Agentic Checkout: Stripe + OpenAI’s new protocol

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 from NYC, where I saw a few folks and did Stripe Tour. The one where they announced all the things, like agentic commerce, or the ability to create your own stablecoin. 

Not to be outdone, Brex went and announced their stablecoin product, which is very different from Ramp’s. Letting companies get paid in stables, or pay credit card bills with stables. Stablecoins are the new default.

And in possibly the most significant Fintech news of the year, Nubank announced it is applying for a US national bank charter. They’re already the primary bank for 70% of their customers. Maybe this is finally the challenger for that crown for the big banks?

Fintech Nerdcon is coming. And we’re building a quest map for folks who lean AI, stablecoin, policy, or open finance. We’re halfway to sold out. Grab a ticket before they’re gone!

Here's this week's Brainfood in summary

📣 Rant: Agentic Checkout - Stripe + OpenAI’s new protocol

💸 4 Fintech Companies:

1. Circuit + Chisel - Teaching AI how to Pay.

2. Shield - Stablecoins for import/export 

3. Pillar - Construction Finance Platform

4. Alguna - The Monetization engine for AI 

👀 Things to Know:

1. Nubank has applied for a US Bank Charter.

2. Brex announced stablecoin payments

📚 Good Read: AI Scaling exponentials are coming

If your email client clips some of this newsletter click below to see the rest

Weekly Rant 📣

Agentic Checkout - Checkout comes to ChatGPT

The quiet part out loud of agentic* commerce is there’s no volume yet. But we know ChatGPT converts clicks into a sale between 2x and 6x higher than Google. I know math is hard, but just to state the obvious If you're a merchant doing $100M at 2% conversion, this could mean $600M as volume to Chat increases.

The market agrees.

When Etsy announced they're the launch partner for a new type of Instant Checkout in ChatGPT, its stock jumped 16% in a day. Maybe it’s AI hype, or maybe it’s ChatGPT's 200 million US users being able to make purchases without leaving the chat. This announcement came with the new Agentic Commerce Protocol (ACP) built in partnership with Stripe.

And Etsy isn't even a Stripe processing customer. That's the point. This is an open protocol.

The game genuinely just changed. Before these announcements, merchants had a horrible choice.

• Reach new customers 

• OR keep control of the transaction and block bots

So what’s different, and why this?

1. The announcement: Agentic Commerce is Live

2. How does it work: product feeds, programmable shared checkouts and shared payment tokens

3. What does it mean for merchants: don’t change your stack, get new customers

4. Where do all the other protocols fit: probably alongside most likely, but its early

5. What should you do about it: As merchants, figure it out ASAP!

Meet “Buy It In ChatGPT” and OpenAI/Stripe’s Agentic Commerce Protocol.

Agentic Commerce is Live and in Production

There are three related announcements

1. OpenAI launched commerce in ChatGPT, and the checkout is powered by Stripe. This lets users find and actually buy things in ChatGPT and reduces user friction. 

2. OpenAI and Stripe open-sourced a protocol to any merchant sell via AI Agents. Called the Agentic Commerce Protocol, it builds checkouts from SKUs.

3. Stripe also launched “Shared Payment Tokens.” This lets any payment method be collected by the agent, secures it, and sends it to the merchant who can send it to their payments processor (e.g., ChatGPT could potentially use your card, PIX, or stablecoin balance. Although today its just Apple Pay, Google Pay and cards.)

Put together, it looks like this

Any task that used to require 10 tabs and Google searching is ideal for an LLM. If you’re looking for a set of Mario-styled ceramic garden ornaments for your upcoming conference for Fintech Nerds. That could be days of scrolling. Or you’re doing something complex, like an office setup. 10 tabs. Pain. Product reviews.

It’s less search and more research. So when a user gets to an answer, they’d more likely convert.

But now OpenAI brought the checkout to the user. 

And the LLM became an entirely new channel for merchants.

How does it work

Stripe and OpenAI have created the Agentic Commerce Protocol “ACP”, which is designed to unite three parts of the commerce journey into a standardized set of APIs for merchants and agents to use. 

Once a user has an intent like “I’d like a ceramic Mario as decor for an upcoming conference,” there are three parts to any journey.

1. Item discovery 

2. Checkout building (Part of the Agentic Commerce Protocol)

3. Payments (Part of the Agentic Commerce Protocol)

Part 1: Item Discovery

The agent can find SKUs (products) in a couple of ways.

• LLMs have already crawled the web, and they can search it too. Most merchant pages are structured to allow the LLM to extract live product data.

