Latest / Agentic Conversations (formally mlops.community) / Inside Just Eat's AI Lab: Voice Agents & Agentic Commerce

Transcript

Inside Just Eat's AI Lab: Voice Agents & Agentic Commerce · Agentic Conversations (formally mlops.community)

• 0:00
  I'm Guthrie Cooper. I had a product for the
• 0:03
  incubator domain and that is part of Just Eat Take Away,
• 0:07
  which is a company that we work for and we're obviously within
• 0:09
  the Prosus group. I'm Niti Sharma, majority of my
• 0:12
  career had been in the grocery and retail.
• 0:15
  So I worked for multiple retailers to build their
• 0:18
  e-commerce website as an engineer, then moved into
• 0:21
  architecture role and then into engineering leadership.
• 0:28
  You know, I could just frame the AI innovation and incubation a
• 0:32
  little bit just to give a little bit oversight of you know what
• 0:35
  this team is all about and how we are structured.
• 0:39
  So currently from the team structure wise we have we begin
• 0:43
  with an incubation within incubation.
• 0:46
  So we have a team called Pre Seed.
• 0:48
  Like Inception. Yes, it's like inception.
• 0:50
  So we we this was one of our learning as well.
• 0:53
  So what we have is a pre seed team and they are all about
• 0:57
  just, you know, bringing a prototype to life in two to four
• 1:01
  weeks. In incubator, we operate like a
• 1:04
  startup. We're really we are inventing
• 1:08
  the next wave or waves of this industry.
• 1:11
  So like right now in the world of AI, we see that, you know,
• 1:14
  every other day there is a new protocol and there is a new
• 1:17
  technology. So we need a bunch of people to
• 1:20
  just, you know, jump on it to try it out, see if it has
• 1:23
  potential, and then we add it to our pipeline.
• 1:26
  And that's why we call it Incubator with an incubator and
• 1:30
  then rest of the teams are basically working on different
• 1:33
  teams. As you know, that world is
• 1:36
  evolving faster than we think. So we just have to come up with
• 1:41
  multiple ideas working in parallel.
• 1:44
  So some of our teams are working on autonomous delivery.
• 1:47
  Some of our teams are experimenting with all the
• 1:50
  innovation things like evolving our existing AI assistant to a
• 1:56
  fully agentic commerce person shopper agent and Guthrie will
• 2:02
  explain more about, you know the vision for the personal shopper.
• 2:06
  And then some of our team are also working on our off the
• 2:10
  platform strategy because we work with bold vision that there
• 2:14
  is a possibility that we'll not have any apps in the market or
• 2:19
  we may have just just super apps.
• 2:21
  We don't know how the you know the consumer will interact in
• 2:25
  future. We launched back in the year
• 2:28
  2000, so we were the first food ordering online platform in
• 2:33
  Europe. It's always been about
• 2:35
  innovation and it continues to be about innovation with the
• 2:39
  process acquisition that happened last year.
• 2:42
  We have jet fuel plugged into, you know that that culture of
• 2:48
  innovation. So 2000 on the web, around 2009,
• 2:52
  we were one of the first apps. So we moved obviously to the app
• 2:56
  ecosystem. That was the big change that
• 2:57
  happened then between 2009 and where we are now.
• 3:04
  You know, technology innovation was a big part of what we were
• 3:07
  doing, but it was also about adding more merchandise for our
• 3:12
  consumers to purchase, right. So the consumer expectation
• 3:15
  continues to evolve over time. And in in 2000, it started as a
• 3:20
  food delivery service because Jet now, I guess for the
• 3:23
  Americans that are listening who may not be so familiar with it,
• 3:28
  it is a food delivery service. You can order from restaurants,
• 3:32
  you can order from supermarkets, you can also order from other
• 3:35
  places too, right? Exactly.
• 3:37
  Yeah. So in the last couple of years,
• 3:40
  we've added grocery. Grocery's been a big growth
• 3:42
  lever for us. We've added electronics.
• 3:45
  You know, our vision is to empower everyday convenience.
