Why AI Agents Still Won't Pay for Anything — Jan Curn, Apify

Jan Curn· Apify20:37

Read the talk

Why AI Agents Still Won't Pay for Anything

The title slide states the talk’s central question.Open full source frame
The title slide states the talk’s central question.

Jan Čurn follows agent payments from the need for a budget to discovery and settlement, using Apify’s experience with x402 to show where the friction remains.

From a talk by Jan Curn

At a glance

Ideas worth remembering

  • Existing accounts and API keys already let agents use paid services; autonomous payment becomes useful when an agent needs to discover and buy a new service during a task.

  • A funded wallet does not establish whether an unfamiliar seller will deliver the promised service at the promised price. Trusted relationships and directories are potential sources of that signal.

  • For variable-cost jobs, a seller must handle the interval between authorizing payment and settling the final charge. Apify’s upfront-charge-and-refund workaround shifted trust to the buyer and incurred an extra transaction.

  • Čurn reports that batch settlement addressed Apify’s double-spending concern through an escrow-like arrangement, but the talk does not detail its implementation.

A crowded field of payment protocols

Agents already use paid tools, but Jan Čurn sees little autonomous discovery and purchasing. That gap matters to the Apify founder: at the time of the talk, its marketplace had about 80,000 community-created tools, called Actors, for web data extraction and automation. Čurn says Apify paid their creators almost $2 million in the preceding month. Autonomous purchasing would give those creators another way to reach customers. 0:12

The Apify marketplace slide illustrates the Actors discussed in the opening section.Open full source frame
The Apify marketplace slide illustrates the Actors discussed in the opening section.

Longer autonomous tasks may require an agent to spend without stopping for approval at every purchase. Payment companies have rushed toward that possibility: Čurn names Coinbase’s x402, Google’s AP2, OpenAI and Stripe’s ACP, Visa’s TAP, Google and Shopify’s UCP, and Stripe’s MPP among a much longer list. The companies have reason to compete for transactions on their existing networks; the number of protocols alone says little about how much agents are spending through them.

Čurn sees a particular opening for crypto, despite describing himself as no crypto enthusiast. His case rests on small transaction costs and the absence of card-style chargebacks: a merchant can choose to issue a refund, but a buyer cannot simply reverse a payment in the way he describes for card transactions. He worries that automated abuse could make chargebacks especially costly at agent scale. These are arguments about the payment properties, not evidence that any particular network is already widely used.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

The volume gap and the payment funnel

How much money was actually moving? Čurn presents figures from x402 and MPP transaction-monitoring sites: approximately $1 million on x402 and $1,500 on MPP over the preceding 30 days. The comparison concerns the blockchain transactions those sites monitored at that moment; it is not a measure of all spending by agents or all possible payment channels. 7:14

A shortage of services does not seem to him to explain the gap: payment directories already list many offerings. Čurn estimates that Apify and several other listed providers together earn far more than $1 million a month from agents using their services, principally through existing accounts and API keys. He proposes a funnel to explain why autonomous purchases are rarer: a builder must first need them, set up payments, give the agent a way to discover trustworthy services, and finally let it buy.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:44 · section reference included

The existing competitor is an API key

For web data, a developer can create an Apify account, put its API key in a cloud environment or connect through MCP, and pay the resulting usage on a credit card. The agent can call the services the developer has already chosen. For this workflow, autonomous payments compete with a convenient prearranged account. 9:43

The reason to add a spending mechanism is a different capability: letting an agent find a service it was not given in advance and purchase it during a task. Čurn expects better models and longer-running personal agents to increase that need. He suggests building trust incrementally, perhaps giving an agent $5 for a job, then $50, then $500 as it proves useful. He considers a bounded monetary loss easier to tolerate than some failures already possible when an agent has write access to a system or sensitive data.

The economics of model use also affect the choice between building and buying. Čurn claims that subsidized token access can make it seem cheap for an agent to construct a solution from scratch. If that subsidy diminishes, buying an existing service may become more attractive. Both the size and duration of such subsidies, and their effect on purchase decisions, are his estimates and forecasts.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:14 · section reference included

A wallet still needs a trustworthy seller

Even a builder who wants autonomous payments has setup work to do. Crypto terminology and wallet creation can be unfamiliar, and giving an agent a wallet takes more effort than paying a familiar service by card. Čurn’s proposed simplification is for an agent platform to integrate payments into its existing account: a user might allocate $50 of account credits for service purchases. He presents this as a possible future integration, not a feature established by the talk.