• Merchants already publish product feeds for ads. Google Shopping XML/JSON dumps that have existed for a decade

(It would be cool if your LLM could trigger these buying suggestions, seeing that you had a birthday coming up, rather than having to think to do that, especially for us, not always good at remembering types)

Chat can already do this. The clever bit is what happens next. 

Part 2: Cart building

Before a cart can get built, the merchant has to register with the Agentic Commerce Protocol. (The merchant exposes an ACP-compatible endpoints). This is their "here's how to talk to my commerce backend" signal.

Then the new bit starts.

1. The user tells the agent, “I want that particular ceramic Mario please.”

2. The agent calls the merchant’s endpoint “1x ceramic Mario” from “mariosRus”

3. Merchant's e-commerce system (Shopify, BigCommerce, etc) receives this as if it were a request from their website (!!)

4. Merchant returns a checkout spec (items, price, taxes) and sends to the agent

5. The agent renders its own checkout(!!), which could be a standard checkout UI, a conversation or even an SMS. Every agent can have its own checkout UX.

Handily, they launched with a pre-made sequence diagram (so I didn’t have to make one 🤌).

Then it moves to payments.

Part 3: Shared Payment Tokens

Now the user has seen the cart and confirms they want to buy a ceramic Mario riding a unicycle for $24.99 from Mario R Us. 

1. The Agent collects a payment method (Card, Google Pay, Stripe Link, stablecoin, UPI, Pix, whatever)

2. The agent creates a payment token (not raw card numbers or wallets but representing them). It can also include fraud signals, customer reputation data and anything else agents and merchants want to share at the point of purchase.

3. The merchant “unwraps” the token. They get the payment credential, any fraud signals, and other data the agent sent (if it's Stripe it could include data from radar)

4. The merchant passes the payment for processing. This could be with Stripe, or any ACP-compatible PSP (e.g., potentially Adyen, Checkout, etc in the future)

5. If a step up is required, the merchant passes that to the agent. The agent then renders the verification (e.g., strong customer authentication in Europe)

6. Merchant confirms order and passes to agent via the ACP specification. The agent displays the confirmation and any order tracking.

7. The agent sends the merchant any information it needs, like email, order data, and payment confirmation.

Done.

Conceptually, we’ve decoupled the checkout.

The consumer bought from a merchant, and ACP connected them and their AI agents.

The UI is now an agent; the merchant back end is still there, they’re just passing tokens back and forth to stay in sync with each other. 

What does it mean for merchants?

Think of this bit as turning AI agents into a new channel. 

The good:

• No tech stack changes for merchants. The protocol doesn’t have an opinion about what back-end the merchant uses, or what UI the agent renders. It just keeps the two things in sync. The agent can render any checkout; the merchant can have any back end.

• This new channel could be high conversion. If your product was stuck on page 1500 of the search engine result, what were you going to do? 

• The Agentic Commerce Protocol supports any payment type. It doesn’t have an opinion about cards, non-cards, etc. 

• If agents allow, you can still see fraud signals, customer info and more. This potentially means merchants' existing fraud and loyalty data could be retained through this new channel.

The unknown:

• Is giving up some customer visibility worth it for higher conversion? One of the main reasons you want a customer to browse your page is to increase their average order value (AOV) with upsell / cross sell.

• Any agent could use the protocol, doesn’t mean they will. It’s noticeable that Anthropic’s MCP has caught on, but as yet, OpenAI hasn’t had a similar win. I think that will change here. OpenAI is the clear winner in consumer distribution, whereas Anthropic had more of a right to win in developers.

• Any payment processor could use the protocol, doesn’t mean they will. Given how merchant-centric this protocol is vs any of the others, I actually think other processors should adopt this. PayPal, Adyen, and Checkout at least.

• Any merchant could use the protocol, but that doesn’t mean they will. This protocol is designed to be incredibly lightweight. But the early adopters are likely e-commerce more than omnichannel (legacy) merchants.

Jeff Weinstein from Stripe had this lovely line when I spoke to him about the launch of ACP.

Isn’t that something.

The known trade-offs

• The Attribution Problem. Where did this customer come from and what else might they want to buy from us? If all you’re doing is passing tokens back and forth, as a merchant, you’re somewhat at the mercy of what an AI agent is willing to share with you. That said. I do think AI agents will want to sell attribution in return for some sort of revenue line from merchants. Maybe someone needs to make the UTM for agents.

• The Fraud Problem. On-page bot behavior is fairly well understood. You have device fingerprinting, velocity checks, and session replay. Stripe solves this by giving the merchant Stripe Radar fraud scores across Stripe's entire network, and giving customer and agent reputations and signals (which is cool, but means buying Radar and is non-trivial as a merchant). Other fraud vendors should adopt this idea quickly.

