the value chAIn - what will it look like when AI is processing payments

Aden: Next up is the value chain: what will it look like when AI is processing payments? It's all about how one business cut its monthly invoicing work by 90%, and what agentic machine-to-machine sales could mean for your clients' finances. So please welcome Pete Horder, co-founder of SubSync, and Jillian Chi, Partner Solutions Architect at Stripe. Pete and Jillian, welcome, and thanks so much for being here. We'll let you share your screen and get started.

Pete Horder: Awesome. Thank you very much, Aden. Jules, thank you so much for joining today. As Aden said before, Jules is from Stripe, based out of Singapore, so a slightly early morning for her, but not too crazy. Yes, I'm Pete from SubSync. Great to see some familiar names joining today on the chat. So thanks for joining us for the value chain.

Pete Horder: What we're going to talk about today is probably a little bit different to some of the other sessions you've had over the last day and a half. What we really want to talk about is your clients and the reality for them moving forward. Most of your clients today have a business that generates revenue based on a human interaction. Someone might want to procure a service or a product, off the back of that an invoice is generated, and sometimes somebody pays it manually in the background. What we want to talk about today is the shift to agents. So what happens when it's an AI agent that's initiated a transaction, what that means for our downstream systems, and we'll also look at a customer example of someone who set themselves up really well for this shift.

Pete Horder: So, straight up the top, what we want you to take away today: how this will start to impact your clients, and what to consider when you're engaging with clients who are going down this path of agentic commerce. We'll go through a few examples. We'll talk about how agentic commerce is different to some of the other AI that you've been talking about in the last few days. We'll look at that customer example, and Jules is going to help us look at some really cool solutions from Stripe that will be powering some of agentic commerce moving forward. And lastly, we've got a little takeaway that will hopefully help you understand how close your clients might be, and how prepared they are, for agentic commerce.

Pete Horder: Okay. I said we'll talk about different types of AI. Over the last few days, there's been lots of talk of assistive AI. So predictive, which is pretty straightforward: we look at data in the past, pull that data together and try and forecast what will happen in the future. Generative, which is where we might be drafting emails or proposals, or even slide packs. And quite often what we're seeing now is that the interaction with those types of AI is conversational AI. I actually used it this morning. I was driving into the office, I uploaded my PowerPoint into ChatGPT, and I said, 'Let's talk through my presentation today. What can you tell me about it?' And so I was chatting back and forth with AI about today's presentation.

Pete Horder: The part that we're talking about now is agentic, where we're actually using AI to achieve a goal. It might be to create a purchase order in a system, or it might be to create a maintenance schedule in an ERP platform. The agentic part of that is where we're actually going to allow AI to make financial transactions on our behalf. So, a pretty big change from what a lot of people have probably been working with to date, and obviously the risk profiles associated with it change a little bit as well. If you haven't seen agentic before, I'll quickly hand over to Jules. She'll do a really simple example of what this could look like in practice. So Jules, I'll let you take over.

Jillian Chi: Thanks, Pete. So before we scare people off by saying that the agents are going to be running all these transactions on their own today: while technologically it's possible, I think where we are today, and how ready we are as consumers, we want the agents to actually facilitate the flow, but the autonomy, the authorisation, still comes from the human. So that is the piece that we'll be looking at today. We're not expecting agents to fully automate the entire flow just yet. Wait, oh, I'm not sharing my screen. Sorry about that.

Jillian Chi: So don't be too freaked out when we're saying agentic commerce, agents are going to take over everything, right? No, humans are still going to make the decision, as you can see in my demo here. All right, so for our demo today, we'll be following a fictional company called Dunder Mifflin.

Jillian Chi: Some of you might be familiar with Dunder Mifflin. They're a really strong paper company that really knows what they're doing. They have been selling paper for years and years. And what they realise is that customer service is really important, salespeople being able to sell paper to people who matter. And with the availability of agents today, they're trying to scale this all across the world. They can't have salespeople all over the world, so what they have done is create an agent, Recyclops, to really help them simulate that experience with their users or their potential customers all over the world.

