Machines of Buying and Selling Grace - Adam Behrens, New Generation
00:00:00.000 | -
00:00:02.000 | - So as a philosopher turned engineer,
00:00:18.180 | I have, for better or for worse,
00:00:19.920 | been obsessed with two questions my whole life.
00:00:22.760 | What is a thing and why does it exist?
00:00:26.260 | So we're here talking about AI in the Fortune 500,
00:00:29.680 | and the future of AI commerce.
00:00:31.660 | So let's ask a hopefully straightforward question.
00:00:35.960 | What is a store?
00:00:37.160 | A hundred years ago, a store looked like this,
00:00:42.000 | something we don't really recognize today.
00:00:43.860 | Inventory was in the back of a shop.
00:00:45.800 | You had to talk to a clerk, tell them what you wanted.
00:00:48.760 | They went and then fetched it and brought it to you.
00:00:51.200 | It wasn't until the 1950s and '60s with information systems
00:00:56.320 | that we were able to actually scale the concept of a store.
00:00:59.360 | And you saw big box retailers like Walmart and Costco emerge.
00:01:03.380 | The inventory moved from the back to the front,
00:01:06.260 | and you now had this concept of browsing.
00:01:08.820 | With the internet, we took that store and then we put it online.
00:01:13.760 | And so we took the scale of merchandise
00:01:16.260 | and matched it with the scale of distribution.
00:01:18.700 | You can now browse anywhere in the world 24 hours a day.
00:01:22.340 | A lot of people think that websites are dead.
00:01:26.360 | But a shocking number of people still shop online.
00:01:29.240 | Walmart had almost 500 million shoppers last month.
00:01:33.260 | Home Depot had 170 million shoppers in its online store last month.
00:01:37.900 | But the thing that we sacrificed was we got a sea of sameness.
00:01:43.300 | Can you really tell the difference here between Adidas, Reebok, Brooks?
00:01:49.360 | And so maybe we can start piecing this together to answer this question:
00:01:55.380 | What is a store?
00:01:56.260 | A store is a location and a protocol that facilitates transactions.
00:02:03.340 | You have merchants that want to sell something, buyers that want to buy something,
00:02:08.260 | and then a system to facilitate that interaction.
00:02:14.280 | So great, now we can start asking the question:
00:02:16.400 | What happens to the store with AI?
00:02:20.840 | If e-commerce digitized the merchandise and the distribution,
00:02:25.060 | AI digitizes the participants and their interactions.
00:02:29.560 | We go from static websites to merchant agents.
00:02:33.160 | We go from consumers browsing to consumer agents.
00:02:37.040 | And we go from low-level payment infrastructure to higher-level intent infrastructure.
00:02:44.060 | But we still have the same goal.
00:02:45.960 | That's not changing.
00:02:46.960 | We're still trying to get to a transaction.
00:02:50.920 | And so what is this actually going to look and feel like, you know, qualitatively?
00:02:55.820 | That's something a new generation that we spend a lot of time obsessing over.
00:02:59.800 | How will a new generation of consumer, both human and agentic, interact with a new generation
00:03:05.820 | of interface that is dynamic and real-time and generative, built on a new generation of infrastructure
00:03:12.840 | that needs to serve those new needs?
00:03:14.740 | So we see two possible futures.
00:03:23.760 | You'll have either AI agents that go to websites, or you'll have agents that have programmatic
00:03:24.840 | access.
00:03:27.560 | Option one, when they go to websites, will look something like this.
00:03:30.600 | Maybe you're in ChatGPT, you want a new TV.
00:03:33.760 | Maybe you're into gaming, so you say, "Hey, I want a new TV."
00:03:36.780 | Your agent goes to a new type of website optimized for that agent that can richly express your intent.
00:03:42.740 | It dynamically cuts the product catalog and the brand style guidelines and returns that content
00:03:49.360 | directly into the chat interface that you're operating in.
00:03:52.780 | When you want to buy, that agent then can go back to the website and go through the entire
00:03:58.540 | checkout process for you.
00:04:02.300 | It gets interesting when you think about programmatic access in this world.
00:04:06.020 | So maybe you have the same starting place, you're in Chat, but instead of actually going
00:04:08.980 | to a website, you have an MCP server or an API that's programmatically accessing every merchant
00:04:15.520 | on the internet.
00:04:17.540 | That API endpoint can then reason over the API call and return back a rich set of UI elements
00:04:24.140 | generated on the fly.
00:04:26.020 | Similarly, when you're looking to buy, instead of going to a website, it's just going to hit
00:04:30.720 | an API endpoint.
00:04:34.480 | And so if that's the future, how do we get from today, this sort of static world of consumers
00:04:40.640 | browsing and websites, to this future world of agentic interactions between buyers and users?
00:04:45.500 | So let's start with the goal.