A funded agent then faces a separate question: will a service at an unfamiliar URL deliver the promised result at the advertised price? Finding a weather API says nothing by itself about its reliability. Čurn compares the missing signal to asking a friend for a barber recommendation: trust can travel through an existing relationship.

Directories and marketplaces can provide another signal by acting as intermediaries. Čurn points to reviewed connectors in Claude’s directory, ChatGPT apps, and Apify’s marketplace as examples of places that can help users find services. His proposed paths are a network of known, trusted services or a directory operated by someone the buyer trusts. Neither a listing nor the ability to pay guarantees that every individual transaction will succeed.

12:1212:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:12 · section reference included

From an HTTP 402 response to settled payment

The final step is billing for the work itself. A fixed price per API call can fit a weather forecast, but a compute job or another agentic job using language models may consume an amount of resources unknown at the start. Those jobs need variable, pay-per-use pricing. Čurn says x402 added batch settlement and MPP added sessions to address such billing.

How does the basic x402 payment path work? The client sends an API request and receives an HTTP 402 response with payment requirements. It signs a payment and repeats the request; the server checks it with a facilitator. The facilitator then submits the transaction to the blockchain and reports settlement after confirmation. That sequence separates checking the payment from knowing that money has actually transferred. 17:12

When to do the work is a business decision. If a server starts an expensive compute job before settlement and the payment fails, it can bear the cost without receiving money. Waiting for settlement before performing work reduces that exposure, at the cost of putting confirmation on the path to delivery.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:12 · section reference included

Pay-per-use exposed a double-spending problem

Apify’s experience shows why variable pricing took more than replacing a fixed price with a maximum. Čurn describes three approaches in x402’s progression:

  • exact: The initial scheme charged a fixed amount. Apify could use it to collect money before starting, but it did not directly express an unknown final bill.
  • upto: This allowed a variable charge, but Čurn says it did not prevent an agent from spending the same funds elsewhere before later blockchain settlement. That left Apify exposed while usage accumulated.
  • Batch settlement: Čurn says the later scheme solved that issue by holding funds on-chain in an escrow-like arrangement, then allowing charges for usage. He does not explain the underlying implementation in detail.

Before batch settlement, Apify worked around the problem with exact and a refund. In Čurn’s example, it charges $10 upfront; if the job consumes $2, it returns the remaining $8 when the work finishes. This protects the seller from waiting to collect, but the buyer must trust the seller to refund the balance. It also requires two blockchain transactions, adding fees. Čurn says Apify added batch settlement only shortly before the talk. 19:12

Čurn expects the payment funnel to become practical as these pieces improve and encourages service providers to investigate selling to agents now. His own example supplies the qualification: accepting a protocol can involve decisions about settlement timing, variable usage, counterparty trust, and transaction costs. The infrastructure is progressing; substantial autonomous payment volume remains his forecast.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

18:12 · section reference included

Resources

  • Announcing MCP connectorsArticle

    Čurn explains how Apify Actors can use authenticated external MCP services, including how its proxy handles credentials and permissions.

Read the complete timestamped transcript
  1. 0:12

    Hello, everyone. Uh, welcome to my ses- session. And, uh, today I'm gonna talk about agentic payments and, uh, in particular, the pressing question. Now that we have all the agentic payment protocols, so the big question is like, why is, uh, nobody is using them? Uh, why don't the, why don't the, the agents pay? And before I start, uh, just a quick introduction. Uh, at Apify, we're running a marketplace of, uh, tools for AI. Uh, we call these tools actors. We have about eighty thousand of them,

  2. 0:43

    and, uh, they cover use cases like extraction of data from social media sites, e-commerce, hospitality, travel, search engines, and general like automation, basically. They help you automate, uh, your, uh, work on the web. And these actors are created by a, a global community of creators. And, uh, so they, they, they build them, publish them on our store, and we sell them to the users, to the customers. And, uh, just last month, we paid almost two million dollars to the creators, uh, for their actors, right? So it's like a thriving two-sided marketplace. So

  3. 1:13

    obviously, I think by now you can understand why we take great interest in agentic payments because for us, it's just another way how we can sell our actors, our tools and, uh, our creators to, uh, to, uh, more people and agents, right? And, uh, so obviously, uh, whenever there was a new, new, uh, uh, way how we can integrate our system, our platform, our marketplace, uh, to let agents, agents buy those tools, uh,