Jillian Chi: So in this case, I'm at Dunder Mifflin's agent, which is Recyclops. There are a couple of prompts here, but in this case today, I'm a customer and I'm running low on paper. So I'll just have a conversation with the agent here, saying that I'm low on paper. What's happening behind the scenes is that the agent will analyse this request and go to the backend system of Dunder Mifflin to see: okay, what is the catalogue available? What can we sell today? And what would actually fit this customer's need? When you're talking to a salesperson, you'll expect more questions. In this case, our agent is not really there yet, so it's shown me a bunch of papers that I do not know what to do with. So I'll just go and tell it that I'm not a paper expert.

Jillian Chi: 'I don't care what I want. I need this for marketing. I want it to pop, because I'm no paper expert. I need you to tell me what is good for me.' So it comes back to say, okay, cool, we want attention, so premium bright is my best bet. I think I'm comfortable with that. 'Let's go with the case of five, then.' In this case, what the agent will do is validate that there's inventory for this specific SKU. It pops out a product card so that I can check out directly from here.

Jillian Chi: Okay, I'll just take a pause here, because some of you might think, okay, this is what I'm seeing on some chatbots on different websites already. And if you're looking at the likes of ChatGPT, they might also be able to provide this experience, where you're looking for an item and they'll pop up a few of these product cards for you. But what is different here is that instead of bringing the customer out to a separate checkout page, like dundermifflin.com or nike.com or something like that, which creates obstacles and drop-off, what's happening here is that the agent is able to provide the entire checkout experience in the UI itself. And in this case, if I choose to check out, they'll be able to bring the checkout interface itself directly here.

Jillian Chi: You might also be thinking, okay, in this case, is the agent then able to see all my payment information? Is that really safe? Should I be concerned? But the fact of the matter is that the agent is actually not seeing any of your payment information. Everything that's within this component sits with Stripe. It's a direct communication between the customer and Stripe itself. The agent has no visibility of whatever's inside here. So in this case, if I already have my Link wallet set up... for those of you who are not familiar with Link, it's a digital wallet by Stripe. What it means is that if I have previously shopped with any of the merchants that use Link, I'll be able to check out with this other merchant directly with Link.

Jillian Chi: So in this example, I have already had a subscription with OpenAI. OpenAI allows customers to check out with Link. And so when I come to Dunder Mifflin, who I have not checked out with before, my payment information and my address information would already be saved there, so it will essentially just be a one-click checkout from here. It needs to verify that this is me, so I'll just put in an OTP and voila, my payment information and address are all in there. All I need to do is click on one button to check out, and we're good to go.

Jillian Chi: So this is essentially the shift that we're seeing with agents, where it's not just able to give you good recommendations based on your needs, but also to facilitate the entire checkout process, so that we are cutting down the obstacles to conversion, bypassing all the redirects that we would typically need when you're checking out on a usual website. So that is agentic commerce as we have it today. So, back to you, Pete.

Pete Horder: Thanks, Jules. Really cool, simple demo. But I guess from here, a lot of you are probably sitting there going, well, what does that actually mean for my clients? They sell something. What does that mean on a day-to-day basis? How's it going to impact you if you're a bookkeeper, what's it going to do to their coding, and your accountant? What are their accounts going to look like if you've got these bots or agents out there making transactions with your clients?

Pete Horder: So, just a slight bit of rewind around generation of revenue. If you look at an invoice back in 2016, very simple sort of thing: they'd probably do some services through the month and generate an invoice at the end of the month, by a human. You'd probably assume you've got, per customer, 12 transactions to reconcile in a year. Fast forward to now, and we know that lots of the commercial models have evolved. People who are selling services are also selling products and subscriptions, which means that from an accounting and bookkeeping perspective, there's a lot more happening in that integration and from a coding perspective, and it's sometimes a lot more challenging to work out what is happening from an accounting perspective.

Pete Horder: Going into 2027 and beyond, what we're really looking at is, okay, those things will continue to happen, but in addition to that, we'll start to see things like agents making purchases. And then we'll also see things like outcome-based billing. So once an event has been completed... and this could be digitally, right? So once something happens online, an outcome is reached, and a transaction is completed. So, lots of different things that will be happening moving forward. What I want to talk about is what considerations there are. What does that mean for customers? There's a number of different things that we're looking at here.

Pete Horder: So, you want to know who you are dealing with. Can the customer be verified? Is this a real person? Where do they exist? What do you need from them from a tax perspective? If it's an agent, do they actually have permission to buy? Do your clients actually want to sell something to an agent if that agent hasn't been authorised to make that purchase? Are you going to get paid? You want to make sure that money's coming through, that the payment method's legitimate, and it's not a fraudulent transaction. If there's a dispute, or you need to make a credit, how's that going to work, and what's that going to do for you as well?