• 3:48
  And if you really think about what that means, everyday
• 3:51
  convenience, consumers have many needs across over the course of
• 3:55
  their day, of the course of the week, of the course of the
• 3:56
  month. And we are increasingly there
• 3:59
  for every single moment that they need to purchase something,
• 4:02
  they need something entered lead.
• 4:03
  Give an example. If you're playing PlayStation,
• 4:07
  you know, with your with your family or your friends and one
• 4:10
  of the consoles breaks, you can order a replacement console and
• 4:13
  you can get it in 20 minutes from Justeed, right?
• 4:16
  So innovation is on the tech side, but it's also very
• 4:20
  important to make sure you have the right merchandise for the
• 4:22
  customer at the right time. Looking at the future.
• 4:27
  So here we are, you know, early 2026 consumer expectations and
• 4:31
  the tech landscape is evolving faster than ever.
• 4:34
  It's one of the most exciting times that I've had in my
• 4:36
  career, right? So, and to Neddy's point, we are
• 4:41
  very interested in ensuring that we are exactly where customers
• 4:46
  are, right? So we want to meet them where
• 4:48
  they are. And that means on our app and
• 4:52
  off our app. So making sure that we're
• 4:55
  present, we call it beyond the app.
• 4:57
  That's the name for our strategic theme, right?
• 5:00
  So whether they're using chat programs, whether they're using
• 5:04
  search. And when I say search, I mean
• 5:06
  people will think Google, but of course, OpenAI is, is battling
• 5:11
  Google for that top of the funnel, right?
• 5:13
  And I think the metrics at this point are 2% of global search is
• 5:17
  done in ChatGPT. The growth rate of that search,
• 5:22
  of that search or that discovery interface is 60% year over year.
• 5:26
  Google is still the majority, but they're growing at a slower
• 5:28
  rate. So the consumer landscape is
• 5:31
  changing. And in Incubator, we exist to
• 5:33
  make sure that we're one step ahead of consumers and we're
• 5:36
  there for them. Voice is not a new platform,
• 5:41
  it's a new mode, right? And the voice feature that we've
• 5:45
  released is built on top of the most innovative tech platforms
• 5:51
  available right now. It is fully conversational.
• 5:54
  We can maybe do a small demo later so I can show you what I
• 5:56
  mean. But it is, it's contextual, it
• 5:59
  has personality, you can interrupt it, you can have a
• 6:03
  conversation with it, and it uses all of that context within
• 6:05
  the conversation and it makes recommendations for you.
• 6:08
  Some of the benefits of that are, you know, so if you think
• 6:12
  about accessibility, right? So people who would prefer,
• 6:17
  there's many customers who prefer to use voice versus use
• 6:20
  the fingers or some other paradigm.
• 6:23
  That's why you see companies like Chachi, BT, Gemini, etc,
• 6:27
  integrating voice. A lot of people are using that.
• 6:30
  During the course of your day, there are moments when it makes
• 6:33
  much more sense to use voice. For example, you're driving your
• 6:37
  car, you're on a bike. And the vision there, if we
• 6:40
  think beyond the app and we pair that with another strategic
• 6:43
  theme that we have which is conversational continuity,
• 6:47
  maintaining that context, maintaining the context of the
• 6:50
  conversation across modes and touchpoints throughout your day.
• 6:55
  That's what we believe is going to be, you know a really
• 6:58
  powerful lever for our growth as a company and it will meet
• 7:02
  customers expectations. Let's talk for a minute about
• 7:04
  these different touchpoints, and basically what I'm hearing you
• 7:07
  say is memory. You just want to be.
• 7:11
  I'm remembering everything in every mode that folks are
• 7:15
  engaging with the app. Yeah.
• 7:16
  So when we say beyond the platform, we talk about our
• 7:20
  platform technically, but the context and the intent from the
• 7:25
  external apps or it could be from our app as well.
• 7:29
  So for example, if you are using ChatGPT, you are basically Chat
• 7:35
  GPT's user. So your intent and conversation
• 7:39
  will remain with ChatGPT. As for the privacy consents and
• 7:43
  everything. But the moment you switch to Jet
• 7:47
  to pay to buy something, that's when Jet app will start your
• 7:54
  context. And with the consent we will
• 7:57
  maintain that context. But when we say conversational
• 8:02
  continuity, what we mean by that is that you are in your car and
• 8:07
  you're hungry and you start a conversation to get something,
• 8:12
  but you have to switch to your mobile later or you might have
• 8:16
  to use your laptop. The same conversation should be
• 8:19
  continued, so the context should be there, the information should
• 8:24
  be there, and the flow should not break.