  4. 1:43

    we, we, uh, we went fast and added those things. So, uh, actually, uh, last year, uh, at AI Engineer World's Fair in San Francisco, I was the only one who was talking about this top- about the topic of agentic payments and, uh, agentic economy in more broadly. Because back then it wasn't really like, you know, well-known. Uh, not many people talked about it. And I think like over the la- last year and a half, uh, there's been a great progress, right? So, uh, this year, uh, on AI Engineer World's Fair in San

  5. 2:13

    Francisco, there was a dedicated, uh, track, uh, for agentic payments. And now, you know, over the past year, I think like, uh, really everyone understands that, uh, agents, in order to complete ever longer tasks, they will actually need a budget, uh, because, you know, at some point, uh, they will ask you to, to approve something and so on. And if you want them to, to be autonomous and really complete bigger task, you might need to give them money. And, uh, there's been a huge progress over the past year and a half, uh, in

  6. 2:43

    this domain. So, um, I think like Gartner predicts, uh, that by two thousand twenty-eight, uh, this will be like a fif- fifteen trillion dollars economy, and that eighty percent of B2B activity between, uh, between, uh, companies will be done autonomously through agents. I mean, that might be a little too much, but we'll see. But anyway, a lot, a lot of players, uh, in the finance, uh, and finance, uh, in fintech, uh, jumped on this bandwagon because they understood-- they understand the size of this opportunity. So we saw a lot of different, uh, tools

  7. 3:13

    and payment standards, uh, appear over the past, you know, year and a half. So it started with L402 by Lightning Labs. Uh, it was mid-two thousand and twenty. It was super early. And last year in April, Mastercard Agent Pay. Uh, then May two thousand twenty-five, x402 by Coinbase. It's probably the most popular agentic payment standard currently. Uh, in June last year, uh, KYA Pay, uh, by Skyfire by Visa. Uh, we have, uh, Agent Payments Protocol, AP2

  8. 3:43

    by Google. Then Agentic Commerce Protocol, ACP by OpenAI and Stripe in September last year. Uh, in October, Trusted Agent Protocol, TAP by Visa. Uh, in January, Universal Commerce, Universal Commerce Protocol, UCP by Google and Shopify. There's more. Agentic Commerce, Agentic Commerce Trust Protocol, ACTP by Alipay. Uh, Machine Payments Protocol, MPP by Stripe and Temple. Uh, that's from, uh, early this

  9. 4:13

    year. And, uh, from April, Agentic Mobile Protocol, AMP, AMP by Alipay. And, uh, last but not least, Agentic Payment Open Protocol, APOP by UnionPay. Oh, there's more. Uh, OKX Agent Payment Protocol by OKX. So we can see this became a heated, uh, battleground, uh, for the future of agentic commerce because all these like players understand, okay, agents are gonna pay at some point. And, uh,

  10. 4:43

    these companies actually want to, uh, you know, have part of the, of the, of this new future and, uh, preferably to use their networks and existing infrastructure to do that. Oh, one more. Agent Pay for Machines by Mastercard. So anyway, there's players, a lot of these, and it's really hard to keep track, uh, of all the new standards happening right now. Uh, and I think like, uh, this whole agentic commerce became like a huge, uh, uh, huge, uh, well, big news for, for, uh,

  11. 5:13

    crypto people because finally, uh, there seems to be like a real use case they've really been waiting for, for a long time, you know, for crypto, except for, uh, like buying illegal substances and, uh, hiding money away from your spouse. And I actually believe that. Like, I'm not like a crypto person myself. I'm not hodling anything, kind of, any crypto. But, um, I think like, uh, crypto is actually very, very well positioned, uh, to win in this like agentic payments wars because, uh, crypto, like, uh, if done well, is, uh, it doesn't have like single

  12. 5:43

    owner. There's like no single company co-co-controlling the, the ecosystem, right? And, uh, and, uh, it has like properties that, uh, make it very useful for agentic payments. So if you, for example, compare, uh, crypto with like credit cards, PayPal or PayPal micropayments or, you know, bank debit, um, it's much cheaper. Like you just can't like- use, uh, use these traditional payment methods for small agent transaction. It just doesn't make economical sense. And also, uh,

  13. 6:14

    uh, the traditional payment methods, uh, they also have chargebacks. Like, everybody who's selling online understands that, uh, you have to count that, like, uh, some people will just abuse your service and then ask for money back, and it's just like, you know, sort of, uh, way of life for, for, for online businesses. But you can't do that in a, in a agentic, uh, world because there's like infinite number of agents, and, uh, they will just like, you know, steal all your money effectively or, you know, take your service for free. So, uh, on crypto, uh, there is no chargebacks. Uh, it's only like refunds if you... But these