Pete Horder: Is the purchase correct? So, is the integrity of the purchase correct? Has an agent managed to find an unstocked item on a website and purchase that somehow? Or if it's an outcome-based billing activity, has that outcome been reached and verified before you actually take the payment?

Pete Horder: And of course, the integration bits. This is where I sit a lot of the time, and lots of things that I concentrate on. Do you need this payment to be integrated into your ERP system to update stock levels? Is there a CRM that needs to get updated? Obviously the financial product needs to understand the numbers and where that revenue's coming from, and the tax. And from an accounting perspective, where do you want all this to go? Do different products go to different accounts in the chart of accounts? Do you need to put some sort of reporting categories around this to make it easier to report on downstream as well?

Pete Horder: So it's not that these challenges are drastically new, but one of the things is that the speed, the volume and the complexity can increase quite a bit. But the positive thing with all this is we've seen technology evolve, and businesses are going to be able to manage this complexity with some level of confidence. This isn't a vibe-coded exercise. These are platforms that are delivering real solutions. So again, I'll pass back to Jules, and she'll give us an example of one of the pieces of technology that we'll be able to use moving into the future.

Jillian Chi: Thanks, Pete. Just let me share my screen. Okay, so we'll go back to our previous example of Dunder Mifflin. As we saw earlier, Dunder Mifflin's agent was really good at helping to understand the customer's requirement and making recommendations, and they really have deep insights around their product, paper. So as they realise that their conversion is going up, and they're having less dependency on salespeople, what they realise is: okay, right now I don't have to just sell paper. I can also be selling this agent to, say, the other paper companies, because they will also benefit from it. So what they have done now is decide to sell their agent to other paper companies, so that their customers in turn can leverage this agent. But then there are new questions that will come up when you're trying to monetise, or gain revenue from, this agent.

Jillian Chi: How do you charge? What exactly do you charge them for? I think with a lot of the agentic platforms that we're seeing today, one common way of monetising it is through subscriptions. With subscriptions, you're just paying a flat fee at certain cadences, like monthly or weekly. But the thing about agentic use cases is that they're very much reliant on tokens. And so if your usage is higher, typically your costs will also be higher. So if Dunder Mifflin is selling their agent, and they're doing it based on a flat fee, it gets a bit difficult for them to manage their infrastructure. Their customers, on the other hand, might also be complaining: 'I might not be using as much, and I'm being charged the same fee as other users.' Or, 'The outcome that I'm getting from it is not exactly what I want, so why am I still getting the flat fee?'

Jillian Chi: And thirdly, with B2B cases like this, something that's very prevalent is the different relationships that you have with your customers. In this case, Dunder Mifflin is selling their agent to different paper companies; we have North Star and we have Coastal. And because I have different relationships with these customers, the kind of contract and the kind of pricing that I want to have for them will also be different. In this case, if I don't have a very good relationship with North Star, they can just go with a standard contract. I'll just charge them a flat fee per month, and for every additional positive outcome, which is an order being placed directly by the agent, I can charge them, like, 15 cents per outcome. And I might throw in two orders for free each month.

Jillian Chi: But for someone like Coastal, maybe I want to have a little bit more flexibility with them. I do like them. They are also one of our suppliers, so perhaps I don't want to negotiate the flat fee, but I'll give a 20% discount on the per-outcome fee. And also, because we have other business relationships, I might want to give them additional credits to burn down from, so that instead of me paying them directly, I want them to use these credits on my platform, so I get to earn back from it. So while Dunder Mifflin is able to build all this up from scratch by spending their time and resources to help monetise their agentic flow, they can also choose to use services like Stripe and Metronome to help abstract away all this complexity, so that they can focus on what matters most to them, which is selling paper.

Jillian Chi: Alright, so how it would look on Metronome is that for each of the customers they have, we will have a set contract set up for them. So here we still have North Star, and we still have Coastal. You can see the contracts look similar for the most part: these are monthly statements, and I want to give them net payment terms of 30 days. All of these are configurable per contract, per customer. But what you're able to see differently is that here we not only have the monthly subscription enabled for $50, I'm also able to set up 20% off whatever I'm charging for a standard contract. So it gives me a lot of flexibility to change up the pricing model and the contracting for each of these customers.