00:04:51.360 | The goal is a transaction.
00:04:54.400 | We want high-quality conversion with users that hopefully are happy, hopefully they don't
00:04:59.060 | return the item that they get.
00:05:01.800 | And so in code, a payment is represented with a payment intent when a user clicks a buy button.
00:05:10.100 | That intent then goes through a set of transformations through the checkout process.
00:05:15.480 | results in success and money movement.
00:05:19.340 | Our first challenge in this agentic world is, what if software is the one that's clicking
00:05:23.440 | this button?
00:05:25.200 | So if you use operator today, it'll mostly error out on these e-commerce websites.
00:05:31.060 | And so there's two solutions that people have come up with.
00:05:35.100 | One is the solution that exists today.
00:05:38.100 | This is kind of the Stripe SDK solution.
00:05:40.060 | So instead of checking out with a merchant, you check out with ChatGPT or the software provider.
00:05:45.920 | That software provider then spins up a virtual card and buys the item on behalf of you using
00:05:51.800 | that new card.
00:05:53.720 | We think the more elegant solution is what one of our partners Visa is working on, which is delegated
00:05:58.780 | authentication.
00:06:00.720 | The agent is able to use your actual credit card and go through the checkout flow for you.
00:06:08.180 | So that solves transactions, but it doesn't get us very far.
00:06:15.500 | What if we can move up a level of abstraction?
00:06:18.300 | What if we can actually go to the buying and selling intent and the preferences?
00:06:23.960 | So a buyer intent today is expressed and it's inferred via keyword searches, click data,
00:06:32.440 | and site metrics.
00:06:36.360 | In the agentic future, we think that this is going to be explicitly captured rather than inferred.
00:06:42.880 | Conversation data is rich with user intent.
00:06:45.280 | And you can actually just ask a user agent what it's trying to do.
00:06:48.800 | You don't have to statistically infer it.
00:06:53.280 | This gives us our second challenge.
00:06:57.480 | If we have fuzzy intent, often people are just searching for, hey, I want a pair of running
00:07:02.400 | shoes.
00:07:03.800 | How do we actually get that to the skew level item, which is the inventory representation
00:07:08.580 | of that product?
00:07:11.940 | Today the main solution is to force it explicitly.
00:07:14.680 | So you actually have to provide a product detail page URL to an agent in order to buy.
00:07:21.280 | We think there's something very interesting that's emerging with the merchants that we're working
00:07:24.580 | with, which is users that come from an AI channel are much higher conversion, much higher dollar
00:07:31.080 | value, and much higher lifetime value.
00:07:34.320 | That opens up a whole new set of possibilities for how to rethink the cost structure of fulfillment.
00:07:39.480 | a product detail page.
00:07:40.480 | Maybe it doesn't actually matter if the user gets the wrong thing if it's easy to return
00:07:45.680 | and they're high value.
00:07:49.280 | On the seller side, that selling intent, as I mentioned, is really represented with a product
00:07:53.960 | detail page.
00:07:54.960 | We all know these.
00:07:55.960 | They're pretty static.
00:07:56.960 | There's just a buy button, a price.
00:07:57.680 | Maybe there's some discounts, maybe there's some bundles.
00:08:02.400 | In the future, this is going to be very dynamic.
00:08:05.080 | We think merchants will need to show real-time product availability, they'll have contextual
00:08:11.500 | pricing and discount that they can serve to a user in line, and the ability to infinitely
00:08:18.560 | bundle products across multiple merchants.
00:08:22.960 | This starts to get complicated, so then we encounter our third challenge.
00:08:26.840 | How do we know if a specific item is actually available across all the thousands of stores
00:08:32.180 | that exist?
00:08:33.180 | So we have a user that wants to buy something, and now we actually have to go find the store.
00:08:39.200 | There's two solutions today for this that we think are suboptimal.
00:08:42.620 | One is to use existing product feed infrastructure.
00:08:46.260 | So a lot of folks use this with Google.
00:08:49.440 | That requires chat products to individually work with every single merchant to get this
00:08:53.720 | data.
00:08:55.280 | The other alternative is you scrape product data from every website on the internet.
00:09:00.680 | We've done that.
00:09:02.080 | We think it's suboptimal.
00:09:04.340 | It's both repetitive.
00:09:05.920 | You have to do it a bunch.
00:09:07.480 | And then you end up clogging websites with bot traffic.
00:09:12.280 | The more elegant long-term solution that we're working on is to actually create a unified API
00:09:16.820 | to access product data across every merchant.
00:09:19.680 | You can think of this like a plaid, but for product data.
00:09:22.400 | So instead of aggregating over financial institutions, you can aggregate over merchants.
00:09:31.580 | And then the last piece at this layer is how do we represent buyer and seller preferences?
00:09:36.100 | Today, this is one-sided and very narrow.
