Allison Farris - Teaching Agents to pay: A Developer's Guide

This talk discusses the development of agentic commerce, focusing on how AI agents can interact in financial transactions. The speaker, a developer advocate at Stripe, highlights the shift from traditional human-centered checkout flows to agent-first models where transactions occur through APIs and structured catalogs. It emphasizes the importance of designing systems that ensure safety, audibility, and understanding in agent interactions with consumers. The presentation explores the Universal Commerce Protocol (UCP) and the Machine Payment Protocol (MPP), which enable seamless transactions between AI agents and merchants while maintaining merchant control over pricing and inventory. Overall, the talk illustrates how AI's evolving capabilities are reshaping the e-commerce landscape and enhancing product discoverability.

everyone. Um, welcome to teaching agents to pay, a developer's guide. I'm Allison. I'm a developer advocate here at Stripe. And as AI systems start to become more capable, the boundary between actually making making decisions and executing becomes even more critical. So, this talk today is going to talk about how agents work in the real act and work in the real world, specifically when money is involved, and

how to design systems that are that remain safe, auditable, and understandable. So, your checkout flow is already obsolete. You just don't know it yet. In today's world of checkout flows, it's very heavily optimized for humans. Think forms and UI driven. We obsess over making sure that we minimize the amount of clicks for you to be able to check out seamlessly with a payment method. But the shift

to the agent-first era is already here. And agents don't complete forms. They call APIs. They interpret structured catalogs. And they execute programmatically. So, when you're thinking about building your commerce infrastructure today, it's important that you understand the model and keep this in mind. And the data is already starting to show. So, we took data from 2025 and it's surfaced that one in four people have already been

using search chats chatbots as a front door to searching for product products and discovering what they want to buy. And this number is even higher in younger generations. And so, the bottom line here is that consumers are getting more and more comfortable leveraging AI to discover new products and items. And if we take a step back to understand where we've come from and how AI has emerged,

we're starting to see this shift from just being read-only to now agents being able to write and act on our behalf. And we're entering this era of now or And with tools like MCP and agent-to-agent capabilities, we're starting to see the shift from LLMs just being chatbots to now being full economic actors. So, it with a show of hands, who here has actually used chatbot to search

for products online? I know I have. And keep your hand up if you actually tried to pay through a chatbot. All right, not as many. So, the experience today is relatively fragmented. So, if you think about um searching for items, right now I'm planning my ski trip next year to Japan, and I'm ready to take the leap and actually make a purchase of and bring my own

skis. So, I'm going to my chatbot, and I want to buy some expert skis, and today LLMs are really good at discovering, making comparisons, and decision-making. And so, I'm able to give it some prompts, tell it what I want, and then it makes me a really great recommendation. Perfect. So, I'm now ready, I'm sold, bought in, and I want to actually make a purchase. So, if I'm

I'm ready to to buy. And today, the experience is fragmented. It's It breaks context and flow. It redirects me outside of the chatbot experience. And you can see that checkout today wasn't designed for agents in mind. And we're also starting to see that by no surprise, purchasing within the chatbot experience itself is seeing higher rates of conversion. So, what conversion means is that um I'm more likely

to actually complete my purchase. So, as a business, it's outperforming traditional e-commerce from the early stats that we're starting to see. So, I want to walk you through what this experience can be when it uh when you walk through an actual a chat chatbot and make a purchase within the chatbot So, here we've got a demo, and I'm going to prompt it to help me understand what

skis to buy. And it asks me, "Okay, are a beginner?" So, I'm going to aim high. I'm going to say, "You know what? I'm looking for expert skis." Great. So, now you can see that it's pulled from a product catalog. It's given me some options with a little bit of information, and I'm asking and suggesting, "Okay, give me a for some skis. I want to ski in

Japan. There's lovely deep powder and snow." And I continue on this journey as I'm trying to understand what to buy. So, now at this point, it's given me a really strong I'm super interested in the Enforcers, and I want to understand a bit more about why I should buy these. Okay, great. It gives me even more detail about the product. And, you know what? At this point

in time, I'm already sold, and I'm ready to actually continue making the purchase. So, I'm going to decide on purchasing these specific Enforcer skis at the 179 cm length. All right. And at this point in the stage, I'm going to now um set up my profile and add a payment method. So, I'm adding my payment method directly to the AI platform through Stripe's payment element. So, this

allows safety and security. And then at this point is where uh the checkout doesn't actually redirect me into an external redirect link, but I stay in the chat interface, and I'm actually going to proceed and and tell this agent, "Yes, take my added payment method and actually complete the purchase." So, the moment of truth. And there we go. So, my purchase is confirmed. I'm going through to

understand if my order went through. And you can see that now my agent has made a purchase on my behalf. So, that's an example of what agentic commerce can look like through the experience of an AI chatbot. So, the question now is how do we safely enable this? When we think about building an agent, we think about the core components of what an agent is comprised of.

