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

AI Engineer · 20 min · 189 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

  1. 00:12Hello everyone.
  2. 00:15Welcome to my session , and today I'm going to talk about agency payments, specifically a hot topic since we have all the patent payment protocols.
  3. 00:24So the main question is: why doesn't anyone use them?
  4. 00:29Why...why don't agents pay?
  5. 00:31And before I start, a quick introduction : at API, we run a marketplace of AI tools.
  6. 00:39We call these tools " actors"; We have about 80,000 of them, and they cover use cases such as data extraction from social media sites, e- commerce, hospitality, travel, search engines, and automation
  7. 00:50in general.
  8. 00:51Essentially, they help you automate your online work, and these actors are created by a global community of creators.
  9. 00:59And they, they, they create them, publish them in our store, and we sell them to users, to customers.
  10. 01:05And just last month, we paid almost $2 million to creators for their actors, right?
  11. 01:10So it's like a moving two-way market.
  12. 01:13So obviously, I think you can see now why we're so interested in magic payments, because for us, it's just another way that we can sell our actors, our tools, and our creators
  13. 01:26to more people and agents, right?
  14. 01:31And, uh, so obviously, whenever there was a new, uh, way to integrate our system, our platform, our marketplace to allow agents to buy these tools, we were quick to add those things.
  15. 01:47So, uh, actually, uh, last year, uh, at the AI ​​ Engineer World Fair in San Francisco, I was the only one who spoke on this topic, on the topic of Asian payments and, uh, the agent economy in a broader
  16. 01:59sense, because back then it wasn't really, you know, well known, uh; not many people talked about it.
  17. 02:04And I think in the last year and a half, uh, there's been a lot of progress, right?
  18. 02:09So, uh, this year, uh , at the AI ​​ Engineer World Fair in San Francisco, there was a special, uh, track , uh, for agency payments.
  19. 02:18And so, you know, over the last year, I think everyone has realized that Asians are actually going to need a budget to do increasingly long-term tasks, because at some point they're going to ask you to approve something and so on.
  20. 02:32And if you want them to be autonomous and actually do bigger tasks, you may need to give them money, and there has been tremendous progress in this area in the last year and a half .
  21. 02:44So, I think Guard is predicting that by 2028, it will be a $15 trillion economy, and that 80% of B2B activity between companies will be done autonomously through agents.
  22. 02:58I mean, it might be a little too much, but we'll see.
  23. 03:01But in any case, a lot of players in the financial sector and the fintech financial sector have joined this bandwagon because they understood the scale of this opportunity.
  24. 03:10So , over the last year and a half, we've seen a lot of different, uh, tools and payment standards emerge .
  25. 03:17It all started with the L42 from Lightning Labs.
  26. 03:20It was mid-2020, it was very early, and last year in April, Mastercard's agent payments came out , then in May 2025, X42 from Coinbase.
  27. 03:30This is probably the most popular agency payment standard at the moment.
  28. 03:35Last June, uh, KYA pay, uh, Skyfire from Visa.
  29. 03:39Uh, we have, uh, Asian Payments Protocol, AP2, from Google.
  30. 03:44Then the Agent E-commerce Protocol, ACP, from OpenAI and Stripe in September last year.
  31. 03:50Uh, in October, Trusted Agents Protocol, TAP, from Visa.
  32. 03:54Uh, in January, Universal Commerce Protocol, Universal Commerce Protocol, UCP, from Google and Shopify.
  33. 04:00There is even more agent commerce: Agent E-commerce Trust Protocol, ACTP, from Alipe; uh, Machine Payments Protocol, MPP, from Stripe and Temple.
  34. 04:11Uh, this is from the beginning of this year.
  35. 04:14And, uh, since April, Asian mobile protocol, MMP, AM from Alipe.
  36. 04:20And, uh, last but not least: Asentic Payment Open Protocol, APOP by Union Pay.
  37. 04:26Oh, that's not all.
  38. 04:27Uh, OKX, OKX's agent payment protocol.
  39. 04:32So you can see that this has become a hot battleground for the future of agent e-commerce because all of these players understand that agents are going to pay at some point, and these companies really want to be a part of this new future, and, uh, preferably leverage
  40. 04:48their networks and existing infrastructure to do so.
  41. 04:52Oh, another agent payment machine from Mastercard.
  42. 04:56So, there are a lot of players, and it's very difficult to keep track of all the new standards that are emerging now.
  43. 05:03Uh , and I think this whole agency business has become huge, uh, huge, uh, well, big news for cryptocurrencies because finally, uh, it seems like there's a real use case that
  44. 05:17they've been waiting for, you know, for cryptocurrencies, other than, uh, buying illegal substances and hiding money from your spouse.
