Agentic Commerce Trends - Arun Kalyanaraman

This talk discusses the emerging concept of agentic commerce, which redefines how users interact with e-commerce platforms through conversational interfaces. The speaker highlights the shift from traditional user interfaces to agent-driven interactions, where an agent can autonomously manage tasks like trip planning or product purchasing without requiring user intervention for each step. They explain how foundational models and evolving protocols enable these agents to execute transactions seamlessly across multiple retailers. The talk emphasizes the changing roles of product owners and engineers as they adapt to a landscape where user intent and action can be captured within the same conversational context, leading to significant efficiency gains and the need for new engineering practices. The ultimate aim is the development of autonomous commerce at scale, raising important considerations regarding the management of agent interactions and economic factors such as token costs.

Okay, as I was walking in and Scott mentioned something about end of software engineering. So, somebody else also asked me the question saying so now that everybody is building agents, do we really have to do anything about software engineering? Are all the distributed systems we built so far not meaningful, etc., right? And I was reminded of um an analogy or a parable that my grandfather once shared

and I thought I'll start with that before I get into the topic itself. So, there was a man who was extremely hungry. And he had very little money, so he decided to use all the money to buy seven idlis, right? And then he ate it really fast because he was really really hungry. And once he finished the seventh idli, he started crying. And he was almost lamenting.

So, somebody went and asked him, "What's the problem? Why are you lamenting?" He said, "You know what? When I ate this seventh idli is when I felt satiated. I was happy. If I had actually eaten the seventh idli first, it's very probable that I would have saved money actually eating the first six idlis, right? So, for anybody who tells you that whatever you've done so far is

not very meaningful or it's not relevant, you need to know that you have to go through a journey before you can, in many ways, truly become agentic. So, with that context, I would go ahead and start my discussion on agentic commerce. It's something we all do, but we don't realize how much things are changing. It's just like how my mother uses WhatsApp today without realizing how much

she has changed in the way she is using a phone, right? And agentic is kind of creeping up on you that way without you truly realize how much is actually happening in the ecosystem. So, historically, whenever you thought about a website or a commerce site and somebody said, "Let's improve the experience. You went to your app developer or your web developer and they actually came up with

a new UI, lesser number of clicks. The entire UI was rewritten. You had a new version of the app, but imagine that is not relevant going forward. Imagine that you essentially can have your entire commerce conversation executed through a particular chat window. And that's that's really what agentic commerce is in many ways, right? So, what really is happening is you can going forward effectively from your app,

your website, or from an app even outside, any surface, even a social media site, have a conversation and have pretty much everything executed from that conversation without leaving that conversation itself, which means intent and action can be captured at the exact same point, which was never the case before, right? Before you had intent somewhere in a search bar, you clicked the link, the link navigated you to

a certain listing page or a display page, and then you took an action somewhere else. So, intent and action were never captured in the same place, but going forward, it's very possible you can do that. And the difference with agentic really is that the agent has agency. It can actually act on your behalf. And I I'll just take a minute to explain that to you, right? So,

let's say you decide that you want to go to Andaman and you want to go there on May May 5th. It's essentially say need to go to Andaman, plan me a trip traveling on May Right? It's what essentially is going to happen in a conversational flow is that it's not going to ask you any questions. It's going to understand that the weather in Andaman is hot. It's

going to find out the right kind of apparel you need to take. If it's even better, if you can imagine, it's going to say, "Okay, these are the tickets that are available for you." If you say my budget is within 500 or 50,000 rupees, it's going to go ahead plan your budget accordingly, make sure that your flight tickets are so much. Make sure that your apparel cost

so much. And if you have layered in preferences, right? It is going to add all of that together. So, normally, imagine you do that, right? You'd have gone to multiple sites. You would have managed your budget yourself. You've done pretty much everything. But, in this particular case, in the true promise of agentic commerce, you wouldn't have to do that. There would be a shopping agent which should

be able to communicate with multiple other agents across retailers to come up with one holistic experience where you can then go ahead and everything planned for you to make that decision. And when you said May 8th, it would identify if you're you know, placing an order for an item it arrives before May 8th. All of those decisions would be made. Imagine the number of clicks you're going

to save. Imagine the number of navigations you're going to save. And the kind of agency that provides you, too. So, what really is happening therefore is that in some ways UI is becoming irrelevant. A lot of our time and energy has been spent optimizing for UI experiences before. I mean, just imagine the number of stories you would have worked on to actually go ahead and do that.