  14. 6:44

    are optional by, by, by the merchants. So I think really like crypto is well suited for agentic payments, and, uh, it is like bright future, um, for that. And actually two... Oh, sorry, two biggest players, uh, uh, or two most popular agentic payment standards currently x402 by Coinbase and Stripe, uh, MPP, uh, on a stablecoin paths are actually like using crypto, right? And, uh, it's not surprising. So let's have a look at the adoption of, uh,

  15. 7:14

    these two systems, uh, in real world. This is from the x402 scan and MPP scan websites, uh, which basically monitors, uh, the transactions on blockchain. And you can see in the last thirty days, uh, in the last thirty days, uh, on x402, there was just only like one million dollars payment volume, and on MPP only like $1,500. I mean, that's basically nothing. It's, uh, it's like nonexistent. So the question is like, where is

  16. 7:44

    everyone? Why is nobody using it? Like, uh, it's been more than, uh, like year and a half now, uh, when agentic pa- payments become a thing, but like agents are still not using it. So let's try to, uh, find answer to that question. So first, maybe there's a low supply of services. Maybe there are just not, not many, not too many things to buy on, you know, for our agents autonomously. Well, I don't think that's the case because if you look, uh, at these like, uh, agentic payment directories, like there's like plenty of services,

  17. 8:14

    you know. And I would say just like our company and Hexa and Perplexity and Farcru are all there, and, uh, I would say we together make far more, you know, uh, than a million dollars per month, uh, from agent, agents using our services, right? So clearly agents are using these tools, pay much more for that, uh, but, uh, through API keys, not, uh, through autonomous discovery and payment. So it must be the demand, right? Like, uh, so clearly the agents just are not interested. They are not buying the services. So let's try to

  18. 8:44

    understand why, okay? And, uh, I think the best way to do that is to look at the, at the agentic payments as a funnel. So it starts with like need. So first, you as the author of the system or creator of the, the agent need to have the need. Like, okay, uh, be- be- because otherwise like you would, you wouldn't allow it or you wouldn't set it up. And then you need to like somehow set it up and once it's, uh, once it's done, uh, the agents need some way to discover the services and eventually buy them, right? And I would say at every step of this funnel, we have some

  19. 9:14

    friction points that ultimately like, uh, lead to the fact that like agents are not buying, um, at the end of the funnel. So let's try to look at it more, in more detail. So first, the need. Well, most people are building agentic systems now, uh, and th- they give them access to services, but they, they, they, they do it, uh, through fixed, uh, accounts or like, you know, accounts, uh, with sign up, uh, they sign up before and through API keys, right? So you just like, I don't know, like you have an agent and if you need to give it

  20. 9:43

    access to web data, uh, you just, uh, create an account on Apify and, uh, and then like put API key into your cloud or, you know, through use, use, use, uh, MCP to connect, uh, to the services. And then so, and then you pay, uh, for, for the use of the services, uh, to your credit card, right? It's pretty straightforward. Everybody can do that, right? So it's, it's really not like that, like agentic payments are, are competing with, with Visa. It's, uh, they are competing with, with API keys because it's just the more convenient way. So what needs

  21. 10:13

    to happen for, for agents or people to really, uh, decide, "Okay, I want my agent not to use only fixed set of services that I, you know, gave it before, but, uh, uh, I want my agent to also discover new services as, as it goes, uh, and buy those on its own," right? So I think there's a few prerequisites to that. First is better models. Um, I mean, as the models are getting better, the agents can complete ever longer task autonomously. You, you trust them

  22. 10:43

    more and so on. So this is sort of like nature progression of things, and I think over time, uh, the models like if, if, if they don't have it already, will, will deserve a trust also with money. And I think like you can start like giving them, I don't know, five dollars for a job and then fifty, and if you know it works, like you can do five hundred. And I would say people already trust agents with, uh, access to their systems, uh, even write access quite often, and sometimes, uh, it leads to disasters or, you know, leakage of data. I think with money it's, it's safer. I mean, like if you lose fifty dollars it's, it's sad, but it's not a