00:09:39.340 | You have siloed user accounts, transaction data.
00:09:42.200 | You have limited LLM memory.
00:09:44.560 | And then businesses don't share really anything about their intent beyond quarterly and annual
00:09:50.040 | reporting.
00:09:53.200 | In the future, we think this will be two-sided and very expansive.
00:09:57.740 | They'll have rich context on users, not just across the purchase, but across every aspect
00:10:02.020 | of their life.
00:10:03.980 | And businesses will be able to express their real-time strategic goals.
00:10:09.620 | What's low on inventory?
00:10:11.420 | What users do they care about?
00:10:13.560 | What strategic change has happened with something like tariffs?
00:10:19.660 | The challenges here are very wide open.
00:10:22.240 | These are market design challenges.
00:10:24.480 | You have the challenges that preferences are very complex.
00:10:26.880 | They change over time.
00:10:28.420 | They often conflict between buyers and sellers.
00:10:31.380 | And there's a disincentive to honestly report your preferences.
00:10:35.460 | The current solution is you naively trust the information that you're given.
00:10:42.300 | Other folks have talked about prompt injection or manipulation of LLMs.
00:10:45.960 | And so that's kind of the world that we live in.
00:10:49.260 | When I worked at Bridgewater on our trading system and market systems, the world of finance
00:10:54.740 | saw this with third-party institutions and market makers that manage those differences between
00:11:01.420 | buyers and sellers.
00:11:03.340 | We think that's the world that this needs to move to.
00:11:07.240 | And so great.
00:11:08.240 | We've gotten a little bit further, but nothing is really agentic yet.
00:11:13.500 | You sort of need this last piece.
00:11:15.580 | This is the frontier, which is you need to add intelligence to the decision-making at each
00:11:20.380 | of those components.
00:11:22.100 | And so consumers and merchants need not just intents and preferences, but they need intelligence
00:11:28.900 | that can reason over them and negotiate them.
00:11:32.980 | And then on the infrastructure side, you need to move from just market-making to actual coordination
00:11:39.120 | and reasoning over these participants.
00:11:42.840 | The logic for generating in real-time the interfaces that each of these people need.
00:11:48.740 | Okay.
00:11:49.740 | So let's see how this is working in real life.
00:11:53.620 | I think a lot of people look at the Fortune 500 and they think, oh, these are big, old,
00:11:59.040 | slow-moving companies.
00:12:02.600 | But what people fail to realize is they had to survive the last 100 and 150 years of dramatic
00:12:08.800 | technological and societal shifts.
00:12:12.020 | And so we find that they're actually quite forward-thinking in terms of this challenge.
00:12:17.380 | And so if you take an example of Samsung, they started 100, 150 years ago as a fish merchant
00:12:23.840 | in Korea with 20 or 30 employees.
00:12:27.720 | It wasn't until the '70s that they started selling televisions, and it isn't really until
00:12:32.320 | the '90s that they became what we know them today, this technological behemoth.
00:12:36.760 | And they're at the forefront now of thinking, how does the brand of Samsung evolve in the world
00:12:41.620 | of AI?
00:12:43.940 | And how are they going to bridge e-commerce to this agentic future?
00:12:49.660 | So the first step is we create an API and MCP server for any chat client to use.
00:12:58.960 | A lot of Fortune 500 companies have complex product systems.
00:13:03.800 | Samsung, for example, has 10 different verticals, each with their own inventory representations
00:13:10.200 | of products, and so we do the work of attracting that into a consistent API with cohesive endpoints
00:13:17.360 | that will work across any merchant.
00:13:21.760 | The second step is then to connect that product data with other data sources at the company.
00:13:27.740 | This is the first step in constructing that seller intent.
00:13:32.100 | The natural starting place with what we're doing is let's just connect the brand and design
00:13:37.880 | system that the company has to actually wrap the products in how they want to be represented.
00:13:44.940 | So we're moving away from these carousel and static representations of products.
00:13:51.160 | That third step then feeds into a container for experimentation.
00:13:57.000 | And so we make an AI subdomain that allows for rapid experimentation of generative interfaces
00:14:05.240 | that can ingest both the product and brand data to serve customers.
00:14:11.860 | And so we're also experimenting with what does conversation look like when it's not just a bullet
00:14:15.820 | point and text list but actual images and products and content.
00:14:23.620 | And then the last piece is handling agentic transactions.
00:14:28.180 | This is actually enabling the payment flow to work on this new surface for bot traffic, which
00:14:34.620 | is a real inversion of the typical posture of a .com website.
00:14:39.880 | And the reason that brands are excited about this is again because users from AI chat, while
00:14:46.060 | they might be small, are much higher intent, they're deeper in the funnel, and they convert
00:14:51.760 | much better.