• 8:26
  That's what we mean by conversation continuity.
• 8:29
  But when it comes to voice agent specifically when you start a
• 8:34
  session, that session context is maintained so that conversation
• 8:39
  is being maintained in the memory and based on tools like
• 8:45
  you know your order history or dietary preferences, you will
• 8:48
  get the right recommendations. Yeah, OK.
• 8:51
  If you're populating things on Chat GPT's apps or Claude apps,
• 8:57
  I guess MCP apps is what they call it now.
• 9:01
  How much of that information are you getting back so that you can
• 9:07
  help the MCP apps give the right things and make sure that you
• 9:14
  are recommending the proper things to the user?
• 9:17
  If you don't have any of that information now, all of a sudden
• 9:19
  it's like you're flying blind or you're just getting the intent
• 9:23
  from ChatGPT or Claude, right? You don't get to decide what the
• 9:27
  intent is, or try and puzzle the intent and put new things in
• 9:33
  front of the user that you think they might want, or upsell them
• 9:35
  on things. Yeah, it, it depends on the
• 9:38
  platform that we're working with, right?
• 9:40
  So for example, Google might have different policies in
• 9:45
  place, different tooling in place versus an open AI, right?
• 9:50
  We saw a talk from Google last night where they're really, you
• 9:53
  know, Google is famously an open platform, right?
• 9:57
  So that memory, that context might be available to us in some
• 10:03
  partnerships more than others. The dream is to have the right
• 10:07
  kind of context, the right amount of memory, the most
• 10:09
  relevant data available across platforms at all times, across
• 10:13
  modes and across platforms. And it even goes beyond, you
• 10:17
  know, ordering agents, ChatGPT, Gemini, etcetera.
• 10:22
  You can think about wearable devices, so for example your
• 10:25
  smartwatch or think about health data, right?
• 10:30
  So food and health are. Obscene.
• 10:32
  They're intertwined. Yeah, exactly right.
• 10:34
  So obviously this, you know, I need to caveat this by saying we
• 10:38
  take user privacy, user data extremely seriously.
• 10:43
  So we never want to do anything. That's obviously and legislated
• 10:48
  for the regions that we operate in and of course user consensus
• 10:50
  is super key. Disclaimer Put that in there
• 10:55
  just so everyone knows that is good.
• 10:57
  Indeed. I mean, it goes without saying,
• 10:59
  but it is super, super important to us, right?
• 11:01
  But it does. It opens you up to a new world
• 11:03
  if you're using health data. Yeah, I just had a workout this
• 11:07
  morning. And now you probably know if
• 11:10
  you're reading my whoop. Yeah, that I did a workout.
• 11:14
  Oh, I should order you some. Yeah, whatever protein rich
• 11:17
  foods. Exactly.
• 11:18
  That actually goes one step further.
• 11:19
  So I'm going to give you a story that combines that health data
• 11:23
  with some magentic proactive use cases as well, right?
• 11:27
  So we know that you, Demetrius, order the same coffee from the
• 11:31
  same partner regularly, right? So if you've had your workouts,
• 11:36
  we know that usually after the workout, you order that coffee,
• 11:39
  right? So we proactively prompt you to
• 11:41
  order that coffee. We might even place the order
• 11:43
  for you or there might be some a healthy friction in there where
• 11:46
  you just get a notification on your on your watch or in the
• 11:49
  app. And you said, yeah, go ahead and
• 11:50
  buy it. On the way to your coffee.
• 11:53
  You are on your bike and you realize I'm going to invite a
• 11:56
  buddy to join me, right? So now you need 22 drinks in
• 11:59
  that order. You can pick up where you left
• 12:02
  off on voice because you're on the bike and you say, hey, the
• 12:04
  order that's being placed, I need an additional cappuccino.
• 12:08
  Can you add that to to the order?