  23. 11:13

    disaster. Like, you know, unless, well, somebody leaks your email, that's much worse. So I, this is already progressing, like models are getting better and, uh, they will have more trust, uh, from people to, to, to do that, uh, to, uh, purchase services. But I think big unlocker will come with, uh, always on agents. Now we saw with like, uh, Meta Muse, you know, OpenAI Dot and so on. I think the personal always-on agents are becoming a real thing, and, uh, if you want those agents to actually do something like you will have to give them budget eventually. So I think we're,

  24. 11:42

    we're getting to this, uh, stage as well. And I think what will also help a lot is, uh, end of the era of token subs- subsidies. Now on hundred dollars per month on Claude you can buy, I don't know, like fourteen thousand dollars worth of tokens if you max it out. Uh, this is crazy because then the agents have no incentives whatsoever, uh, to, you know, to look for existing services because they can build anything from, from scratch because it costs nothing. But I think at some point this will end maybe with OpenAI and Anthropic's IPO, and suddenly the agents will be, okay, maybe it's more

  25. 12:12

    economical for me to buy a service rather than build it from scratch or figure out way, you know, to, to, to get it done. So I think this is like, uh, the unlock is already happening. And then s-second, uh, step in a, in a, in a, in a funnel is set up. Okay, I want to use it now. How, how do I do it? I think for a lot of people, like crypto is still fairly, fairly opaque. You know, there's like a lot of different like lingo. There's like different words and, you know, uh, terms. And like I would say like most people, uh, really haven't tried to set up crypto wallet themselves. Uh, maybe they use like

  26. 12:42

    Coinbase for trading, but really like give their agent crypto wallet, you know. It's, it's, you know, it needs like, you know, some extra work from them. And it's, it's, it's a little more difficult, you know, than, I don't know, using a credit card to pay for a service because everybody has credit cards already. But I think, uh, this, this, this will also like, uh, soon be solved, you know. Either with better ex-better user experience. I think the agentic payment tooling is getting better like every month really. And I think big unlock would come if, uh, these major agentic payment agentic platforms, you know, like

  27. 13:12

    Claude or, or ChatGPT or, you know, these, uh, personal agents, if they actually integrate, uh, crypto payments directly itself, right? So basically you would just like use your main account credits from Claude and say like, "Hey, you can, uh, use fifty dollars to pur-purchase services." I think this would be like a big unlock, and I think it's coming s-sooner than later. All right. So then the agent has a crypto wallet or, or, you know, agentic payment system set up, and it can buy services. So the big question is now, where do, where can it find the services it can trust? And

  28. 13:42

    trust is important because if you're buying a service from some, from someone, you need to make sure, okay, that they will deliver the service they promised they will do, they will deliver for the price they, they, they say they will, right? And if you don't have this trust, you wouldn't transact with anyone. And like historically, how are, how are people building this trust, you know? Like first, like they did start to transact with people they, they, they knew, right? So they like, you know, exchange services and later with money. But it was always someone they knew, so they knew, they knew they can trust and they will sell them good thing. Or

  29. 14:12

    they will transact with someone, uh, someone else, like they knew, uh, and trusted, like, uh, recommended them, right? So it's like, kind of like if you ask, "Oh, um, can you recommend me a good barber?" Like, uh, to a friend. Uh, you'll be like, "Okay, uh, you, you can use this one," you know, and then you can trust the person, and then it, it's all great. And over time, you know, as the world like, you know, got, got bigger, uh, people started to use like a trusted like, uh, middleman, uh, to, to tell you who, who is like good seller and who is not, right? So for example, like different

  30. 14:42

    marketplaces or to-town squares where, you know, if there was some-someone like, uh, selling in a town, uh, you know, on the square, you could probably trust, okay, they are, you know, they are fair. And you know, we have like different sort of like reviews and, uh, and, you know, different agencies, uh, for consumer protection and so on, you know. So basically, these elements like give trust to the, to the, to the economy that you can buy services even from, from strangers because there is like signals you can use. But in the agentic world, there is nothing like that. How can you know that like this agent or this service

  31. 15:12

    on, on some random URL that, that, that, that claims, "Oh, I have API for, I don't know, the perfect weather data." Can you trust it? Well, it's kind of hard, but I think the same principle will apply, you know. So either like network, network of trust that like, oh, I know these services, uh, are good. You know, so basically this is a pool you can use or, uh, basically using trusted directories and marketplaces. And we can already see this happening. For example, uh, Claude, uh, has, uh, has a directory of connectors,