00:14:55.120 | And so we think every retail brand, every merchant needs to adopt this posture.
00:14:59.520 | And with any big change, we think the right approach is to start with the question of what and how.
00:15:08.140 | We don't think stores go away.
00:15:10.080 | We just think that they evolve.
00:15:11.340 | We think they return to the original form that they were in.
00:15:14.920 | And we think that form is actually a conversation.
00:15:20.200 | Thank you.
00:15:21.200 | All right, we do have time for questions, if anybody wants to chat for a bit.
00:15:33.520 | And if not, yeah, go ahead.
00:15:35.880 | Can you use the microphone over here, please?
00:15:38.240 | Thank you.
00:15:39.240 | I'm not that tall.
00:15:43.380 | I don't know if you noticed that, but we are talking about machine customers, maybe, just
00:15:48.920 | like there is a book called "Where Machine Becomes."
00:15:51.180 | From customers.
00:15:52.180 | Yeah.
00:15:53.180 | Yes, from the Gartner.
00:15:54.180 | So I work in a bank, and now it's very usable, what you said.
00:15:59.740 | But what is your projection in terms of how this is going to be, just like the break even,
00:16:06.180 | right?
00:16:07.180 | When it's going to start seeing this in our daily basis?
00:16:11.880 | Because today is only--
00:16:13.000 | Yeah, it's happening fast.
00:16:15.680 | I mean, even if you look at a product like ChatGPT, it is starting to bring shopping experiences
00:16:22.920 | into the product.
00:16:24.240 | The thing that's missing is you don't actually have the full journey within that application
00:16:29.680 | yet.
00:16:30.680 | And so you still link out to a website.
00:16:32.680 | And that's the thing that we're interested in exploring, which is, will you still link
00:16:38.180 | out to a website?
00:16:40.500 | I don't think that goes away immediately.
00:16:43.580 | And then every brand that we talk to, there's a strong desire to own a Surface in this new
00:16:48.680 | world.
00:16:49.680 | So we think you can start with a web-like Surface that is built in a way that's transportable
00:16:55.240 | so that when these chat products want to bring shopping into the application, that you can
00:17:00.240 | actually just bring that data and components directly in.
00:17:03.640 | You don't actually need to re-architect and rebuild.
00:17:07.480 | And so we're really excited about that.
00:17:09.020 | It's almost like this inversion where instead of going to the website, the website is going
00:17:11.720 | to go into a thousand different places.
00:17:17.960 | Yeah.
00:17:20.160 | Are credit cards the right payment mechanism for the agentic economy, or do we need something
00:17:23.580 | new?
00:17:24.580 | I didn't want to open that can of worms in the talk.
00:17:30.500 | I think conceptually, there's a strong argument for stable coins and crypto to be the native payment
00:17:39.560 | rail for AI, mostly because the agents can actually live within the wallet.
00:17:46.860 | I think practically, consumers use credit cards.
00:17:51.480 | And so it's the most likely bridge to get to that world.
00:17:55.780 | And then there's a third alternative, which is the agent itself just owns a perpetual credit
00:17:59.800 | card and you top it up, which is interesting.
00:18:03.200 | Thank you.
00:18:04.200 | There's an interesting parallel in places like China and Brazil where they have the form
00:18:10.500 | of super apps where everything happens in that one place.
00:18:13.500 | Yeah.
00:18:14.500 | Do you see Claude and ChatGPT trying to become the super app and shopping just takes place entirely
00:18:19.280 | over that platform?
00:18:20.420 | I think that's their goal.
00:18:23.500 | And then it becomes this question of how do the rest of us, how do merchants have control
00:18:33.240 | in that environment?
00:18:35.120 | And I think the one thing that model providers are very open to and what's different than the
00:18:40.840 | internet is, is the goal is to get the user to the right outcome.
00:18:45.800 | And so they have a, interestingly, they have a different incentive than tech companies in
00:18:52.180 | the past.
00:18:53.180 | And so we think there's actually kind of like a nice alignment between the goals of a merchant
00:18:58.060 | and the goals of these chat products.
00:18:59.980 | And do you foresee any form of revenue share happening there?
00:19:02.980 | Definitely.
00:19:03.980 | I think we don't think it's going to be advertising.
00:19:07.300 | It'll probably look a little bit more like either affiliate revenue or, you know, if you provide
00:19:16.800 | high quality data and you can be attributed to a good answer, I think there'll be some portion
00:19:22.900 | of that that these model providers will give back to merchants.
00:19:26.480 | Great.
00:19:27.480 | Adam, thank you so much.
00:19:28.480 | Yeah.
00:19:29.480 | That was a great talk.
00:19:30.480 | Thank you.
00:19:31.480 | Yeah.
00:19:32.480 | Thank you.
00:19:33.480 | Thank you.
00:19:34.480 | We'll see you next time.