• 12:10
  That can be done through voice mode, right?
• 12:12
  And then wait. Can I stop you there because
• 12:14
  please, I think that is a beautiful vision.
• 12:16
  But the friction or the problems that I generally have when I'm
• 12:20
  in that experience is that I'm on my bike.
• 12:24
  I can't open my phone and open the app or I do, but it's a pain
• 12:28
  in the ass like to be riding, trying to focus on the phone and
• 12:32
  then find the app. And then So what I'll try to do
• 12:36
  normally, and this is where it fails is I'll say Siri, open the
• 12:41
  app, open, just eat. And Siri sucks.
• 12:45
  So it gives me some random app where it says, oh, I don't have
• 12:48
  access to just eat or you need to scan your face ID for that.
• 12:52
  And I'm like, I'm on the bike. I can't scan it right now or I
• 12:56
  will have to whip it out. And it's kind of that same
• 12:58
  thing. So I've see the vision.
• 13:00
  I think it's amazing, but you're almost beholden to how bad the
• 13:06
  other voice agents are out there.
• 13:08
  The ecosystem and the way the ecosystem interacts is super
• 13:12
  important. And that, that friction point
• 13:13
  you just described now is exactly why our domain incubate
• 13:17
  exists. We see those friction points as
• 13:19
  absolute gold. So that's an opportunity, right?
• 13:21
  So we would take that use case, let's assume and we could
• 13:25
  probably quite safely assume that we do a bunch of research,
• 13:27
  which is a really important part of our process.
• 13:30
  And the friction point you've raised comes up as as a common
• 13:33
  thing, right? We have really tight
• 13:36
  relationships with many of the companies you mentioned and
• 13:38
  others, right? And they're looking to solve
• 13:40
  those problems as well. So we would team up with them to
• 13:43
  make sure that we can find a way to get that right.
• 13:46
  And it could be through integration with your Apple
• 13:47
  Watch. That's the level that we're
• 13:48
  operating in. Those those connection points
• 13:51
  are the key because it's very easy for me to sit here and map
• 13:54
  out the happy path. Yeah.
• 13:55
  But the work is, you know, understanding those friction
• 13:58
  points and. Exactly where it fails.
• 14:00
  And if it fails before it even gets to you.
• 14:03
  Yeah. Then you don't know why.
• 14:04
  It's like a silent failure. Yeah, yeah.
• 14:07
  Yes. And that's where I think identic
• 14:11
  AI commerce comes in play because what we one of the thing
• 14:15
  with identic a commerce is that we should be, we should
• 14:18
  integrate all the possible surfaces for the frictionless
• 14:22
  commerce. And one of the surface is this,
• 14:26
  you know, scenario where the problem begins with your inbuilt
• 14:31
  app in the iPhone. How do you actually improve that
• 14:36
  site? First, because you need that as
• 14:39
  an entry point. Even if you solve the problem
• 14:42
  with building multiple agents for ordering, decision making,
• 14:47
  recommending, but at the same time the basic fundamental
• 14:54
  blockers are still there, as a customer you won't be satisfied.
• 14:59
  So this is where I think these new protocols where the where
• 15:03
  you have system to system integrations are going to add
• 15:09
  some values and will provide the frictionless commerce.
• 15:13
  And system to system? You mean like MCP?
• 15:15
  Yes, MCP agent to agent basically.
• 15:19
  And then there is there, there is also, you know, potential and
• 15:24
  anticipation that instead of having multiple apps, you will
• 15:28
  only have one lifestyle app, for example.
• 15:31
  And that lifestyle app can actually read when you're bike,
• 15:35
  when you're riding, when you are free.
• 15:38
  And then based on that, you know, make sure that your phone
• 15:41
  is unlocked, for example. But that is like all these
• 15:44
  things needs to be think about. And then based on the research,
• 15:49
  there should be some solutions in place.
• 15:51
  What we know as of now is that there are multiple surfaces,
• 15:55
  there are variables like, you know, with your meta glasses, we
• 16:00
  can integrate our app or if you are just using Gemini all the
• 16:05
  time or if you're just using ChatGPT.
