Welcome to Fintech Brainfood, the weekly deep dive into Fintech news, events, and analysis. You can subscribe by hitting the button below, and you can get in touch by hitting reply to the email (or subscribing then replying)

Hey Fintech Nerds 👋

I took two weeks off, and the GENIUS Act is now law, JPM and Citi talked stablecoin strategy. Revolut may be raising $1bn at $65bn (not far from Stripe!). Circle filled with the OCC to be a trust bank. JPM decided to hike its API fees (the dark side is strong). ChatGPT Agents launched. Anthropic launched financial services. Perplexity launched Comet, and the whole Windsurf, OpenAI, Google, Devin thing happened. 

It feels like we’re in one of those phases where decades are happening in minutes. For more on stablecoins, see “Things to Know” this week.

This week’s Rant consolidates all my ideas on agentic commerce into one place. There’s a lot in there, maybe grab a large coffee for this one.

🎧 Stripe broke the internet when they spent $1.1bn for Bridge. But this was a small part of a bigger master plan. In this Exclusive Tokenized Interview, President of Stripe Will Gaybrick takes us behind the scenes of that, and much more.

Weekly Rant 📣

Apps Are Dying. Agents Are Buying.

Why agentic commerce will reshape payments, merchants, and everything else

Imagine needing to find late-night sushi in Soho. Instead of opening Maps, then Yelp, then OpenTable, then my credit card app, I’d ask my LLM one question. It gives me a map, restaurant options, and lets me book a table. Three apps became one conversation.

The AI-first browser isn’t the new app. It’s not browsing web pages in the traditional sense. It’s the personal assistant baked into the internet. 

This is how a platform shift begins. 

Perplexity's Comet browser is a hint at what this could look like

Three apps became one for the Sushi example. With Comet, anything on the internet can become a single conversation. Instead of separate apps or pages for maps, reservations, and payments. 

We’re still early. We’re still finding our feet.

ChatGPT Agent is another model that spins up a virtual computer for agents to use. Will Agents have their own computer, or will they live in the browser? I don’t know.

But one thing I know for sure: Agentic commerce won’t resemble e-commerce or mobile commerce. The experience of shopping itself may become more invisible. What once required multiple app screens or web pages will soon be a single, fluid conversation.

This requires two fundamental shifts:.

1. AI will dissolve apps into “applets” Apps once felt like destinations. Now, every calendar, email, or map is just an on-demand assistant: summoned for a moment, shaped by the question, and dismissed without a trace. Today it’s via chat and a browser. Tomorrow, the UI’s could be assembled dynamically.

2. Merchants websites will dissolve into Model Context Protocol (MCP) (or something better) making pricing, SKUs and payment available to AI Agents and LLMs.

Here’s how it plays out

1. AI is turning Apps into Applets or Liquid Software

2. The checkout is becoming invisible creating two new problems

3. The bear case - we’re too early

4. The bull case - we’re 100x bigger than mobile commerce

5. The three phases of agentic commerce

6. The security unlock’s we need to get there

7. And life after the checkout

(Some background reading if you’re new to agentic commerce, the four models of agentic commerce and the checkout is dead).

1. AI is turning Apps into Applets or ”Liquid software.”

AI is a platform. Every app on your phone will disappear into your AI agent experience. 

This reduces the cognitive load on the user and is more helpful than searching for numerous restaurants near a location or clicking through top 10 lists. 

Today that’s: 

• Going to an LLM and querying it, and having it search the internet. 

• Or it's websites baking an LLM into their existing e-commerce or app experience.

Tomorrow that becomes: 

• Agents see what you see: Chat and voice baked in to the internet: Your LLM or agent can see what you see. Seeing via the browser (like Comet) is just the first step. Next up. Wearables.

• Liquid software: AI creates app-like experiences on demand, then discards them because it has the ability to quickly create new code and UI’s.

The combination of seeing all of your context and liquid software building an experience for just this moment and throwing it away is what makes AI the platform.

It’s less app store. And more of all the apps in one. Except in this world, context is shared between them

If AI becomes the platform, commerce faces a two-sided challenge.