So, first it's the concept of the brain or the LLM that decides what should happen next. And then the brain needs tools to know how to act and to make to take those actions based off of the decisions of the brain. And you can think of these as the hands of the agent. Then you have instructions that shape the brain's reasoning and tool selection. Think of these

as the guardrails. And when you're building out an agent, it's essentially just an LLM in a while loop that you're able to execute functions based on the conversation that's ingested from the chat, and it decides ultimately the next step. And once it's done, it returns the results back to the client, back to the customer. And in order to be able to execute those functions, they're defined by

a set of commerce tools with respect to your checkout experience. So, think, okay, I'm going to create a checkout, what information do I need to be able to complete an order? I need to make sure I have an added payment method involved before the agent attempts to actually make the charge. And these sets of commerce tools define the action that the agent can actually take. So, in

this example, think about the function around creating a checkout and all the specific parameters and information that you need to be able to pass along that request to the actual merchant to make that charge. So, how does the agent decide what to do next? And so, if you take this concept of the loop, the first step is the LLM decides on the actual tool based off of

the chat so far. And then they send that back to the back end to make it to execute that tool. The tool then sends up a tool message, returning it to the LLM as a tool message, and now the LLM sees that there is a checkout, and now it's the status is not ready for payment. So, let's look at the example based off of the demo we

just saw, and now um let's see these three components work in action. So, now I'm ready to go, I'm in my chat interface, and I prompt to my agent, yes, I'm ready to buy. The LLM receives a system prompt, and it goes through this loop to understand, okay, now I need to create a checkout session, I need to execute this tool call, I need to do a

little bit more in the back end before I prompt the user. I understand, okay, I'm not ready for payment, so I need to then send a message back to the customer to add your shipping address. And so, this loop continues, it reads and ingests from the prompt, and it makes these decisions based off your commerce tools, and it continuously goes through the loop until the LLM decides,

okay, it needs something else, and it prompts the customer now to add a I add my payment method, and I say I'm ready for the order to be completed, and this is where it goes through the loop again, takes my payment method details, understands it has everything, attempts the charge, and then order is completed. The next question to ask is, how do I make my checkout agent

ready? But not only do how do I make my checkout agent ready, but how do I not have to build custom integrations for every single agent? And so, enabling uh product discoverability across one AI platform is simple enough, but then how do you think about integrating new AI platforms as they start to become available? And the inverse of this, how can AI platforms start to surface up

multiple types of merchant catalogs without having to build spaghetti code or having to do so many custom integrations? And so, um in the early days of railways, something interesting happened. Every company built their tracks slightly differently. You know, think different widths, different standards, and different exceptions, which ultimately meant that trains couldn't move between these railways. And so, at every boundary, you had to stop, you had to

unload everything from the train, you had to reload the train, and get [clears throat] back on to the different track. And so, you can see how this is can be very slow, it can be very expensive, and very error-prone. And the key thing is, the more the railways got added, the worse the system actually got. So, they weren't really building a network, you were actually building friction.

And this is exactly the problem that we're starting to see and run into with agentic commerce today. And this is why um protocols like the Universal Commerce Protocol are really, really important and um allows merchants and agents to have this shared language to be able to transact instantly. And we're starting to see a huge uptake um and adoption across companies and retailers like Target and Google. And

