge than Mercury. Circle, Ripple, Paxos, BitGo, and Fidelity Digital Assets got conditional approvals for a trust charter in December. Owning costs less than paying a third party, and it puts the fund flow under Ramp's control.

🧠 I’m so interested in this model of aggregating treasury above the banks while also supporting tokenized assets directly. CFO’s still need fiat at banks as they scale. But often, that is done in complex treasury management software. Ramp can eat that and start to win some of the assets on platform too as stablecoins or MMFs.

3. Mercury crossed $1bn in annualized revenue, up ~50% YoY

CEO Immad Akhund told the Newcomer podcast Mercury now runs at $1bn annualized revenue, up around 50% YoY, with 450K+ customers, four years of profit and, by his count, one in three US startups.

🧠 Revenue is accelerating: Mercury went from $500M in ‘24 rev to a $650M run-rate in Sep ‘25, and is now growing ~50% YoY (Per JC Stefano)

🧠 Customer growth keeps accelerating: +40% in ‘24, +50% in ‘25, then 300K to 450K+ in under 10 months. Although some of this is helped by the average customer size getting smaller / going more mass market.

🧠 The $1bn came before the software started charging. Books and Payroll are free until January, so this revenue is deposits, cards, and treasury, and Sacra puts most of it on interest from roughly $20bn of deposits. Subscription billing starts on 1 Jan 2027 at $35 a month for Books and $100 a month plus $50 a seat for Payroll.

🧠 The product velocity here is astonishing. In the last five weeks, it launched Books (double-entry accounting), launched Payroll today (W-2, 1099, benefits, all 50 states), and let agents move money from ChatGPT or Claude through its MCP and CLI, with a human tap to approve each transfer.

🧠 I’ve been asked three times this week which banks let agents do things for consumers? Jamie Dimon blocked outside agents on 13 Oct, citing "rogue agents"; Mercury launched agent payments on 15 Oct, including on Mercury personal.

🧠 The launch velocity speaks to a company accelerating, in part, because of AI. Books on 16 Sept, agent payments on 15 Oct, Payroll on 20 Oct. Mercury bought part of that speed (Central, acquired in April, became Payroll) and built the rest.

🧠 Curious to see if they can keep this feature velocity when they become a bank. Being a bank hasn’t slowed Revolut or Nubank down. Also love that it’s such a pure play to become an actual bank rather than a bank in all but name like the trust charters. Neither approach is wrong, but as an observer, we get to see how both paths play out.

4. Visa and Mesh are making stablecoin balances spendable at 175m merchants

Consumers will be able to pay in-store and online directly from the crypto, stablecoin or fiat balances they already hold in their wallet apps. Any platform that integrates Mesh Pay can offer it, and Mesh names Kalshi, Félix Pago, MAJORITY and KAST as partners (which of them goes first is still TBC). The MESH SDK is also in apps like Coinbase and Polymarket.

🧠 Meet the ghost card. Mesh issues a "ghost card" (a Visa credential the user never sees) in the background, tagged to the KYC of the wallet holder. You tap to pay from your wallet balance, and the merchant gets a regular Visa payment.

🧠 Platforms with lots of users can give them card spending without ever launching a card program. A card or virtual card as a form factor became another thing to sign up for, manage, or add to your Apple Pay. This just works from your wallet and is much easier for the wallet owner to package.

🧠 MESH is a company that’s everywhere, and nobody knows what it does. Their core product is a network that makes crypto interoperable across 300 wallets and blockchain networks. As a result, they have an impressive client list of wallets that now have an alternative to launching a card program. I'm curious to see if this product from MESH gets traction and competes meaningfully with Rain or Wirex.

5. The OCC finalised its GENIUS Act stablecoin rules

The OCC published its final rule implementing the GENIUS Act today, the first federal regulator to finish its part. It takes effect January 18, 2027, the same day as the Act. The issuer yield ban stays, however the proposed “presumption against third-party rewards” is gone. Meaning these rules are likely to continue the status quo for a while.

