The Four Models of Agentic Payments

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 👋

Nubank. Just. Keeps. Winning. Wow, 22% more customers, revenue + 50%, net income + 85%. What happens when they launch in the USA?

Meanwhile, Chase will block payments originating from social media or going to marketplaces. That’s smart. We’re in a scamdemic ™ after all.

Crypto’s biggest scandal since FTX went down last week. The Argentine president, memecoins, and insider trading are all alleged, and it looks bad. We desperately need crypto regulations. ASAP.

I loved how Visa has started to describe itself as “Visa as a Service” and is actively unbundling itself to support more payment methods. Smart strategy.

PS. Heading to Fintech Meetup in a few weeks? Me too! See you there?

Here's this week's Brainfood in summary

📣 Rant: The Four Models of Agentic Payments

💸 4 Fintech Companies:

1. Orin - Zendesk + AI Agents and workflow for Fintech

2. Affinity Africa - Savings & Loans for Micro SMEs in Ghana

3. ApplePie - The Franchise lending Marketplace

4. Comulate - Accounting Automation for Insurance

👀 Things to Know:

1. Memecoins are dead. Argentina’s president killed them.

2. The Scamdemic: Chase to block all Zelle payments originating on social media

📚 Good Read: Varo's CEO steps down

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

Weekly Rant 📣

The Four Models of Agentic Payments

Agents that can find, book, and pay for your next vacation are the canonical use case for Agentic Payments. Yet it seems like everyone is lost in "how will it work though?"

There are four models.

1. The Agent performs checkout with a human in the loop

2. The Agent gets authority from a user wallet (like Apple Pay)

3. The Agent has a limited ability to pay (like a virtual card)

4. The Agent has its own wallet and funding (e.g., a stablecoin wallet)

This brief piece explores which is the most promising and, therefore, where you should be spending your time.

Recap: What is an AI Agent?

As I wrote in AI Agents will be Wall St firms.

I’ve used this example before. Imagine you ask your AI agent to find a flight to Antarctica and book the cheapest one. It might look like this:

So you’re immediately wondering.

How will they pay for things?

Great question.

Model 1: The Agent performs checkout but waits for user input

Metaphor: Your human assistant needs step-by-step instructions to complete a task.

Open AI Operator (and the countless alternatives) takes human instruction and allows users to browse the internet to perform that action.

The process: 

• You give your AI Agent an instruction like “find a hotel for me tomorrow night in Austin.” 

• It then hits the internet, performs a search, and gives you the options. 

• You select an option, and it heads to the checkout page

• It asks you to input your card information and uses that to pay

You can test this today with Open AI’s operator. 

How it works under the hood:

• Think of it like early Plaid (screen scraping)

• It’s taking user credentials (for Plaid it would be bank login, for checkout its card info)

• It then pushes those credentials at the secure page

• It then reads the page and plays back the result to the user 

The problem? 

The agent always gets stuck on the payment screen, and needs user input.

The magic completely breaks if you have to manually enter your card information. The checkout page is one of the most painful bits of e-commerce for merchants and consumers. 

Model 2: The Agent can use stored credentials the user approves.

Metaphor: Your human assistant gives you options and asks for approval before paying.

The agent can browse the Internet and pre-populate login or card information that is stored securely “on file” or “in wallet.” (If you’ve used Shop pay or stored your credentials at Amazon this will be familiar)

The process: 

• You instruct your AI Agent to “find a hotel for me tomorrow night in Bognor Regis.” 

• It then hits the internet, performs a search, and gives you the options. 

• You select an option, and it heads to the checkout page

• It pre-populates the card information and asks you to approve payment.

How it works under the hood:

• Think of it like Plaid today (tokenized access)

• It takes a user's permission to receive a “token” from the users bank.

• It then pushes that token for secure data retrieval. 

We have a similar experience today in e-commerce and browsers. Your browser often asks if you want to store your card information or password. Then blam, it just appears. You still click “log-in” or “pay,” but with less friction.

This model has less friction and much more scope for even better UX. 

Imagine talking to your AI Agent like an EA. It finds you your flight and trip information and gives you the options. Then you can simply say, “Yes I like that go ahead and pay.” Then it can use the “card on file” in the browser or e-commerce page.

The problem? 

The agent still gets stuck on the payment screen, it's just a better UX.