• 16:08
  So you did mention about, you know, how this user interaction
• 16:11
  will work and how would Jet be benefit with this kind of
• 16:15
  integration where you're actually using Jet GPT and your
• 16:19
  intent and your context and your behaviour and everything is with
• 16:23
  ChatGPT. We Jet is there just to, you
• 16:26
  know, provide you the the best recommendations.
• 16:29
  So I think it depends what kind of user base we are talking
• 16:33
  about. So the the Jet already has the
• 16:37
  user base. We already have, you know, large
• 16:40
  amount of customer data and we do have their order history
• 16:44
  based on the consent and everything.
• 16:46
  Similarly, ChatGPT has their own or OpenAI there has their own
• 16:51
  user base. Some of the users are already
• 16:52
  Jet customers, some not. So we need to think about, you
• 16:56
  know, these things individually and build the use cases and
• 17:00
  solution around those use cases. But this, this is still new.
• 17:05
  The kind of contract, the kind of standards and the kind of
• 17:08
  collaboration we'll have in this B2B scenario will define how
• 17:13
  this, you know, relationship will work.
• 17:16
  But we need to keep customers privacy at top.
• 17:19
  We need to take the right consent with them in this
• 17:23
  identity commerce. We shouldn't just allow, you
• 17:26
  know, sharing data between two vendors or two companies without
• 17:29
  even letting our customer know that this is what we are going
• 17:32
  to do with their data, even though it's for customer
• 17:34
  convenience. Yeah, I do see a world where
• 17:37
  'cause I think about how I use three or four different LLMS
• 17:42
  daily and sometimes it's just out of convenience that I'll use
• 17:48
  one or the other because it's there.
• 17:50
  It's like, all right, I'm using anti gravity with some with
• 17:54
  Gemini, and then I've got Claude and I've got a hot key where I
• 17:57
  just press Caps Lock and I instantly start talking to
• 18:00
  clawed. And there's probably a subset of
• 18:05
  users that you said like they're Jet customers and they to
• 18:09
  appreciate the tight integration.
• 18:12
  And so they're going to be able to configure into their
• 18:15
  different LLMS like, hey, I'm a Jet customer, here's my whatever
• 18:19
  Jet customer ID and you want to be able to have a very tight
• 18:23
  coupling so that anytime I ask for anything, it's like a
• 18:28
  default that it will just go and use Jet.
• 18:31
  But then there's another subset of customers where it is like
• 18:36
  searching for a restaurant on Google and you get the blue
• 18:38
  links and some of those are sponsored and some of them are
• 18:41
  not. Yeah, there's a spectrum and I
• 18:44
  think we need to be prepared for a world where there'll be many,
• 18:48
  many different types of user appetites, right from your early
• 18:51
  adopters, your Dimitra style users all the way to, you know,
• 18:55
  we serve all demographics and we're in 16 countries, right.
• 18:59
  So like Nelly said, there's on a long enough time scale, perhaps
• 19:02
  the apps, the apps, the front end of the apps disappear and
• 19:06
  it's just like a fragmented, you know, ecosystem of consumer
• 19:11
  tools. Probably the reality will be
• 19:14
  somewhere in between those two. So we need to make sure that
• 19:17
  we're able to serve consumers no matter how advanced in
• 19:20
  technology usage they are or what they want, right?
• 19:24
  And you should be able to change that over time as well.
• 19:27
  So if you're a ChatGPT user and then you decide actually I'm not
• 19:31
  into talking to Allah limbs anymore, I want to go back to
• 19:33
  the app, you should be able to do that, right?
• 19:35
  For sure, or you switch from ChatGPT to Claude or to Gemini.
• 19:39
  Absolutely, Absolutely. Yeah, I think we all have our
• 19:43
  favourites. We still, even though there are
• 19:46
  so many LLMS, we still think like, you know, some of those
• 19:49
  really understand us. And I do have my favorite
• 19:52
  because I feel like I, I get the right kind of responses and I've
• 19:57
  been heard and understood properly.
• 19:59
  And that's, that's the vision actually.