2. Checkouts becoming invisible creates two problems.

When you search for products in ChatGPT or Perplexity, you can also purchase directly in that experience. 

The new user flow is quite different. The AI Agent or LLM is collecting SKUs, possibly from multiple merchants. It already has the shipping information, and may already have a payment method (like a card on file).

This creates a two-sided challenge:

1. For agents - how do I pay?

2. For merchants - how do I show up in the agents?

How will agents pay? There are four models emerging: 

As of now each of these models is experimental, limited in volume, and still emerging. We don’t know the security standards we’ll use or how this all interacts with existing wallets, cards, and payment rails. 

Every major payments company, and AI lab is racing toward agentic commerce. Visa's Intelligent Commerce approach shows how this will work:

Daydream is an interesting early example of agent-first commerce. As the New Consumer explains

This experience could be an upgrade for those of us overwhelmed by choice paralysis. It’s not a default for us yet, but neither was using a mobile phone. We also don’t know how to make this secure or what the default payment method will be.

These problems are solvable. The question is, who solves them first?

Merchants are reaching out and asking me, “What should we do?” And there are no easy answers; there are just good domains to explore. 

How will Merchant appear to Agents?

When websites disappear into conversations, merchants face a discovery problem:.

• Merchants who spent decades optimizing for Google's algorithm now need to optimize for Claude's reasoning

• SEO becomes AEO (Agent Experience Optimization)

• The race to become the default recommendation agents present

Merchants can't just buy ads anymore, they need to become the right answer. But the right answer for who and in what context? And who gets to decide what’s right? 

When someone asks ChatGPT "where should I travel with a young family," they're not looking for your SEO ranking top 10 list. They want trade-offs, considerations, and context. The agent that gives comprehensive, nuanced answers wins the recommendation.

This means merchants need to optimize for AI reasoning, not human browsing. Instead of SEO-optimized "Best Family Destinations" pages, you need content that helps agents understand when and why to recommend you.

MCP Services become the new store fronts.

Instead of landing on a web page agents land on something more optimized for them. (For brick and mortar merchants it would be another “new channel”).

The Model Context Protocol (MCP) server concept, innovated by Anthropic and becoming and increasingly common way to make all kinds of 3rd party tools available to LLMs and agents. Perhaps merchants can and will use something like this? 

From “🧠 Model Context Protocol - Flash for AI”

To visualize this look at Ogment AI, which “makes your products shoppable on ChatGPT, Perplexity and Deep Seek.” It helps merchants create an MCP server that can handle checkout, conversion analytics, price customization, and customer service. It’s sort of “Shopify for MCP servers.”

In the future, merchants would have mcp.yourdomain.com and then a product like Ogment handles the super complex set of prompts and custom code to help manage all of their pricing, service, and fulfillment needs.

So instead of running a storefront, merchants will run a store engine. The presentation and visualization of products could be rendered to the customer by different LLMs or agents, but the pricing, images, returns policy, and perhaps even your fraud rules could be delivered to the agent via an MCP server.

MCP Discovery is unsolved. How does the buyer agent discover and use the store? The idea of MCP is instructive, even if the frameworks and ideas are early.

There’s also the possibility that agents themselves will become entire businesses. 

Anthropic recently created a vending machine that was run entirely by AI. It could manage inventory, customer service and collect payments. And while it hallucinated items, and gave too many freebies to be profitable, and even went into selling tungsten cubes… this was an early experiment. (See this weeks Good Reads for where this could be heading). 

This platform shift seems inevitable. But are we too early?

3. The bear case - We’re too early

Are we too early?

Before Apple Pay or mainstream e-commerce, companies like Nokia were attempting to integrate credit cards into mobile phones in the late 1990s. 

Mobile payments had many false starts.

Nokia, Sony, and BlackBerry all built mobile web capabilities and attempted to create commerce experiences. These “feature phones” largely existed before 3G made mobile internet fast enough for everyday use. So they used an online protocol called “WAP.” I worked in payments at the time and saw so many companies demoing their “mobile commerce” experiences and innovations.