That is going to change. And that changes a lot for a product owner. It changes a lot for the engineer. And it changes a lot for business. Because then business models also have to evolve. What happens to targeted ads, right? When the intent capture and the action happens in the same conversational window. It's a completely different paradigm for a targeted ad. And you need to consider all

of that when you're looking at how agentic commerce is going to evolve into the future. So, it's really shifting left if I were to go ahead and paraphrase all of this. From a UI or an app that you built or you won't which still could be relevant to another surface wherever. It doesn't matter. And you being able to interact with it. Your website or your app or

your capabilities. going on to the next slide. This is a video which I want to play in terms of what we have done so far with Hello. >> [music] >> Hello. So, this is on ChatGPT. >> So, you can see how much we have progressed over the last year in terms of how we have used conversations to actually drive guest behavior. So, the question now is, why

is this a big thing today? Now, if you really ask the product owner, if there are any product owners here, they would say that this is what we always wanted. I mean, we never wanted multiple interactions to happen. We didn't want multiple screen transitions to happen. My engineer never built it for me. And if you ask an engineer, they'll be cynical and say, "Yeah, my product owner

has always had this story without understanding why it's not possible." But, what's changed for us really in the last 8 to 12 months? Is that the foundational models have become really, really better. Almost every day there is something new which is coming up, and it's sometimes overwhelming just to keep pace. But, the fact is, they are very they're really, really good in terms of what they're able

to do. The entire protocol ecosystem has evolved, right? I mean, in terms of how AI can talk to your underlying systems, in terms of how agents can talk to agents, also in commerce where you have Google coming up with UCP where they're clearly defining how a commerce transaction can happen with multiple retailers so that you don't have to do point-to-point integrations or even with things like A2P

new AP2 protocols coming in where you're able to go ahead and establish trust. Now, commerce is one transaction where maybe it doesn't work. It has to be trust-based completely. So, you have protocols, for example, there where you're able to go ahead and capture mandates and effectively make sure that the mandate carries through from intent to checkout to placing an order. So, all these protocols that are evolving

they have helped us kind of accelerate this across in a very, very rapid way. And finally, the unprecedented scale. Now, the proof of all this lies in the pudding. I mean, you can build something great, but does it even get used? Now, if you really look at some of the conversion numbers that we see from versus clicks, the conversation conversion numbers are significantly higher than the click

conversion numbers, which means our customers or our guests, they want to have conversations. They don't like clicking on a facet, going to multiple links, and doing what they want to do. And you just look at the number of retail AI token processing that has happened through 1 year. This is across all of retail, but it has grown 11x, and the base number was 8.3 trillion tokens a

month. So, that's massive in terms of the number of guests or people who are actually using tokens and therefore conversations to communicate with a shopping journey. So, that's a very interesting point to note. So, you have the technology ecosystem evolving, you have protocol ecosystems evolving, and you have guest behavior or consumer behavior, which is really shaping this shift towards more and more agentic commerce, right? Now, just

to summarize everything that I said. Now, even when you look at our example, it's still conversational commerce in some sense, right? Because you have a AI which assisted the human. When you talk about something truly agentic, you will have to have it act on behalf of the human. Just to give you an example, let's say you say I'm I really have this particular blue shirt less than

5,000 rupees. Well, then it decides that yes, it's available, but it's not going to be delivered by 8th May. A truly agentic system will make a decision to go ahead and call a recommendation agent and say if it's not this blue shirt, maybe it's this blue shirt you should buy. And it makes the decision on behalf of the human. So, it's not that it's assisting a human,

it's moving away from to taking actions. Similarly, before humans decide even in that checkout flow, you had the add to cart button, which was still something which the human did. Going forward, it's very likely that you can have and delegate that to an agent for an agent to do it. The scope is broadly single-track so far, right? Search for an item, discover an item, product Q&A. But

now you're talking about end-to-end journeys that can be orchestrated, and you can go ahead and have that coordinated not only across one retailer, but potentially across multi-retailers, right? And personalization is a big thing because so far we've always dealt with session context. In this session, what did the guest do? But imagine what is the currency going forward? It's your entire chat history. You you will have long-term

memory and short-term memory, which are going to go ahead and be available for the agent to have that much more of a personalized experience going forward. And the autonomy level obviously increases from it being low autonomy to high autonomy. And so, that's broadly how it's better than a chatbot. You know, when you really look at agentic commerce, it's pushing the envelope in terms of where it's going

to go maybe in the next couple of years, So, now if you just to take you through a flow of how it happens, right? There is a guest customer layer, which could be any layer, which could be a chat interface. There is a Gemini shopping agent layer, which will go ahead and reason, plan, and orchestrate like, "Okay, this is about maybe planning a travel, this maybe about