🧠 A final rule is hard to undo. An executive order can't easily reverse one that implements an act of Congress, so this is the framework for years, whoever sits at the OCC.

🧠 The yield ban stays, and it covers every holder. The OCC refused a split between retail and institutional holders because the statute makes none.

🧠 The banks wanted the rebuttable presumption but it’s gone. It would have assumed any yield paid by an affiliate, a yield-as-a-service provider or a white-label partner came from the stablecoin issuer, and made the issuer prove otherwise. The OCC dropped it, called the market nascent, and will judge each arrangement case by case. That leaves a lot of room for discretion, for this OCC and the ones that follow it.

🧠 No safe harbors either. Existing rewards programs, and white-label partners all get case-by-case review. The OCC flagged white-label partners as the ones most likely to be paying on the issuer's behalf, so it will look at those hardest.

🧠 White-label stablecoin issuance is permitted all the same. White-label issuing is where an issuer (e.g. Paxos) creates a branded stablecoin for distributors (like PYUSD). The rule writes white-label distributors into its definitions.

🧠 The OCC said in March an issuer can share profits with a white-label partner, and the final text says payments for real services are generally fine. So picture Klarna, Western Union or MoneyGram with a branded dollar minted by Bridge, Paxos or M0. The brand collects its cut of reserve revenue. Whether it can pass that cut to holders depends on what the payment is for.

🧠 If the brand just forks over 4% on balances, that looks like interest where someone struck out the word and wrote "rewards" over it. Points and cashback are the easier argument, and we haven't seen much of that yet.

🧠 The banks fought the CLARITY Act so hard that it ultimately stalled. In hindsight, I wonder if that was a mistake, given how this rule is playing out.

🧠 The rest of the rule kept what the proposal set out. Issuers must hold reserves at least equal to the coins they have issued, redeem within two business days, hold at least $5m in capital, and tell the OCC within 24 hours if redemptions top 10% of a brand's float in a day. Existing issuers with an application pending can get a waiver of up to 12 months.

🧠 All told this looks like the status quo. GENIUS bans yield, and rewards and white-label pass-through are very much alive and well. We’re some ways off seeing where this all stands because applications haven’t started yet.

While we’re considering stablecoins.

6. Tala brings unsecured stablecoin credit to El Dorado's wallet in Bolivia

Tala will underwrite loans inside El Dorado, the USDT wallet with 1m+ users across 13 Latin American countries. Customers see a pre-qualified offer in the wallet and receive USDT in seconds, with loans disbursed and repaid on Tempo. Bolivia launches first, with more Latin American countries to follow.

🧠 Unsecured credit has been the unsolved problem of on-chain finance. DeFi lending runs over-collateralized, where borrowing one BTC can mean locking up two. Under-collateralized lending, funded by deposits, is how most credit in the world works, and it is a far better business for the lender.

🧠Tala didn’t come from DeFi; it came from lending to frontier markets. Tala’s secret sauce is underwriting from signals on a consumer's phone: how they behave, transact, and use mobile money. That business has distributed more than $10bn to 14m+ customers across Africa, Latin America, and Asia.

🧠 The stablecoin wallet is a distribution channel for that engine. Tala brings the capital and carries the credit risk; El Dorado brings the customers and the relationship. Any wallet can plug in without building a lending business of its own.

🧠 Stablecoin wallets have found ultra product-market fit in markets like Bolivia. More than 8 in 10 workers are informally employed and consumer credit is 10% of bank lending, so banks can't see most borrowers. A wallet that watches how someone saves and moves USDT every day can.

🧠 This completes the stack for non-bank wallets. A customer can now store stablecoins, earn rewards, invest, and borrow, all in one wallet, on-chain. That is a full financial ecosystem running in production in Latin America.