  32. 15:42

    right? It's like sort of like verified connectors that the, you know, that actually like Anthropic like reviews pretty carefully. It's ha- it's hard to get there, you know. Uh, so they provide already this layer of trust for the directory of these tools, and same for ChatGPT. They have apps and, uh, you know, Apify is a marketplace as well. So basically there is a trusted middleman that tells you, "Okay, this is the service you can trust, you know, and you can, you can use it, right?" So I think this is already happening as well. And, um, uh, again, like a big unlock would be like if these

  33. 16:12

    like big agent providers, uh, enabled like payments out of the box, you know. Suddenly, I think the, the, the, the whole thing would really start. And last is to buy, right? And, uh, traditionally these agentic payment systems, they only supported this, uh, super simple, uh, use case, and it was like fixed price per API call. But in this world, this is not sufficient. You know, there's only so, so many things you can buy for fixed price. Maybe like the weather forecast. But to complete a real job, you actually need like, uh, pay-per-use.

  34. 16:42

    I mean, like for example, to, to launch some compute job or some agentic job that uses LLMs it-itself, right? And, uh, and, uh, so you need like variable pricing. Uh, and, uh, thankfully, just, just, uh, uh, earlier this year, x402 added this option. It's called batch settlement. And, um, MPP added s-sessions. Uh, and the goal of these, uh, payment schemes is to enable, uh, agents, you know, to, uh, handle also this like pay-per-use billing. So let's, let's have a quick look on the how, how, uh,

  35. 17:12

    x402 implements pay-per-use because I think this is sort of a nice illustration what the, the early adopters have to go through, you know, to, to, to, uh, make it work. So just a quick look how x402 works. Like first, like, you know, you send API request. You get like four OH two error back. Uh, with some, with some payment, uh, signature, sorry, payment required header. You create a signature, repeat your request with some money. Uh, then the server, uh, verifies it with the facilitator, which might be Coinbase, for example. It's all good. And then it will work there.

  36. 17:42

    Well, we'll see. Uh, then you s-settle the transaction. Uh, Coinbase, the facilitator, submits the transaction to, to blockchain, and transaction is confir-confirmed. You get like, uh, settled response back from the, from the, from the facilitator. And then you know actually the money has been transferred, and you can do the job, right? So the risk is like if you do the job a little earlier and then, uh, the money does-doesn't arrive, like you can lose money. So it depends on your service. Sometimes, uh, you just want to

  37. 18:12

    make sure that, that you charge before you actually do the service if there's like non-trivial cost, maybe like, you know, uh, some, some compute and so on. And then you get payment response. So this is all good, uh, but this is like really just, uh, for the simples- simple scenarios. But ha-ha-handling pay-per-use was pretty much, uh, like a little more challenging, you know? So first, in May, uh, x402 launched a exact payment scheme. This was the first scheme for fixed- fixed- fixed pricing. And then they released a promise that we will release upto

  38. 18:42

    payment scheme, basically allowing like these variable costs. They released it earlier this year, but unfortunately it didn't really work exactly for our use case. I'll get to that very quickly. And then finally, uh, they introduced batch settlements, uh, in 2026. So the problem with the upto payment scheme was that even though, like it allow you to, to charge var- variable cost, there was no- nothing preventing the buyer, the agent, to spend those money twice, you know? Because the transaction was only like settled on blockchain, uh, later. So basically, upto wasn't really useful

  39. 19:12

    for- for, uh, for pay-per-use use case, which was a bit of a bummer. So we work around it by using exact and refund. So basically we charge like sort of like, I don't know, like $10 upfront and then, you know, let's say consumed $2 and o- and once, um, once, uh, the work was done, we refunded the rest. But then again, that means that the agent need to trust, uh, the seller, which again, goes back to the original point that it's- it's- it's difficult sometimes. So, and it, uh, two blockchain transactions, so that means like more fees and so on. So this wasn't really

  40. 19:42

    ideal. So finally, the batch settlement payment scheme, uh, solved this issue. We added it just recently, I think last week. Uh, I don't have time to go through the detail, but, uh, it uses, like, um, crypto magic to kind of like escrow the money on blockchain and then like you can like charge it like, you know, uh, for micro transactions later. And final remarks. Uh, so I think the agentic payments funnel is really being enabled right now, and, uh, soon, uh, it will really like start like a real volumes on the, on

  41. 20:11

    the, on the, and real start of the agentic e- economy. And I think it's a really good time now to, uh, figure how we can sell your services, uh, in this new world because the implementation might be a little more difficult than you think. And if you lay too long, if you wait too long, maybe, uh, you will miss the train. And that's all. Thank you so much.