• The Reactivation Problem. In agent commerce, someone buys without ever hitting your domain. No pixel. No session. No browsing behavior. You can email them (the agent can pass email address). You can't retarget them on Facebook. Is that trade-off worth 5x conversion? For the long tail, absolutely. For big box retailers, optimizing AOV through upsells? TBD.

There are also other protocols here. Google, Visa, x402 are all making a play to be the relevant protocol for the future of commerce.

The immediate benefit is for long-tail merchants.

When Stripe and OpenAI announced this, the first company to reach out wasn't a tech darling. It was Slice, the vertical platform for local pizzerias.

Think about that. Local pizza shops have been digitally stuck for decades. They can't afford Shopify customization. They can't compete with Domino's on tech. They're stuck with phone orders and maybe a janky website.

But "Hi robot, please send pizza to my house, the one I like, but this time extra spicy"? That works. No web development needed. No checkout optimization. Just an API endpoint and suddenly they're in the game for the entire Chat customer base.

The ceramic garden ornaments shaped like Mario? Page 1500 on Google. But when someone says, "I need something for my conference about video games," the LLM finds it.

Where do the other protocols fit?

But this new protocol follows Google, Cloudflare, Coinbase, Visa, Mastercard and countless others throwing their hat into the protocol arena.

Think about the subjective motivation of the companies creating each protocol.

• Google is trying to defend its shopping position and create new experiences.

• Cloudflare is thinking more about content monetization

• Visa and Mastercard have solved for trust and consumer protection (underrated!)

• OpenAI and Stripe have gone right at buying things.

• ATXP is agents buying from MCP servers (which have broad adoption)

None of these answers are wrong. This phase sort of reminds me of the early phase of the internet. HTML had a forgotten competitor Gopher, then we had CSS, then JavaScript, and the evolution continues to this day. 

But if you were looking for a clear “winner,” then Model Context Protocol (MCP) has been it. When that launched, every AI influencer on the planet created content around it for weeks. (Complete with exploding head emoji).

The test for ACP will be if it endures and gets others to adopt it.

What should you do about it?

Your customers are already using these agents. This is a new channel, not a replacement. Your website still matters but if you're not where your customers are, someone else will be.

You should explore how this new set of protocols integrates with your existing stack. The key questions to answer are

1. Can our existing providers support this protocol? Ask your processor: "What's your agentic commerce strategy?" Etsy proved you don't need to process on Stripe to use the protocol, but you need SOME answer.

2. What changes do we need to make to render a checkout? The Agentic Commerce Protocol (ACP) is a super lightweight lift.

3. What signals do we want to get from that checkout? This is the key question and probably requires a 1:1 relationship with OpenAI to really nail

4. What KPIs are we testing? Is conversion rate higher than the web? What does this do to CAC and AOV? Are we getting net new customers?

This is a no-brainer for long tail merchants who use platforms like Shopify or Stripe, especially tech-forward or ai-native sellers. It’s the first attempt to do something in “agentic commerce” that’s deeply merchant-focused instead of just being enablement.

Checkout is dead. Long Live Agentic Checkout.

This is the most merchant-centric protocol I’ve seen for agentic commerce.

This protocol lets you:

• Keep your tech stack

• Keep some customer data (email at minimum)

• Potentially get 5x conversion

• Access new distribution (ChatGPT's 200M users)

The trade-off is real (you lose pixels, attribution gets fuzzy). But for the first time, the trade-off might actually be worth it.

And this shift is coming fast.

Stripe put over 100 people on this. Not 10. Over a hundred. That's the size of most scaling startups when you staff something that heavily, you’re chasing future revenue and future markets with conviction.

Just when you thought the checkout was dead.

OpenAI and Stripe re-invented it as an embedded experience for agents.

The checkout is dead. Long live agentic checkout. 

ST.

* This isn’t really an agentic flow, it’s a human driving an LLM, but heyo.

4 Fintech Companies 💸

1. Circuit + Chisel - Teaching AI how to Pay.

Circuit & Chisel are teaching agents how to pay and get paid. The goal is to make any agentic workflow accessible to any other agent, and have it pay for tool usage without having to make a subscription. 

🧠Chaging per tool call makes much more sense than requiring a subscription when its machine to machine. Imagine you build a language translator tool. If another AI wanted to call that tool, they’d need to be able to pay you, but if its just a one off task, signing up for a recurring subscription is prohibitive. C&C has a thoughtful set of developer docs on how this mechanism can be baked into the MCP Protocol standard which has already demonstrated strong adoption.

2. Shield - Stablecoins for import/export 

Shield helps international buyers pay quickly and affordably without having to use the correspondent banking system which is often slow and expensive. 