PS. Chrome should just do this. You have Google Wallet, Chrome, and Gemini for the love of Christ Google. 🤦‍♂️

Model 3: The Agent has a Virtual card

This is where it gets much more fun. 

Metaphor for Virtual Cards: Your human assistant has a virtual card and can just pay for things.

The virtual card process: 

• You instruct your AI Agent to “find a hotel for me tomorrow night in Stevenage.” 

• It then hits the internet, performs a search, and gives you the options. 

• You select an option, and it heads to the checkout page

• It uses its virtual card on file and pays

How it works under the hood for virtual cards:

• Like a modern spend management card.

• The agent is given a budget and set of controls in the card about what is in and out of policy.

• It then uses that card when it goes to pay

Stripe kinda does this today. First you do a standard payment, then you get to use the virtual card model. Here’s how Jeff described the flow to me.

1. Consumer uses shopping agent’s search (e.g. Operator or Perplexity) to discover goods to buy

2. The consumer clicks “buy” button, fills out shipping & billing information.

3. This experience is hosted via Stripe Checkout (or any Stripe payment acceptance, with any payment method, in any currency). 

4. Money settles in Stripe balance.

But the subsequent flow is where it gets clever

5. A virtual card gets created via Stripe Issuing, 1 per transaction.

6. This card is not exposed to the consumer and can have multiple different controls to limit fraud (e.g., amount, MCC, expiration). It has a webhook where the app/platform can review the transaction or call human in the loop approval process.

7. The purchase is completed with the merchant by submitting the cards in a PCI-compliant way.

8. End merchant sends good to consumer.

Want to know something wild? This is live. On Perplexity Pro. Today. Go search for something like TV’s, and up pops a checkout page, and if you have Stripe Link credentials, it’s already pre-populated.

The problem with virtual cards? I’m not sure yet. This seems like low-hanging fruit. You could argue you’re at the mercy of what possible card controls exist. The card network data is limited. Vs is a stablecoin that is truly programmable, global, and instant. They also quite like a checkout page which isn’t ideal.

The problem with virtual cards? I’m not sure yet. This seems like low-hanging fruit. You could argue you’re at the mercy of what possible card controls exist. The card network data is limited. Vs is a stablecoin that is truly programmable, global, and instant. They also quite like a checkout page which isn’t ideal.

Model 4: The Agent has a stablecoin wallet

Metaphor for stablecoin Wallets: Your human assistant has an “account” and can buy, sell, or earn income online just like you would.

This example brings it to life. This AI agent can book and pay for appointments WITHOUT a checkout flow. Agentic payments are here.

The stablecoin wallet process: 

• You instruct your AI Agent to “find a hotel for me tomorrow night in Knaresborough.” 

• The agent handles a phone conversation 

• The user can give audio permission

• There is no checkout

• Payments are as seamless as with Uber 

How it works under the hood for stablecoin wallets:

• The agent is added to a wallet (in this case an Eth compatible wallet)

• The agent is given a budget and set of controls in the card about what is in and out of policy.

• It then uses that card when it goes to pay

The problem with stablecoin wallets? They’re still early, and they’re not a guarantee of success. Today the users are mostly holding dollars to avoid currency devaluation in the global south. 

But what makes stablecoin wallets exciting is, agents can live inside that wallet, and be controlled by it. (As I wrote in Wallet Wars Pt4: Crypto Agents)

The B2B Opportunity for Agents is Huge

While we've focused on consumer use cases, businesses may adopt these models even faster:

• For Model 2 (Stored Credentials): Corporate travel booking where an agent finds flights within policy and requests manager approval. Folks like Navan are already doing this kind of thing.

• For Model 3 (Virtual Cards): Procurement teams giving agents specific budgets for office supplies or software licenses, or to complete tasks like running SEO or outbound SDR campaigns.

• For Model 4 (Stablecoin Wallets): A Treasury agent could manage cross-border payments between business units without currency conversion delays

What do I need a checkout for?

When you think about it. Checkouts are a very early 2000s skeuomorphic way to pay.

The function they serve is 

• Collect the list of SKUs and quantity a user wants to purchase

• Collect their shipping information (if not in-store)

• Collect some sort of money (with cash, card, pay by bank or stablecoin)

• Under the hood they’re also using data to try and prevent fraud too

What happens when that is more ambient?

Couldn’t that function be served much better with tokens and wallets?

Spoiler alert: yes, it could. 

The CEO of Coinbase seems to think so too.