  45. 05:26And I actually think that, like, I'm not a crypto lover myself, I'm not hiding anything, I do n't own any cryptocurrency, but, uh , I think that cryptocurrency actually has a very, very good chance, uh, of winning
  46. 05:38this agency fee war because, uh, cryptocurrency, like, if done right, it doesn't have a single owner, there's no one company that controls the ecosystem , right?
  47. 05:47And, you have, and, uh, it has properties that make it very useful for agency payments.
  48. 05:53So if you compare cryptocurrency, for example, to credit cards, PayPal or PayPal micropayments, or, you know, bank debit, it's much cheaper, you just can't use these traditional payment methods for small agency transactions.
  49. 06:09It simply does n't make economic sense.
  50. 06:11And also, uh, traditional payment methods, uh, also have chargebacks, and anyone who sells online understands that, uh, you have to consider that some people will just abuse your services and then demand a
  51. 06:24refund, and that's just, you know, a way of life for an online business.
  52. 06:29But you can't do that in the agency world because there are an infinite number of agents and they just want to steal all your money or take your services for free.
  53. 06:39So, uh, there are no chargebacks in cryptocurrency , only refunds, but they are optional for sellers.
  54. 06:46So I think cryptocurrency is really well suited for agency payments, and that's, uh, a bright future for you.
  55. 06:53And actually the two, sorry, two of the biggest players, uh, or the two most popular agency payment standards that are currently being explored, Coinbase and Stripe, uh, mpp, uh, on stable coin pass, are actually
  56. 07:05similar to using cryptocurrency, right ?
  57. 07:09And this is not surprising, so let's look at the implementation of these two systems in the real world.
  58. 07:17This is from the websites X42 scan and MPPP scan, which basically track transactions on the blockchain.
  59. 07:25And you can see that in the last 30 days, uh, in the last 30 days, uh, on X42, the payment volume was only about $1 million , and on MPP, only about
  60. 07:37$1.5 thousand.
  61. 07:38I mean, it's practically nothing.
  62. 07:41Uh, it's like it doesn't exist.
  63. 07:42So the question is, where is everyone ?
  64. 07:44Why doesn't anyone use this?
  65. 07:45It's been over a year and a half , uh, since agency payments became a reality, but agents still aren't using them.
  66. 07:55So let's try to find the answer to this question.
  67. 07:59So, firstly, there may be a low supply of services.
  68. 08:02Maybe there just aren't many things for agents to buy autonomously.
  69. 08:06Well, I don't think so, because if you look at these agency fee directories, there are a lot of services out there, and I would say our company, Exa, Perplexity, and Farcru–they're all there, and I would say that we collectively make a lot more, you know, than a million dollars a month from agents using
  70. 08:25our services, right?
  71. 08:27So it's clear that agents are using these tools, paying a lot more for it , but through API keys, not through autonomous discovery and payment.
  72. 08:36So it must be demand, right?
  73. 08:38So it's obvious that agents are simply not interested, they're not buying the services.
  74. 08:42So let's try to understand why, okay?
  75. 08:45And I think the best way to do that is to think of payments as a sales funnel.
  76. 08:49So it all starts with a need.
  77. 08:52So first you, as an external participant in the system or the creator of the agent, need to have a need, like, okay, because otherwise you would n't allow it or
  78. 09:01configure it, and then you need to configure it somehow, and once that's done, once that's done, agents need some way to find services and eventually buy them, right?
  79. 09:10And I would say that at every step of that funnel, we have some friction points that ultimately lead to agents like that not buying, um, at the end of the funnel.
  80. 09:20So let's try to look at this in more detail.
  81. 09:24So, first, the need, well, most people now create agent systems, uh, and they give them access to services.
  82. 09:31But they, they, they, they do it through fixed, uh, accounts or, you know, accounts with registration, uh; they register earlier and through API keys, right?
  83. 09:40So you just, like, I don't know, you have an agent, and if you need to give them access to web data, you just create an account on AIfy and, like, then, like, put an API key in your cloud or, you know, through the MCP to connect to the services, and then you pay for the
  84. 09:57services through your credit card, right?
  85. 09:59It's pretty simple.
  86. 10:01Anyone can do it , right?
  87. 10:03So it's not really like agency payments compete with Visa.
  88. 10:09They compete with API keys because it's simply a more convenient way.
  89. 10:13So what needs to happen for agents or people to really , like, decide, " Okay, I want my agent to not just use a fixed set of services that I, you know, provided them with before, but, like, I want my agent to also
  90. 10:26discover new services as they go along and buy them on their own," right?
  91. 10:32So I think there are several prerequisites for this.
  92. 10:35Firstly, these are better models.
  93. 10:37What I mean is that as models improve, agents can perform each longer task autonomously; you trust them more, etc. So it's kind of a natural progression, and I think that over time, uh, models, if they don't already have it, will
  94. 10:51be trustworthy, with money too.