🧠 I met Tala founder Shivani Siroya in Vegas at Money20/20 this week. What they’ve built in terms of underwriting tech and their partnerships in the stablecoin space are exceptional. You can hear more from Shivani on this episode of Tokenized.

had the pleasure of judging these at this year’s Commerce Ventures America’s Got Access competition.

1. Beltic — Agent identity and permissions for businesses

Beltic helps merchants, marketplaces, and financial apps get visibility into their agent traffic, and set rules for what those agents are allowed to do on their site/platform. From simple blocks to writing specific policies for known agents on what actions and resources they can take. Beltic also aims to show the intent behind the agent's action so a merchant or platform can attempt to match that to something that makes sense for a user.

🧠 Most companies can’t tell genuine agent traffic from bad bots; that's partly why they’re blocking them. Cloudflare already verifies bot and agent traffic, Beltic adds the ability to decide what they can do. If you stick this with the recently announced Personal Agent Protocol (PAP), which is a way for agents to formally request what they want, and you have something fairly complete. I feel like this should be a feature in every major commerce platform (Shopify, Woo, BigCommerce, etc.).

2. Cache — Embedded agents for banks for consumer cashflow.

Cache partners with banks to embed an agent in their mobile and online banking that automatically moves money between your existing checking and savings accounts. The money stays in your accounts, so you don't need a new bank account. The service is also available standalone (direct to consumer).

🧠 Smart distribution strategy. Not all customers will be comfortable letting Muse move their money. We’re super early in the adoption cycle, so banks being able to offer a competing experience for the mass market makes sense. I think the calculation here is banks would rather have the customer relationship than lose it to a third party agent, even if they end up not profiting as much from inertia.

3. DenZen — Layoff protection for renters and homeowners

DenZen offers protection that pays your rent or mortgage directly to your landlord or lender if you lose your job and have an eligible claim. It offers coverage through selected property managers, landlords, and lenders, built into the process of signing a lease or taking out a loan.

🧠 A layoff can become a housing problem very quickly. The landlord or lender has a reason to want protection too, because their customer’s lost income can become their missed payment. That gives DenZen a sensible distribution channel and a moment to offer it: when someone commits to a large monthly housing bill. Keeping people in their homes is a pretty useful financial product.

4. Kiro Money — A financial coach inside your banking app

Kiro lets banks, credit unions, and fintechs put an AI financial coach inside their own app. Customers can ask questions such as “Can I afford to save more?” and get guidance using their balances, spending, and goals. It aims to see customer intent early and improve the chances of cross sell for its customers.

🧠 This is the chatbot for credit unions. I can imagine credit unions wanting this, especially if the assistant is as helpful telling a member to keep their money as it is suggesting another product. Both this and Cache find themselves competing with bigger incumbents like Personetics, which now offer similar capabilities, but sometimes there’s power in a clarity of message from a startup who does just that thing. Kiro also aims to get live super fast (with case studies of less than two weeks).

1. Companies are using FDEs because they don’t have an AI-friendly knowledge base.

The authors led FDEs for Palantir. FDEs are so crucial because most companies don’t know their own processes. They describe a new company capability, “ContextOps” (think DevOps or BizOps). Whoever owns the process designs it, but AI continually runs it, flags exceptions, and helps adjust it. The goal is to increase the “executable coverage” (the number of tasks an agent can execute without support).

🧠 A lot of companies are building agents. Few are building learning loops around those agents. Well worth a read.

2. Brand as Software.

Every brand is dying by 1000 cuts. The design team creates templates but ends up building a queue for people to form into, because their capacity was constrained. AI can now create assets much faster, but brand guidelines aren’t built to move that fast. It helps constantly remember (and build context) of where a brand team changed assets so it can learn to do so itself. So now at Ramp, instead of contacting the brand team, you @ the design agent in slack, and it carries the brand’s consistency based on the continual learning its doing with the team and humans.

Remember, if you're enjoying this content, please do tell all your fintech friends to check it out and hit the subscribe button :)

Want more? I also run the Tokenized podcast and newsletter.

(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