Is he right?

Maybe. Time will tell.

The Next Frontier: Agent-to-Agent Payments

What happens when your shopping agent and a merchant's sales agent negotiate directly? 

Imagine your agent finding the best price for a product by negotiating with multiple merchant agents simultaneously:

• Comparing features, 

• Requesting discounts

• and finalizing transactions without human intervention on either side.

This creates entirely new market dynamics, where algorithmic negotiation becomes the norm, and liquidity pools managed by agents become the backbone of commerce.

This would be pretty hard to build on virtual card infrastructure and with today’s authentication model.

For this to work, we need Model 4 (stablecoin wallets) with sophisticated permission structures. Agent-to-agent economies could completely reinvent wholesale and B2B transactions before ever touching consumer markets.

The new tech’s winning paradigm is impossible to pattern match.

The reality is there is no pattern. With tech, there rarely is.

• In the early days of the internet, many assumed that the AOL portal model of browsing the internet would win. It didn't. Browsers did.

• In the early days of mobile, many assumed the browser model and HTML5 would win. It didn't. The App Store did.

We still don’t know what the AI Agent form factor is. So how will they pay?

This is a little clearer.

Payments won't evolve uniformly. We'll see: 

• Legacy systems sticking with traditional checkouts (Model 1) 

• Newer relationships starting with approval flows (Model 2) 

• Complex purchases using virtual cards with sophisticated controls (Model 3)

• New forms of commerce using stablecoin wallets (Model 4) 

What usually matters is execution. The Netscape browser out-executed AOL by offering users something they wanted more. The App Store out-executed HTML5 by working and providing a super clean and useful experience that fit the mobile form factor.

The winners will be those who match the right model to the right context, rather than forcing a single approach everywhere. 

What's certain is that payment flows are becoming more agent-centric, more programmable, and less visible to the end user. 

Whether through virtual cards or stablecoins, the checkout as we know it is living on borrowed time. 

ST.

PS. Thank you to Jeff and Steve for giving me color on how Stripe’s agentic payments work 🙏

(Pro tip: Claude will generate .mermaid javascript diagrams from a list; here’s a little Vercel app that converts that to Excelildraw. Thank me later).

4 Fintech Companies 💸

1. Orin - Zendesk + AI Agents and workflow for Fintech

Orin provides customer service agents for the regulated world of Fintech and is trained on 1,000s of real-world cases. Behind the scenes, they're a workflow product that integrates various customer service and data systems to provide consistent customer output. Being AI, they can scale around demand spikes.

🧠 Priced at $500/month, it's like hiring offshore talent but without the training ramp. That's pricey. The argument likely goes these are higher-quality outputs than human agents. However, I noticed the website has three case studies, all from the same company. I'm all for customer service agents + workflow + industry context, but tidy your website folks ;)

2. Affinity Africa - Savings & Loans for Micro SMEs in Ghana

Since launching in October 2024 (wow), Affinity has provided affordable banking access to over 50,000 micro SME's. Users can make payments, make transfers, save, and, over time, gain access to loans and investments. Affinity is distributed through an agent network. Today, Affinity has a nonperforming loan (NPL) of 3%.

🧠 African markets are steadily becoming a growth engine for the world. As the West and China struggle with aging populations, Africa is young, digitally savvy, and now tech-enabled. It's still very much the hard-mode market, with little infrastructure. But the recipe of agent network + local knowledge is the one/two punch to look for.

3. ApplePie - The Franchise lending Marketplace

ApplePie is a pure-play franchise financing marketplace that has distributed over $3bn with over 100 brand partners. As a marketplace, they help brands find franchisees and manage financing. They have brands like Orange Theory, Anytime Fitness, and Little Ceasars.

🧠 Franchises are a difficult investment that needs hand-holding; this is smart. Until 2016, Shaq had 155 Five Guys locations and still owns a handful of Auntie Ann's and Big Chicken's (which he co-founded). They consistently return high yields but need upfront capital and a lot of paperwork to get started. Imagine how many investors would get access if this marketplace could be distributed via Robinhood (for example).

4. Comulate - Accounting Automation for Insurance Brokers

Complete helps insurance brokers collate PDF documents from carriers, reconcile them against databases, and post them directly to the insurer's ERP. It can identify upcoming payments and billing errors and perform revenue forecasting.