  95. 10:53And I think you can start by giving them, I don't know, $5 for a job, and then 50, and if you know it works, you can do 500.
  96. 10:59And I would say that people already trust agents, uh, access to their systems, uh, even the right access, quite often, and sometimes, uh, it leads to disasters or, you know , data leaks.
  97. 11:08It's safer with money.
  98. 11:10I mean, for example, if you lose $50, that's sad, but it's not a disaster like you have, unless someone posts your email, which is much worse.
  99. 11:18So, I think it's already progressing, the models are getting better, and people will have more trust in them to buy services.
  100. 11:25But I think the big unblocker will come with the always- available agents that we've seen, like with Metamuse, you know, Open Eye Dot, etc. I think personal, always-on agents are becoming a reality, and if you want those agents to actually do anything, you're going to end up giving them a budget.
  101. 11:42So I think we are reaching that stage as well.
  102. 11:45And I think the end of the era of token subsidies will also help a lot.
  103. 11:50Now for $100 a month on Claude you can buy tokens worth, I don't know, about $14,000 if you spend the maximum.
  104. 11:57This is crazy because then agents have no incentive to look for existing services, since they can create anything from scratch because it costs nothing.
  105. 12:06But I think at some point it will end, maybe with an open AAI rather than an entropic IPO, and suddenly agents will be okay, maybe it will be more economical for me to buy the service than to build it from scratch or find a way to implement it.
  106. 12:19So I think it's kind of like the unlocking is already happening, and then the second step is the sales funnel is set up.
  107. 12:25Okay, I want to use this now.
  108. 12:27How do I do this?
  109. 12:28I think that for many people, cryptocurrency is still quite opaque.
  110. 12:32You know, there are a lot of different terms, there are different words and, you know, terms.
  111. 12:35And, I would say, most people, uh, have n't really tried to set up a crypto wallet on their own, uh.
  112. 12:41Maybe they use something like Coinbase for trading, but really, if their Asian crypto wallet, you know, it, you know, requires some extra work on their part , and it, it, it's a little more complicated, you know, than, I don't know, using a credit card to pay for services, because everyone already has credit cards.
  113. 12:57But I think this, this, this too, well, will be resolved soon, or with a better, better user experience.
  114. 13:02I think agency payment tools are getting better every month, actually.
  115. 13:06And I think the big unlocking will happen if these major agent payment platforms, you know, like cloud technologies or JGPT, or, you know, these personal agents, if they actually integrate crypto payments
  116. 13:18directly, right?
  117. 13:19So essentially, you would just want to use your main account with Cloth and say, "Hey, you can use $50 to buy .services."
  118. 13:26I think this is going to be a big unlock, and I think it's going to happen sooner rather than later, okay?
  119. 13:31So, the agent has a crypto wallet or, you know, an agent payment system set up, and they can purchase services.
  120. 13:37So the main question now is: where, where can he find services he can trust?
  121. 13:42And trust is important because if you're buying a service from something, from someone, you need to make sure that they're going to provide the service that they promise, they're going to provide it for the price that they, they, they, they say they're going to provide it, right?
  122. 13:53And if you don't trust, you won't transact with anyone .
  123. 13:56And historically, how do people build this trust?
  124. 13:59First they started transactions with people they , they, they knew, right?
  125. 14:03So they, you know, exchanged favors and then money, but it was someone they knew.
  126. 14:08So they knew they could be trusted and they would sell them a good thing, or they would trust someone, uh, someone else that they knew and trusted who recommended them.
  127. 14:18So it's like you asked, "Oh, um, can you recommend a good hairdresser for me?"
  128. 14:22For example, to your friend, you say, " Okay, you can use this," you know?
  129. 14:27And then you can trust this person, and everything will be great.
  130. 14:31And over time, you know, as the world got bigger, people started using a trusted intermediary to tell you who's a good seller and who's not, right?
  131. 14:42For example, different markets or city squares , where, you know, if someone was selling in the city, you know, in the square, you could probably trust: " Okay, they're
  132. 14:52honest."
  133. 14:53And , you know, we have different reviews and different consumer protection agencies and so on.
  134. 15:00So, essentially, these elements lend credibility to the economy that you can buy services from strangers because there are signals that you can use.
  135. 15:07But in the world of aging, there is nothing like that .
  136. 15:09How can you know that, for example, this agent or service at some random URL that claims, “Oh, I have an API for, I don’t know, perfect weather data,” can you trust it?
  137. 15:19Well, it's quite complicated, but I think the same principle would apply, you know?
  138. 15:24So, either, like, network, network, or trust, like, "Oh, I know these services are, uh , good," you know?
  139. 15:29So, essentially, it's a pool that you can use, or, uh, essentially, use verified catalogs and marketplaces.
  140. 15:36And we already see this happening, for example, uh, cloth, uh , has a connector catalog, right?