Why does this matter?

Are we in the Nokia WAP era (right idea, wrong timing) or the iPhone era (platform shift complete)?

Without real volume it's impossible to tell. 

Social commerce never really happened.

More recently, we’ve also seen social commerce come and go. In the 2010s, “social media” like Facebook was also going to become a shopping destination. In 2025, it’s not working. Meta eliminated its checkout on Instagram in the past few weeks, and TikTok is restructuring its shop. 

Social commerce failed because embedded checkout added no value - it was easier to go from ad to page.

Agentic commerce has no real volume yet. It could go the way of social commerce.

Every new payment type usually comes with new security and fraud risks.

While I’d argue mobile payments did better, the early days of e-commerce were the wild west for fraudsters. We’d moved from a world where someone had to spoof your card, or clone it to make an in-person transaction, to one where all they needed was your PAN (16 digit) number and expiry.

Agentic commerce could have 100x this risk. 

AI Agents are trivially easy to prompt-inject and steal data from:

AI agents with data access, exposure to untrusted content, and communication abilities create massive attack surfaces. They’re vulnerable to prompt injection. We’re already seeing bad actors use white text on white backgrounds to fool AI Agents in spam and email clients.

And MCP is full of security holes: 

For example, a single MCP server for Supabase - could leak your entire SQL database, and CATO networks demonstrated how the Atlasian MCP server can be attacked to steal data too. 

Agentic commerce opens entirely new attack vectors. We're only beginning to understand how to secure them. There’s also the fraud and AML risks to consider. In a conversation with Jeff Weinstein from Stripe a few months ago he verbalized the idea of the 2x2 of new fraud actor risks we have.

• A bad human could perform an agent takeover (ATO) to use your agent to buy things with your credentials.

• A bad human could also use compromised agents from the dark web or elsewhere that appear to be good agents to merchants to perform attacks

• A good human could install or use a bad agent that appears to be good (in much the same way fraudsters create imposter websites or phishing attacks today).

• A good human and good agent would need some way of consistently verifying / authenticating the link between human and agent and the agents reputation.

Agentic commerce clearly faces a boatload of new risks, but generally we unlock new commerce experiences as we start to build new security models. They’re not a barrier to scale, they’re the way we unlock it.

But the bull case is compelling.

4. The bull case - Agentic Checkout becomes much bigger than mobile

eCommerce will do $6.8 trillion of sales in 2025, and is growing at 8% CAGR.

Agentic commerce will capture some of that. 

It will also grow the overall figure.

Agentic commerce could be 10x bigger by removing the friction of search, purchase, and delivery management.

Agents can make previously fragmented experiences more coherent.

Social is primarily a communication and entertainment destination. Search is search. These experiences are disconnected and fragmented.

LLMs and agents are where we get work done, search, and they’re becoming a place where software can mould around your conversation and handle complex workflows.

When your AI handles everything from calendar scheduling to expense reports, adding purchase decisions isn't a stretch. If anything it feels inevitable.

Agentic commerce will make the experience of shopping better / invisible.

Things like the weekly food order can still be time consuming, there’s so much opportunity for a personal chef AI, personal shoppers rotating your wardrobe, the personal assistant who helps you with supplements or nutrition.

The incentives are uniquely aligned to solve agentic commerce.

• AI labs have all of the funding: Open AI has raised nearly $58bn, Anthropic $18.2bn. We’ve never seen newcomes this well equipped to become platforms.

• Big tech is spending to stay in the game: Google, Meta, Microsoft et al are all jockeying for a piece of the new pie.

• Payments companies have all of the incentive: Payments companies grow when total processed volume (TPV) grows. If agentic commerce is a catalyst for more payments they’re going to aggressively lean in.

The reason you see companies like PayPal, Stripe and Visa trying to do something in agentic commerce is because nobody wants to miss the platform shift, and the best way to learn is by doing. 