🧠Shield is doing $40m/month in transfers using stablecoins. Import/export SMBs are the ideal clients for a stablecoin based Neobank. SMBs are often poorly served by cross-border payments tools, especially if they deal with markets outside the top 60. Inside the top 60, Wise and Airwallex have you covered. But from Thailand to Cambodia? For domestic users I do low key wonder if this is tariff evasion as a service. 

3. Pillar - Construction Finance Platform

Pillar helps construction companies know exactly where their money is going, from revenue to expenses and has real-time margin tracking. It will reconcile bank transactions and flag any gaps before they become issues.

🧠 There’s a massive demand for infrastructure. Data center build out is staggering, and a lot of that is happening across Europe. These things have to go faster than the decades they used to take and software like Pillar is one key unlock. I just wonder if SaaS has much of a shelf life if AI could build this type of tool in a few years time.

4. Alguna - The Monetization engine for AI 

Alguna helps developers launch any pricing model (like subscriptions or usage), generate quotes, accurately meter users, manage billing in any currency, collect payments, recognize all of that revenue, and get data from your revenue. 

🧠This is for power users, and you can see it in the use cases. Like unifying sales and product-led growth, complex calculations for fintech products (e.g., based on partner bank and card network COGS). It even helps salespeople build a complex quote quickly because (say it with me) “time kills all deals.” There’s a lot of “Stripe Billing for x” companies, but this is the best I’ve seen. 

Things to know 👀

1. Nubank has applied for a US Bank Charter.

The application is at a preparatory stage, and the bank says part of its goal to expand beyond Latin America. They’ve assembled a stellar board for market entry. Roberto Campos Neto (former President, Central Bank of Brazil) as Chairman, Brian Brooks (former Acting Comptroller of the Currency), and Cristina Junqueira (co-founder) has relocated full-time to run US ops

🧠This is the PRIMARY bank for 70% of active customers, 83% of which are monthly active. Can it crack the USA? They have $3.7bn quarterly revenue, zero branches and insane unit economics (costing $0.70 to run an account per month). 

🧠 Nubank solved the hardest problem in banking: winning primary relationships at scale without physical infrastructure.

🧠 What happens when they enter a market where. 33M+ Latin Americans already know the brand and the product is battle-tested. Consider that they have the capital to last a multi-year grind of getting licences and market adoption.

🧠 The US is a SUPER competitive market. Nubank is now playing on hard mode. Can US banks compete on customer experience when their infrastructure costs 10x more to maintain? Or can Nubank get distribution?

🧠You’d imagine the LATAM diaspora and remittance are the start. But where they go after will be fascinating to see.

2. Brex announced stablecoin payments

Brex will enable its clients to accept stablecoins and receive USD instantly. Customers will also be able to pay card balances with stablecoins. 

🧠 Historically, there’s a big gap between receiving stablecoins and it showing up in your operating account (e.g, Brex). If you received stablecoins into a wallet, you’d then likely need to convert that, and wait for a wire before you had the funds available in your treasury.

🧠 Brex has a lot of crypto native customers. People forget crypto is an industry with public companies like Coinbase and Circle; a subset of these giant companies are Brex customers. If you’re a stablecoin native, your fiat operating account should be too.

🧠A core use case for stablecoins is global treasury management. If your treasury is in stablecoins, why wouldn’t your operating account let you accept those and pay your bill with them?

🧠While banks are debating if they should do stablecoins, Fintech is shipping. The market isn’t waiting. The demand is phenomenal. And if you’re not seeing it you’re just not looking int he right places.

🧠People forget that stablecoins are a platform for lending (and other products). Consider how Coinbase's integration with Morpho allowed people to borrow against Bitcoin, or lend to borrowers and receive 10.6% APY. What might Brex and its peers do here?

Good Reads 📚

1. AI Scaling exponentials are coming

There’s a lot of talk of an “AI bubble,” given the sheer amount of capex heading into AI data center build-out. But the author Julian Schrittwieser pushes back. Based on current trends, 

• AI task length and token usage (therefore, data center demand) continue to explode

• Models will be able to autonomously work for full days (8 working hours) by mid-2026.

• At least one model will match the performance of human experts across many industries before the end of 2026.

• By the end of 2027, models will frequently outperform experts on many tasks.

🧠 The market is overbought on historical trends; this will impact things. Companies like Oracle are issuing massive bonds for build out. Big tech is loading up on risk.

🧠 Another 3 or 4 “deepseek moments” could be on the horizon, where tokens can be processed much cheaper and faster with new research

🧠 Human demand for intelligence will never go down. But if it became suddenly much cheaper, damaged confidence that could still trigger a risk adjustment. 

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