• 20:01
  The more you know about your customers, the more you know
• 20:04
  about their intentions, their feelings, the better you can
• 20:07
  serve them. And that is why one solution
• 20:09
  cannot be. For every customer, we need to
• 20:13
  just look at, you know, the kind of behaviours they have and then
• 20:17
  solve them accordingly. And that's what I think in, in
• 20:20
  Jed, that's what we live by that customer convenience and meeting
• 20:23
  customer where they are not just physically, but also
• 20:27
  emotionally. And you know what, what kind of
• 20:31
  what kind of things they're looking for, be it food, be it
• 20:35
  any convenience, any other convenience from general
• 20:39
  merchandising And the example, which is that we just shared.
• 20:42
  So this is the kind of convenience we want to provide,
• 20:45
  but there should be that alignment with you as a customer
• 20:49
  really needed because we don't want to over serve at the same
• 20:52
  time. And I'll share some examples
• 20:55
  where we feel like, you know, this will be instead of customer
• 20:58
  convenience, it will become customer grievance.
• 21:01
  And push notification is one of the example where you keep
• 21:04
  sending notifications for different vouchers and offers.
• 21:07
  And you know, you buy this from here and you get this sometime
• 21:12
  that is like too much overhead for a customer to keep track of
• 21:15
  these things, even though you are providing the convenience.
• 21:19
  So we have to maintain that balance when we think about, you
• 21:23
  know, what exactly what type of problems we want to solve and
• 21:27
  how we want to come across as a brand as well as how we want to
• 21:31
  come across as as an organization who is actually
• 21:34
  solving and fixing the real challenges people are going to
• 21:39
  have in AI world and even exist today.
• 21:43
  That is exactly what I thought about when you were saying
• 21:45
  proactive. And I was like, oh, it's so
• 21:49
  dangerous because as soon as you send two notifications that
• 21:54
  aren't relevant, I turn off notifications.
• 21:56
  Yeah, yeah. But if it's relevant, it's like,
• 21:59
  oh, damn, that's actually kind of nice.
• 22:01
  I just see it with all of the apps that I have in general, I
• 22:06
  will turn off notifications because it's usually just like,
• 22:10
  oh, save 20% if you do this. And it's trying to get me to use
• 22:14
  the app, which isn't necessarily that I don't want to save 20%.
• 22:18
  It's just that in this moment in time, I don't have a use case
• 22:22
  for what you're trying to push me like.
• 22:25
  So it doesn't make sense that I would use your app.
• 22:28
  It's not that I'm not using it because I don't want to.
• 22:32
  It's just that I don't have a use case right now.
• 22:35
  And so you have to be really smart about when to do that.
• 22:39
  Yeah, very few companies have have cracked this right.
• 22:43
  And I really think it is such a huge opportunity.
• 22:46
  The last thing you want to do is get people to turn off
• 22:48
  notifications because then you've lost a child.
• 22:52
  But if you get, if you get context right and you get the
• 22:56
  moment right, and you understand your customer base and their
• 23:00
  history to the right level, it can be done right if I can pick.
• 23:04
  Up this the story where I left off, so yeah, I totally cut you
• 23:07
  off, sorry. No worries.
• 23:08
  It was a good time to jump back into it.
• 23:10
  So, so, so coffee story is done. You've had your workout on the
• 23:15
  way home, you're on the app and you want to place a grocery
• 23:18
  order, right? So you can use the traditional
• 23:20
  app to place that grocery order or you can go to our chat
• 23:22
  assistance, right, which is text based assistance, much like
• 23:25
  ChatGPT BT it has the voice mode.
• 23:27
  So the voice mode is really just voice, the voice version of that
• 23:31
  chat app. So we have both of those modes.
• 23:33
  So you can use that app to build your grocery list for your
• 23:36
  evenings dinner with your family.
• 23:37
  So you could say, hey, I want to make a carbonara for the family.
• 23:43
  That's how many people they are, any dietary restrictions,
• 23:45
  etcetera. And it would pull that list of
• 23:47
  ingredients together. It might even surprise you with
• 23:50
  some extra ingredients that you didn't know were in the recipe,
• 23:52
  for example, beyond that, so let's go back to you.
• 23:57
  You've got a Whoop on, right? So let's assume we have an
• 23:59
  integration with Whoop or or Apple.