🧠 Insurance always felt happy to be stuck in the 1960s, but finally, thats changing. Lots of paper, suits and briefcases. Brokers are a great client segment because there are so many of them. Every Insuretech needs a Fintech back end.

Things to know 👀

1. Memecoins are dead. Argentina’s president killed them.

When Argentine President Javier Milei endorsed $LIBRA, the price rocketed then crashed spectacularly. His claim? Funding entrepreneurs would boost Argentina's economy. The reality? A local NGO discovered 40,000 citizens lost approximately $4 billion. Argentina's stock market plummeted 5% Friday.

The scandal deepened when Kelsier (led by Hayden Davis) apparently convinced both Trump and Milei teams to launch memecoins, then allegedly manipulated Solana-based token launches. Davis' crew stands accused of siphoning $200M through insider trading and sniper bots, implicating major exchanges like Jupiter. (Lex Sokolin's breakdown is essential reading)

🧠 Crypto regulation isn't optional anymore. Anyone can buy worthless memecoins while actual innovation like OpenAI remains inaccessible to regular investors.

🧠 The serious crypto community is fleeing the memecoin dumpster fire. Industry veterans Nic Carter and Austin Campbell have declared them gambling, not investing.

🧠 This garbage destroys real economic value. A 5% stock market crash from memecoin shenanigans shows the tangible damage of treating finance like a punchline.

🧠 The timing couldn't be worse as stablecoins finally demonstrate genuine utility. My hope: justice served, fraudsters jailed, and we refocus on meaningful innovation.

🧠 Remember: Terra/Luna called itself a "stablecoin" while its founders earned prison sentences. 

Meanwhile,

• Legitimate stablecoins continue gaining serious traction.

• We're close to stablecoin and crypto regulation

• Consumers desperately want access to investable assets

Crypto is ready for its Spotify vs Napster moment.

Lex Sokolin's analysis exposes the full scope of this disaster. Check it out here.

2. The Scamdemic: Chase to block all Zelle payments originating on social media

The block will come into place from March 23rd in response to a high risk of scams and fraud. The company said 50% of all fraud claims originated from social media. Chase and Zelle point out that scams and fraud are a tiny fraction of overall payments. The company also says it may request more information when you add a payment recipient to Zelle

🧠 This isn't an overreaction; it’s best practice for all RTP payment types.

• Block risky sources of transactions OR

• At least ask for more information from the user

• Adding that bit of friction is something UK banks have been doing for several months now

🧠  I've been saying for a while now, we're in a scamdemic ™ This comes from structural vulnerabilities:

1. Social platforms weren't designed as financial gatekeepers

2. Real-time payments create irreversible payments with no clear liability framework

3. The "buyer beware" approach to consumer protection isn't working.

4. Scammers know this and are exploiting it

🧠 Banks may or may not be held liable, but there’s a broader set of consequences that make this a huge issue

• SDNY vs Citi, and CFPB vs Early Warning Services law suits are one small piece of a bigger problem.

• Consumers are losing money and complaining to their banks and when that happens, so do press headlines that the markets read, and that impact stock prices

• When consumers lose out there are always consequences. Even if the CFPB isn’t moving today, administrations change.

• Consider too that the states will pick up where CFPB left off. e.g., NYDFS is positioning itself to maintain oversight in the interim

Good Reads 📚

1. Varo's CEO steps down

The first Neobank to get a de-novo charter, Varo has announced their CEO Colin Walsh will step down. While revenue is growing (22%), Varo still lost $65m on ~$154.8m revenue. Varo is growing its user base, but those users aren't growing their deposits, and they are still making most of their income from the interchange. It's now running out of runway and needs to figure out what's next.

🧠 This is a Neobank in all but charter. When the majority of revenue is interchange years after the charter, you have a problem. The lending product just hasn't stuck the landing.

🧠 I wonder if they're an acquisition target? If I were a bigger consumer Neobank with more mature lending products, M&A might be a better option than the de-novo charter.

🧠 Let's give Colin and Varo their flowers for taking the hard path and getting a de-novo charter. It's still way too hard to walk through the door and make that happen, no matter how well-run you are. You have to wonder how much that slowed their growth as others took different paths.

🧠 The CEO stepping back can be good for everyone. It worked for Monzo. Tom Blomfield is a true founder but wasn't enjoying running a regulated bank. TS Anil is a banker who gets Monzo, and they've gone on to amazing things, as has Tom (at Y Combinator).

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 :)

(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