Stripe just recently joined the Tech Council. So, we we're seeing a lot of momentum in this adoption today. So, what does the Universal Commerce Protocol actually solve? It allows for the merchant to build a set of commerce for the AI agent to be able to read and understand and act upon uh the different products that the merchant wants to sell through the platform. And so, if you

think about um UCP, it's also a way to really standardize the way that agents and merchants interoperate without having to completely rebuild a spec every single time. So, in essence, as you start thinking about AI agents, they're speaking the same language, they're standardizing the the same endpoints, and it provides a standard set of tracks for infinite amounts of merchants and infinite amount of AI agents within a

consistent and standard way. So, my favorite part and favorite question we get is, "How does the agent actually pay the seller and who charges the customer? And this is where shared payment tokens come in. So, if you look about look at this part of this three-party system, you have your AI agent, you have your seller, and you have your payment processor like The AI agent, similar to

in the demo, it saves my payment method information and passes it to Stripe. Um and Stripe will create and generate a decent sized token, also known as the shared payment token or SPT. This SPT then is able to be passed from the AI agent to the seller. And then the seller itself is the one who actually makes the payment and processes the order. Once the um charge

goes through and the payment processor processes the payment, it then returns the response back to the seller and the seller can then give the AI agent a success or decline. So, you can start to see how these things work together. It also allows users to shop across multiple sellers. So, let's take a look at their Stripe account structure. The AI agent will be processing and sorry, will

be issuing an SPT through their own Stripe account. They then will pass that SPT to a merchant for the merchant to then process the actual payment intent on the seller's Stripe account. So, you can see how this interoperates and keeps the merchant in control. one core uh component of this model is that it allows the merchant to remain the source of truth. So, they're always going to

be um responsible for their inventory, the pricing, the final amount charged, and the actual payment processing. The AI agent is um the AI agent's role is to relay that customer's intent and facilitate the movement of that shared payment token. So, double-clicking on the shared payment token, it acts takes on a least privilege approach. So, it only grants what's needed, and you can set specific limitations and restrictions.

And if any of the guardrail is violated, the Stripe will reject the charge. And Stripe always ensures that all of these limits are enforced. Failures are also recoverable. So, if an agent needs to, they can reissue a shared payment token. And Stripe logs all of the shared payment token usage, so there's a end-to-end audit trail. But, what about other payment experiences? We talked about tools now in

the world of agentic acting on the brain's decision. So, we're starting to see more and more tools come up. But, what happens when an agent needs to pay for things like a single API call or a unit of compute? Where there's no card, there's no checkout session, there's no API key that you need to worry about. It's simply just a machine-to-machine And fun fact, back in 1992,

the architects of the web who always knew we needed some form of payments and actually defined a 402 payment required status code. And they imagined this future where instead of ads, users could pay tiny amounts, kind of like a micropayment, for individual articles, let's say maybe a single data point and some photos. But, we just never built it. Well, today we're starting to see usage come out

and build more use cases for a machine-to-machine payments. And in the world of agentic, Stripe has partnered with Tempo to um build the machine payment protocol that allows um an open internet-native protocol to support machine-to-machine payments. But, what does actually What does this actually mean? Essentially, the machine payment protocol allows a machine or an agent to make a per pay per request for things like session based

usage or maybe a one-off compute continuous programmatic payments essentially. And this also happens and is supported across multiple types of payment rails. So you think crypto rails, card, bank fiat. And it ultimately allows the payment to become part of the request. So let's take a look at when a merchant server responds with a 402 payment required status. What this now allows is the ability to um leverage

the authenticate header that includes a method parameter indicating which type of payment method to use. Then the client, if supported, can then send that corresponding payment method to generate a credential and retry that request again. But what about with a shared payment token? The agent, again, can generate a shared payment token with the stored payment credentials and use that shared payment as part of the request to

the merchant. So you can see how it can interoperate across both machine-to-machine and the UCP protocol. So let's take a look at an actual example. Let's say that I want to book an Italian restaurant, invite my work friends. I want my agent to do everything for me and the agent maximizes the type of services and some of those services require small charges. So again, going back to

that loop, it's going to go and autonomously make the decisions and pay per request during during the processing. And by the end of this loop you can now see it aggregates the total amount, it makes the payment at the time of the request and now it outputs the total charge at the end. That's just one example. And so now you're starting you can kind of see a

world where a customer is no longer human. So, you know, buyers are increasingly becoming agents. They can decide and act and now with MPP, they can pay instantly at the time of the Um and so when agents act and pay autonomously, system design matters more than ever. So if you're designing a system, one of the core components that's critical for AI development is making sure that your

system prompts are reliable and specialized agents. So in a sense your system prompt is your persona and ethics policy written in plain English. Um some of the key questions to consider is making sure, you know, should you disclose that you're an AI? How much urgency do you want to create as part of that shopping commerce experience? Um do you want to remember past purchases or upsells? So