Jillian Chi: And say for the credits that I was talking about earlier: in this case, Coastal, let's see how their credit is doing. I've given them a dollar credit before. It seems like it has been used up, and they're due for more credits from me. So what I can do is just go into their profile and add a credit from there. Again, all of this is available through APIs; we're just doing the demonstration through the dashboard. But what you can see here is that I'm allowing them spend-based credits. Okay, we're more generous today, we'll give them a $5 credit. It should be used before the 20th, and I'll give it high priority. What this means is that the next time they're using my agent, whenever there's a successful outcome, instead of me having to bill them directly for it, it will burn down from this credit that I have provided them.

Jillian Chi: Great. So coming back to this example here, let me break down what this UI is about. Earlier, when we were looking at the Dunder Mifflin agent, that is the front end, fronting the customer. But this one is for me to see the successful orders that have come in from these different businesses.

Jillian Chi: So if I run a batch order just to simulate what is happening, you can see that for both North Star and Coastal, their customers are checking out with different items as they go along. Some of it is being accepted, great. But some of it is not a successful outcome, for various reasons. It might be an inactive SKU. It might be cases where the paper size is unspecified for whatever reason, and it will need a human review for that. And what that means for our agent is that we're not directly charging them for it, because they're not getting a conversion; they're not getting an order directly from our agent.

Jillian Chi: What happens here is that for the 10 orders that we're seeing, very similarly for North Star and Coastal, although it's different items being checked out, we have six successful outcomes and four that have been excluded. So, great: the customer is not paying for anything that we have not processed successfully. And all the successful events go to Metronome. If we head back to Metronome, we'll be able to see under their invoices that all these events are being accounted for. This screen might be a little bit small, so you might want to view it in full screen, because I'm splitting the screen here just for us to get good visibility of the differences between a standard contract and a more negotiated contract.

Jillian Chi: But as you can see, based on the different prepaid outcomes that we have, the total being charged to the customer at the end of the day will be different. And all of this does not need to be managed manually by Dunder Mifflin. It's all done very automatically, systematically, through Metronome itself. And once the invoice is due for payment, since this is integrated directly with Stripe, what you'll be able to do is view it on Stripe, which I've opened here already, and send the invoice direct to the customer so that the payment can be received. If their payment method is already saved on Stripe, this payment will be cleared and charged automatically at the end of the month.

Jillian Chi: But if I'm choosing to send them an invoice and have them pay following their payment terms, this can also be orchestrated by Stripe. In this case, for North Star, I've sent them an invoice once before, and I want to resend it. Let me just resend it again, to remind them that they have an invoice due, and they should be able to receive the invoice in their inbox. There you go. This is a 30-day payment term, but because the invoice has already been due 21 days, it has been updated as such. And all the information, all the breakdown of whatever usage has been made, will be detailed in the invoice automatically.

Jillian Chi: If I choose to pay the invoice here, this will redirect me to a Stripe-hosted invoice page, where I'll be able to see what I'm paying for and also the payment methods. Something cool about our hosted invoice page is that if you're selling to customers all over the world, we have what we call adaptive pricing. Right now I'm sitting in Singapore, and Dunder Mifflin's account is in the US. So while Dunder Mifflin will want to accept the payment in USD, I want to pay in SGD for whatever reason: my card has better rates when I'm paying locally, or I just prefer to pay in my local currency. So in this case, as a customer, I have the flexibility to choose if I want to pay in USD or SGD, and complete the payment directly in the Stripe-hosted invoice page.

Jillian Chi: So to sum things up: with these new business models coming up, a lot of additional complexity has been added to being able to monetise and generate revenue from them. The best way for you to go about it, rather than spending all this time and resources building things out from scratch, is to rely on people who do it best, to cover the end-to-end flow for you, from billing, the metering, the pricing models, the contracting, and even receiving the payments, using both Metronome and Stripe, so that you can focus on the growth of your business. Yeah. So this is a general flow of how things will work, but back to you, Pete, for more.

Pete Horder: Thanks, Jules. I'll just jump back to my slide deck. Sorry. There we go. Yeah, thanks, Jules. A great example of how some really complex and somewhat sophisticated pricing models can be built. This is the sort of thing that you see if you've got a Claude subscription and you're paying for tokens or anything. It's actually Metronome, which is a Stripe product, that sits in the background and helps power all those invoices that you get each month for all your tokens and so on.