The scale of experimentation here is staggering. OpenAI launched AI agents to 200 million users on 17th July. Unlike browser-based agents, OpenAI's approach spins up virtual machines - giving agents their own device, files, and access to your calendar and emails. A lot could go wrong here.

The internet itself will adapt around agents.

Cloudflare just launched “pay per crawl,” the idea is if they detect a bot, that bot can pay the content owner (e.g. the New York Times). It resurrects HTTP 402 - a forgotten web standard for payments:

We might just get micropayments on the internet yet.

Agentic commerce will unlock new payment flows that don’t happen today.

Agentic commerce makes micropayments a reality. The move by cloudflare would bake this capability into 20% of internet traffic. We historically don’t do micropayments on the internet, because a single payment often costs at least $0.20. Stablecoins haven’t made that cheaper and lack broad adoption. 

Today you either hit a paywall and subscribe, or in 95% of cases you don’t. That’s a missed opportunity. If those were payments, that would be new economic opportunity unlocked.

With HTTP 402, we have the beginnings of a way for agents to pay. If that gets expanded upon agent-native payments could unlock new use cases like: 

• Multi merchant orchestration: If you buy a laptop from best buy, cable from amazon, and a setup from geek squad that’s three transactions and experiences.

• Buy when x happens: Set a rule like buy Pepsi Max from whomever has it at the lowest price, or buy Jeff’s favorite coffee if it drops below $12.

• Agent to Agent commerce: Your vacation agent negotiates with various booking sites, travel agents and entertainment venues to book an entire break.

Security problems are solvable. 

We’ll seen a pattern play out where the card networks, banks and other payments ecosystem participants gradually coalesce around a new security standard. That could be x402, Visa Intelligent Commerce, MCP, Google’s A2A standard, HTTP 402, w3c verifiable credentials, or something else entirely. 

And solving them is usually what unlocks the next era of commerce. 

My sense is the way we will do that is by embedding wallets and security closer to the chat / applet UI and experience.

Security is the unlock.

5. The Three Phases of Agentic Commerce

1. Make the chatbot the checkout: Humans paying via AI interfaces (e.g. embedded in ChatGPT)

2. Take the Agent to the web page: Agents paying on human interfaces (e.g. existing checkout via Comet)

3. Make merchants agent-ready: Agents paying on agent-friendly interfaces (e.g. MCP). 

The tweet also posits several complexities like - a human doing 10x $2 transactions in 10 seconds is likely fraud, an agent doing it could be legit. New commerce flows like micro-transactions or rapid transactions will impact every downstream payments processor, bank and non-bank. There’s work to do.

• Step 1 is already happening, with Shopify embedding in ChatGPT.

• Step 2 has started with Perplxity Comet or experiences like Daydream

• Step 3 requires a security model. 

That’s the next frontier.

Imagine a “Perplexity wallet” baked into the Comet browser similar to Google Wallet, that makes payments simple, secure and streamlined. The merchant “server” would recognize the agent through HTTP 402 (or similar) and initiate a secure handshake (through intelligent commerce or similar).

A significant aspect of what a physical credit card does is to make payments more secure. It has chips, numbers and references back to your real identity.

We don’t have that pattern for agentic commerce.

As I wrote in the wallet wars pt 2, your AI Agent needs three things to be useful online:

So much of our focus in agentic commerce is on the payment; it’s not on the consumer side of authentication. If you spend any time in payments, you quickly find yourself dragged into conversations about EMVco standards, and the idiosyncrasies of ISO8583 messaging from different banks and processors.

That might sound nerdy. But there’s a point to it.

All of that technical jargon is making a security and trust promise. What makes payments work is that we have a consistent experience for the happy path, as well as the billions of possible unhappy paths and edge cases.

The AI Labs, agents, and tech companies are probably already starting to think this way. 

6. Life after the checkout

When mobile payments finally happened it was the combination of

1. The broad availability of NFC chips in smartphones

2. Complex agreements between Apple and Card Issuers in the early 2010s

3. Clear technical standards that had been built over decades

4. Support from the card networks, issuers, acquirers and whole ecosystem

The equivalent for agents doesn’t exist. Yet. But we’re standing on the shoulders of lots of bits of technology that have existed for a while. HTTP 402, w3c’s verifiable credentials and soon, the card networks will add their piece to the puzzle.