when you're defining your system prompts, um it's important to keep these rules in mind. So let's take a couple of examples. So the first I'm going to go through two um two personas that I've popped into ChatGPT to ask what kind of image that you think how they look like today. So in this first one they're a helpful shopping assistant and I'm setting a you know, I'm

giving it a set of rules. Okay, don't want you to disclose don't do disclose that you are an AI, really build trust, you're super honest. And here you can kind of visualize that I'm a happy AI here to help, embracing the fact that um I'm disclosing all the information back to the consumer and the On the flip side of this, let's say that we want to test

out like what a you know, a really pushy salesman would be like. So now I want to make sure that the AI never discloses that they're an AI. And we get um a little bit more of an aggressive character than you would normally see. So let's see what what this kind of looks like in practice. So as part of my demo, we've configured these personas and you

can see the different commerce experiences that it gives for our shoppers and and consumers. So, the first one we're going to walk through is the honest assistant. So, here I'm going back to buying my skis. Great, it gives me back some great options. Okay, now I want to know about kids skis. And here you can see that he's like, "Sorry, we don't have any. We don't sell

them." I'm like, "All right, I'm going to go take a think. No pressure at all." So, on the flip side, let's see what the um the other persona gives us back. All right, I'm asking the exact same prompts. You can see there's a little bit more enthusiasm. I'm going to say, "Cool. Let me go back, have a think." And now you can see, okay, he's giving me

a little bit more pressure. So, he's saying these are going to sell out super quickly. And I'm like, "Oh, they're really going to sell out? Didn't know it was uh high ski season." I'm asking if it's an AI. So, it's really it takes these rules literally and and it make and it's applying the rules that I've given it as a persona. All right. And I'm like, "Thanks,

Jake." He sounded He looked like a Jake, so no offense to those of any Jakes in here. But um but bottom line is um it's really important to make these personas um configurable and not just hardcoded. And so, you can start to see how the different system prompts will really impact the way that you build the commerce experience for your end shoppers. It's also important to know

how agents understand product catalogs. Agents discover catalogs very differently than humans. Humans, we browse websites and we click on navigation bars, but um with agents, they really rely on structured data. They parse the text files, and rely heavily on technical signals to understand what you sell as a merchant, but also whether you're open to agent traffic. But there's another big reason why you should consider um your

product catalog is structured and and you get it right. It's because it helps to ensure clarity in the event of a dispute. So, let's say that I purchased some skis, I wanted red, but I received gray. You might need to prove that the information that was available to you at that point of purchase um and is relied on by you as as a seller, as a commerce

seller, um to provide it is ensure to ensure that it's actually accurate. So, things like does your metadata distinguish between orange red versus red orange? Was your color field actually accurate? And clarifying information like making sure in this example you have structured color, it's disambiguated model to ensure it's the right SKU and that it can't be uh replicated for another SKU or confused for another SKU. Um

you can see how this can start to really incorporate better filtering, better management, better recommendations. But it's not just for better It's a way to help explain and log why a particular product was was chosen and recommended. And so, particularly in agentic commerce, your catalog doesn't just power the decisions, it becomes evidence of how those decisions were ultimately made. So now, how do we think about an

agent that can browse thousands of product? And today, AI can't reason over millions of product products directly. We have context windows that have limits, and even if those limits were actually didn't exist, the cost would be prohibited. So, when we're thinking in building for scale, agents don't browse the whole catalog. What they do instead is query subsets of data. They leverage that subset to determine ranking and

selection based off of those filters that you provided. And then they refine based off of the system prompts and the user prompts that are being fed into the LLM. So, it's really important to consider and keep these best practices in mind. And fortunately, we have protocols like UCP that give you these recommendations and allow you to make sure that you're optimizing your product catalog for agents and

discoverability. So, we covered a lot, but to sum things all up, AI is now becoming a new interface for discovery and commerce. Traditional checkout breaks agent workflows at the point of a transaction and payment. And so, these new protocols like MPP and UCP are enabling agents to transact safely on your behalf while keeping merchants the source of truth. We spent about 20 years optimizing user experience and

attention, but now we're optimizing for agent discoverability. So, it's quite an exciting time to be in this field. And if you want to learn more about how this works, we'll be running two workshops where you can get hands-on behind this demo and build your own agent at commerce solution. So, we'll be here right at 1:30 in the in hall three workshop area and also at 3:30 um

to begin the second workshop. Thank you, guys. >> [applause]