  141. 15:42These are kind of like verified connectors that are actually, like, entropic, like, checked pretty thoroughly.
  142. 15:48It's hard to get there, you know.
  143. 15:50Uh, they already provide this level of trust for the catalog of these tools, and the same...They have apps, and, uh, you know, Appi is also a marketplace.
  144. 15:59So, essentially, there's a trusted intermediary that tells you, "Okay, this is a service you can trust," you know, and you can, you can use it, right?
  145. 16:06So I think it's already happening.
  146. 16:08And, uh , again, the big unlock would be if these big agency providers, uh, enabled payments out of the box, you know?
  147. 16:17Suddenly, I think this would all really start.
  148. 16:20And the last thing is to buy, right?
  149. 16:22And you , traditional, these agent payment systems, they only supported this, uh, super simple, uh, use case, and it was kind of a fixed price per API call.
  150. 16:31But in this world, that's not enough, you know?
  151. 16:33There are only a certain number of things that can be bought for a fixed price, perhaps the weather forecast.
  152. 16:38But to do real work, you actually need to, for example, use a payer to run some kind of computational task or some kind of agent task that the LLM itself uses , right?
  153. 16:48And, uh, you need variable pricing.
  154. 16:51And, uh, fortunately, just, uh, earlier this year, X42 added this option; it's called batch calculation , and MPP added sessions, uh.
  155. 17:02And the purpose of these, uh, payment schemes is to enable, uh, agents, you know, to also process this, like paper billing .
  156. 17:09So let's take a quick look at how, uh, x4 implements the use of a payer, because I think it's a good illustration of what the early adopters went through, you know, to make it work.
  157. 17:19So let's just take a quick look at how the X42 works.
  158. 17:23First, as you know, you send an API request.
  159. 17:25You get a 402 error, uh, with some payment, uh , signature, sorry, header " payment required."
  160. 17:31You create a signature, repeat the request with a certain amount of money, then the server checks it with an intermediary, which could be, for example, Coinbase.
  161. 17:39Everything is fine, and then you work there .
  162. 17:42Well, we'll see, uh, then you settle the transaction, uh.
  163. 17:47Coinbase , the intermediary, sends the transaction to the blockchain, and the transaction is confirmed.
  164. 17:54You get a response from the intermediary, and then you know that the money has actually been transferred, and you can get the job done, right?
  165. 18:02So the risk is that if you get the work done a little early and then the money just doesn't come through, you could lose it.
  166. 18:08So it depends on your service.
  167. 18:10Sometimes, uh, you just want to make sure you charge before you actually provide the service, if there are non-trivial costs, maybe you know, some calculations, etc., and then you get a response about the payment.
  168. 18:21So that's all good, uh, but it's really only for simple scenarios, but handling the use of the payer was quite, uh, a little more complicated, you know.
  169. 18:30So first in May, uh, exports launched a precise payment scheme.
  170. 18:34This was the first scheme for fixed pricing and then they promised that we would release a payment scheme up to 100% which would basically allow for variable costing.
  171. 18:45They released it earlier this year, but unfortunately it didn't quite work for our use case.
  172. 18:49I'll get to that quickly.
  173. 18:51And then finally, uh, they implemented the basic settings, uh, in 2026.
  174. 18:55So the problem with the 100% payment scheme was that while it allowed for a fairly variable cost, there was nothing to stop the buyer, the agent, from spending that money twice, you know, because the transaction was
  175. 19:08only settled later on the blockchain.
  176. 19:09So, essentially, the up to 10% scheme wasn't useful for the pay-per- use case, which was a bit of a bummer.
  177. 19:16So we get around this by using exact and refund.
  178. 19:19So basically, we charge something like, I don't know, $10 up front, and then, you know, let's say, we consume $2, and once the job is done, we give back the rest.
  179. 19:30But again, that means agents have to trust the seller, which again brings us back to the original point that sometimes it's complicated, so you have to transact through blockchain, which means more fees and so on.
  180. 19:41So it wasn't exactly ideal.
  181. 19:42So finally the rate-based payment scheme solved this problem.
  182. 19:46We added it recently, I think last week.
  183. 19:49I don't have time to go into the details, but it uses cryptomagic to sort of escrow money on the blockchain, and then you can charge it, you know, for microtransactions
  184. 20:00later and final remarks.
  185. 20:02Uh, I think the agent payment funnel is really starting to take off now.
  186. 20:08And, uh, soon, uh, the real volumes will begin, and the real beginning of the agency economy .
  187. 20:14And I think now is the time to figure out how you can sell your services, uh, in this new world, because implementation can be a little more complicated than you think.
  188. 20:22And if you wait too long, if you wait too long, maybe, uh, you'll miss the train, and that's it.
  189. 20:29Thank you very much.