When I can turn 3 apps into one conversation, I’m a lot more likely to have a good Sushi dinner tonight than living off three protein bars and calling it a night. 

That experience is so compelling it leads me to believe it is an inevitable outcome. 

Three apps became one conversation for my sushi dinner. Soon, three separate shopping experiences will become one conversation for everything else.

When that happens:

• Merchants who optimize for AI reasoning will win over those optimizing for human browsing

• Payment companies that enable agent-to-agent transactions will capture more volume than those built for human-to-human

• The platforms that can authenticate both humans and their agents will own the next layer of commerce infrastructure

This isn't a question of if - it's a question of when and who.

• For merchants: Start thinking about how agents will discover and interact with your products. SEO is becoming AEO whether you're ready or not.

• For payment companies: The next wave of growth comes from enabling new transaction types, not optimizing existing ones.

• For developers: MCP servers are the new APIs. Security models are the new UX.

The checkout page died when commerce moved from destination to conversation. What comes next is commerce that's invisible, intelligent, and everywhere.

That's life after the checkout.

ST.

4 Fintech Companies 💸

1. Ogment AI - The MCP Server for Commerce

Ogment helps brands make their products shoppable on ChatGPT, Claude, Perplexity and DeepSeek. It aims to help them take more control of the customer conversation, increase customer conversion, track queries, intent and analytics. Merchants get to keep real-time customer data, and deliver it securely from their existing back end to the customer. 

🧠This is Shopify for the MCP Era. The clarity here is stunning. There’s a lose agreement that merchants need their own MCP servers. What Ogment has done is define how those servers work and interact with existing back-office systems (like inventory, customer service, and shipping rules). 

2. Checker Finance - The liquidity network for stablecoins

Checker builds “liquidity pathways” between 75 currencies with over 50 providers. This connects OTC desks, banks, stablecoin middleware platforms, PSPs and lenders into a single API. Developers can make a simple API request, and the platform will route the order for “best execution.”

🧠 A key bottleneck to stablecoin scaling is liquidity. Even with stablecoins you have to setup a complex treasury function to solve the liquidity bottleneck. Checker aims to change that by connecting it all with a single API. It reminds me of platforms in traditional finance like t360, where multiple large banks and brokers can publish their currency prices and take orders.  

3. Ellia - AI Analysts for PE 

Eillia has a aries of “analysts” that do company research, valuations, market research and company scouting. It presents this information in a dashboard of deal intelligence, and integrates with proprietary data sources like Capital IQ, Crunchbase and LinkedIn. They have clients like Fuel ventures using this in production.

🧠There are so many of these. Did every class of ‘24 intern rage quite and start an AI assistant company? Also what happens now Cloudflare made 20% of the internet impossible for bots to crawl? If anything the moat appears to be exceptionally good prompt engineering, evals and packaging that into a product. And that sounds simple in theory, but in practice it’s very challenging. That’s your IP. And doing it well shows.

4. Aethereum - Insurance and Loans as a service for digital assets.

Companies, funds and institutions engaging in onchain finance can use aethereum to store digital assets, take out loans and manage risk through the automated insurance product. The services uses AI to underwrite a crypto asset portfolio, provide loans and manage all customer KYC/KYB and repayments. The idea is fintech or banking customers can embed this in their existing offering.

🧠If an increasing number of customers have stablecoins, how many would want stablecoin backed lending? My guess is, in the global south quite a lot. This seems predicated on ensuring the assets live in the aethereum wallet in order to underwrite them fully. Meaning a user of say, Ramp (a Bridge customer), might have to transfer their stablecoin collateral, into the “loan collateral wallet” before they could have their loan underwritten. Which makes me wonder what the moat is here. Why couldn’t bridge build this?