Pete Horder: But what I want to do now is think about what that means downstream. It's great to have all this flexibility and be able to generate all these different prices and discounts, and issue credits and draw down on those credits and everything, but you've somehow got to account for that. You've got to integrate into your financial platforms, and then how do you manage credits and so on as well?

Pete Horder: We're going to look at a customer example. Choco are actually a platform based out of Berlin. They sell software that is mostly paid for by subscription at the moment. That software helps manage food warehouses and distributors. They've got everything from small family businesses to larger distributors, so because of that, they've got a lot of different pricing models. They went through a re-implementation of Stripe recently to cater for all these different types of pricing models. In addition to that, they also had lots of things managed regionally, so they were trying to do different things in different accounting platforms and ERPs in different regions across the world.

Pete Horder: And the net outcome of that is that each month they were spending hours and days generating invoices, working out what they needed to charge customers, and then on the back end of that, their finance processes to understand how they should be accounting for it were equally as complex.

Pete Horder: So what they've done is re-implement Stripe, which manages the subscriptions; they do usage-based billing, and it manages all the recurring payments. SubSync is then the glue that translates all the information that's happening into their financial system. It's real time. There are rule-based allocations, so these are defined up front; it's not making them up as it goes. It's repeatable, so it's that same process followed over and over again, so you know exactly what's going into your accounts and when it's going into your accounts. It manages errors. And as I said before, it'll manage credits and refunds and all the other sorts of things that generally catch people out, and tax as well. Managing your tax and how your tax gets into your financial system, which in this case is Xero, can be incredibly complex. So they've now got a single chart of accounts in Xero, and they've got accurate reporting.

Pete Horder: Overall, there was a 90% increase in their finance operations. Over 150 manual hours saved each month. So a really good outcome for now. But again, they're just working on a subscription-based and usage-based model at the moment. The thing that I want to pose today is: hypothetically, what happens if they move to some form of agent billing or outcome-based billing in the future? This could mean that their clients get more value, and they've been able to increase their customer base, because what they're providing is of higher value at lower cost. But what will that mean for the downstream systems?

Pete Horder: So if we look at the solution, thankfully, it's actually the same. They've prepared themselves in a way that they've got this flexibility: they can work with Stripe on the front end to create all sorts of different, complex and varied scenarios to go to their customers and gain revenue, but in the background, everything remains the same. There might be some configuration changes along the way, but you don't have to fundamentally change your whole infrastructure in the background to cater for a growing and changing market. In Stripe, you might need to start to consider what happens with payment tokens, but overall it's relatively much the same.

Pete Horder: So I guess from here, you're probably going, wow, there's a lot of different things that I haven't looked at, potentially. Maybe some of you are quite advanced, or quite a way down this. But how do we know if your clients are going to be ready for this technology and this sort of step forward? What we wanted to do is really just give you three quick things that you can do now to be ready.

Pete Horder: And we've got a free assessment that you can do at the end for each client, to understand where the gaps are, what they might need to do, and what areas they need to improve in. So I think a really good thing to do now is, one, identify a client who's already transacting online.

Pete Horder: They will have the opportunity in the future to transact differently, to engage agents and sell to agents. And then go and have a look at the chart of accounts. Have they been managing their chart of accounts themselves, and has it got a little bit unruly? Have they got multiple instances of Xero, or QuickBooks or something, set up? Is there a bit of a mess in the background that needs to be cleaned up a little bit? And then the other part is to look at that plumbing, which is kind of the boring bit in the background, but making sure that there are no manual reconciliations, and not having to do monthly journals, output reports and so on. Remove all those complexities that are going to take more and more time in the future.

Pete Horder: That is pretty much all we have for today. We've got plenty of time left, which is great, so hopefully we can do some Q&A if there are any questions. I mentioned we've got an agentic commerce readiness score that you can get for your clients, so if you use that QR code there, you can share it with your clients. There's 20 questions; it'll help identify where in their process flow they might have some challenges in the future, once we get to agentic commerce, and give some ways and suggestions on things they can do to improve that in the meantime. So, Aden, I'll pass it back to you, just to see if we've got any Q&A. We've gone through this a bit quicker than I thought, but that's okay.

Aden: Well, thank you all, that was awesome. Thanks, Pete and Jillian. We really appreciate it. And thank you, audience, for joining us for the AI in Practice Summit.