accessories, this maybe about all of these other things. And then you have a protocol layer through which it's able to interact with the retailer's ecosystem. The retailer would also have agents. We would have let's say a super agent, and then you would have a domain agent and sub-agents under it. In our case, we could have a search agent, we could have a checkout agent, you could have

uh inventory agent, all of that available. And there are multiple levels at which this orchestration happen, and the agents can talk to each other, and then this entire agent-to-agent communication can happen, which then places an order on your behalf. But the important thing I would want to highlight is the need for an agentic platform, even for a company like like Target, right? Because you're going to be

in this place where you're going to build more and more agents going forward. And all of us have been there when we started building, moving to microservices, it's easy it's easy to build too many. And it's easy to build an agent. So, let's be honest, it's not terribly hard today. So, you can have pretty much every team build an agent, right? And at that particular point in

time, you're going to have so many agents that it's going to be really really difficult to manage. So, you need to have an agent platform somewhere where you have an agent registry, where discovering agents become easy. And then you need to have everything else that comes with it. How are you going to have agent-to-agent communication happen? How are you going to have memory management done? How are

you going to make sure that you have the right input-output validations? How are you going to incorporate the responsibility AI? How are you going to give the agent access to the right tools? Everything that needs to happen there, and plus what model are you going to use for what purposes? Because tokens are not cheap, they are expensive, and at some point in time the per-unit economics will

have to play out in all of these initiatives, right? So, an agentic platform becomes therefore foundational, and either you build it or you buy it in some shape or form, it's going to be required for you to go ahead and power of these things at this is not so much to highlight how Target has evolved, but the but to highlight how it is important that you consume

the six idlis before you get to the seventh idli, right? I mean, I did speak about how capabilities are going to be super important going forward for a retailer. In the absence of a UI, the kind of item data you have, the kind of catalog information you have, the kind of price and promotion information you're able to go ahead and publish, and how frequently you're able to

publish That is going to be super important. So, if you're going to have data in isolated systems, if you don't have a semantic layer on top of it, it's very likely that your agent is going to return garbage because it doesn't know what to make of it. If you don't have composable services which they can interact with, it's very likely that you're not going to go ahead

and scale this at all. So, at some point in time, investing in foundational capabilities, and wherever you are in that journey in your companies, is going to be is continuing to be critical if you're going to go ahead and not only for the agent e-commerce, but scale agents for any use case going forward. So, it's to highlight the point that you will have to evolve to this

um through a particular journey. So, where are we now? I mean, as I said, a lot of our teams are building single agent transactions. It's quite simple to do it. We're also leading into multi-agent commerce right now. We're working actively to see how we can bring some of these use cases to life. We hope that somewhere towards end of 2026, early 2027, we will be able to

go ahead and power autonomous commerce at scale. And there are questions that we will have to answer, right? I mean, about customer relationships, about how are you going to go ahead and manage agent-to-agent communication to make sure that nothing goes out of wire, right? Also, about token costs. I mean, that's going to be a very, very important factor as you think about how this is going to

scale. I mean, conversion could be great, but the number of tokens that are going to be required to power some of these are going to be prohibitively expensive. How does the unit economics add up? So, some of these things are going to be ongoing discussions that we'll have to optimize for. Next slide. So, finally to end with, if there are three things that I would want you

to take away from this, this is not something that you can just jump into. It's something that we will all evolve to. So, all you engineers there will continue to have a lot of work to do through many years till we get that. Like Scott mentioned, the engineering practices are going to continue to be important even as you do this, right? I how you do security in

an agentic world will look different. How you will do performance testing will look different. I mean, we always had performance testing model based on clicks. Tomorrow, it's going to be There are going to be no clicks. So, how are you going to go ahead and scale? How are you going to performance test? So, all of those things are going to happen, and you will have to mature

your engineering practices and your underlining tech ecosystem to actually take full advantage of what is going to happen in the future. Open protocols are continuing to evolve and they'll be foundational to accelerate. And finally, anybody who's skeptical about it, this is actually happening now. I mean, there are retailers talking actively to some of these large companies to see how we can bring that to life. There are

internal investments of fairly significant scale happening to see how we can actually power this. So, this is something which will definitely happen. And if I were to finally conclude, we've spent many years of our engineering where we've built for engineers for human for humans for humans to use, right? I mean, that's been the paradigm. And going forward in the next two or three years, we're possibly going

to build for agents and for agents to use. And I think that's the mindset which will be different as we actually go through this journey. So, I'll be still around for a couple of hours. I'm at time, so if there are any questions, feel free to reach out and ask. But, thank Thank very much for your time.