Things to know 👀

1. JPM and Citi Talk Stablecoins strategies

From CNBC: “JPMorgan Chase CEO Jamie Dimon says he doesn’t get the appeal of stablecoins, but he also can’t afford to stay on the sidelines.” Failing to do so could cede ground to fintech players who are looking to recreate elements of the regulated financial ecosystem.

From Yahoo Finance: "We are looking at the issuance of a Citi stablecoin, but probably most importantly is the tokenized deposit space, where we're very active," says Jane Fraser.

🧠 Analysts want to hear about stables given the meteoric rise for $CRCL ( ▲ 2.07% ). Every bank CEO is being asked about them. During the same week GENIUS is heading to a final vote and the CLARITY Act is in the Senate.

🧠 Here's what most people missed: Tokenized deposits become the settlement layer for everything. Stablecoins, on-chain FX, and Tokenized securities. Everything. They will be much bigger than the current closed loops.

🧠 Kinexsys and Citi Token services help existing clients move funds to other Citi or JPM markets, instantly, 24/7. But they don’t go everywhere. In some exotics, stablecoins have better unit economics.

🧠 Stablecoins are instant and 24/7, but often get stuck with poor on / off-ramp liquidity. Tokenized deposits fix this. They use the same networks and infrastructure as stanblecoins. This makes them available for instant, 24/7, onchain FX. Swap one JPMD dollar for 1 USDC.

🧠 Tokenized deposits are currently only available for existing bank clients, but that’s not the point. The point is making stablecoin, long-tail markets available to institutions who want it. And in turn, solving the liquidity bottleneck stablecoins had.

🧠 The Regulatory Unlock is key: "This is where the Genius Act is also something that we are enthusiastic about, particularly because it gives a level playing field as well." says Fraser. That term “level playing field” is code for “we think we can compete.”

2. Anthropic launches Claude for Financial Services

Claude maker Anthropic helps pull together market data from external sources and internal warehouses like Snowflake and Databricks, to make them available to analysts. They’re pairing Claude Code (the software engineering suite) with pre-built connectors to internal and external data sources as well as its Claude 4 models.

🧠 The case studies give a lot of signal:

• AIG compressed underwriting reviews by 5x. Accuracy jumped from 75% to 90%.

• NBIM saved 213,000 hours—that's 20% productivity gains.

• Bridgewater streamlined analyst workflows.

• Jump Trading calls Claude "exactly what's been missing."

🧠 This could commoditize an entire category of Fintech companies. There are 100s of YCombinator companies that do some of these use cases with similar results. If your customers can get Claude plus your data sources directly, what exactly are you selling?

🧠 Anthropic is positioned to dominate the enterprise. They just hired the guy who built GTM for Salesforce, Microsoft, and Cap Gemini (Paul Smith). They’re the default model used by engineers and engineering tools (e.g,. Cursor). They’re also partnered strategically with Amazon and available in the marketplace.

Good Reads 📚

1. What happens when an LLM runs a store

This piece by Nekuda recounts the recent work by Anthropic to build an autonomous vending machine and business. It would handle customer service, sales, email human helpers and manage inventory. It failed quite spectacularly, giving away freebies, hallucinated financial data (like a fake venmo handle). Nekuda’s diagnosis is fascinating.

It lacked a competing system prompt to “be helpful” in Claude, it had too few guardrails and scaffolding and too many different types of autonomy. If we can create self driving cars, why can’t we create self-driving online businesses? Nekuda predicts we will see autonomous stores within the next 2 - 3 years, but to get there we need autonomous payments. 

🧠 You can learn so much about agentic commerce by reading what these folks write. The Nekuda blog is becoming my favorite on the internet.

🧠 Their argument goes that the Visa Intelligent Commerce solution will give agents the guardrails they need to operate within boundaries. I don’t think it’s that simple. That’s one tool to limit what can be transacted but we need others.

🧠We already have entirely automated youtube and tiktok accounts. Generating content, collecting ad revenue, rinse and repeat. That’s sort of an autonomous business. With enough n8n, evals and prompt engineering you can do incredible things already.

Tweets of the week 🕊

That's all, folks. 👋

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

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