• 24:02
  Watch. Speaking of shitty
• 24:03
  notifications, yeah, whoop. Oh my God, dude, I am.
• 24:07
  I had complained about this like every other podcast, but they're
• 24:10
  so passive aggressive on me. They're like, your sleep score
• 24:15
  could be better. Yeah.
• 24:16
  I'm like, I've already feel like shit, whoop.
• 24:20
  You don't have to remind me. And you've slept already, so you
• 24:23
  can't fix it. Yeah, it's not like I can
• 24:25
  retroactively go back in time. Yeah.
• 24:27
  So anyway, sorry, that's just the grievance I have with.
• 24:30
  I hear you. I hear you.
• 24:31
  That's that's actually that's. What I said, it can easily
• 24:34
  become a grievance. The convenience can be a
• 24:37
  grievance as well. Yeah.
• 24:39
  So agree convenience can also become delight.
• 24:43
  Yeah, right. So, So what I'm getting to is
• 24:45
  like you order that carbonara, those ingredients and you get
• 24:50
  imagine the assistant nudges you in the right way with language
• 24:53
  that is tailored to you so that it won't become a grievance to
• 24:57
  say, hey, you know, we noticed your heart rate variability was
• 25:02
  a little different this morning versus your baseline.
• 25:05
  You might want to consider skim milk in that pasta, right?
• 25:10
  Does that actually work or? Did you just make that up?
• 25:14
  That doesn't work yet, but well, let's let's no, I mean.
• 25:17
  The Skim. Milk in pasta for my.
• 25:19
  Heart rate variability because. I'm not a doctor.
• 25:23
  Don't play one on the Internet. Too quick.
• 25:24
  Too quick. But you can put milk in pasta
• 25:27
  sauce. Yeah, yeah, yeah.
• 25:29
  But anyway, yeah, that integration like understanding a
• 25:32
  360 view of you and saying I see this happening in your all of
• 25:39
  these data points that I have for you and here's my
• 25:42
  recommendation. Exactly.
• 25:44
  And we can play, we can test that out in the voice app today.
• 25:50
  We don't have the health data plugged in yet, but you can give
• 25:53
  it that data in the conversation.
• 25:55
  So you can say, hey, I'm trying to lose weight.
• 25:58
  Can you just tailor this recipe a little bit for me?
• 26:02
  Or you know, yes, I was ordering a burger and I said yeah, I want
• 26:07
  the healthiest option possible and it made some recommendations
• 26:10
  around button. Choices.
• 26:12
  Did it say just no burger? That's not you want a healthy.
• 26:15
  Burger after having a dinner. It offered me a vegetarian
• 26:18
  option. Oh, there you go.
• 26:21
  Oh yeah, all vegetables. Indeed, indeed.
• 26:24
  Just. Vegetables, yeah.
• 26:25
  I think food is the integral part of our being.
• 26:29
  Yeah. It touches upon the health, it
• 26:33
  touches upon the mood, social and the lifestyle and the social
• 26:37
  interaction. And in some culture, food is the
• 26:39
  language of love and connection. So making it valuable and
• 26:45
  ensuring that you know, there is a technology as well as the
• 26:49
  intent clearly defined and balancing out for the for, for
• 26:53
  an individual is tricky. But at the same time, it's kind
• 26:57
  of, you know, a social cause where you feel satisfied and
• 27:01
  complete that you are actually adding value to somebody's life
• 27:05
  by helping them to choose the right options when it comes to
• 27:09
  eat. And that's what I think the
• 27:12
  vision for our personal shopper is to ensure that we understand
• 27:17
  or this agent understand you fully as an individual and self
• 27:22
  is the recommendations which are very much tailored for you.
• 27:26
  We, we know the hyper personalisation is quite a
• 27:29
  terminology, but in reality, in the world of food and grocery
• 27:34
  and, you know, providing the customer convenience, the hyper
• 27:38
  personalisation is that tailored individual approach for, for,
• 27:42
  for that particular person who is using it and, and availing
• 27:46
  those services which we are providing.
• 27:49
  And, and that's, that's the vision.
• 27:52
  So it could be, you know, the, the persona could be a single
• 27:55
  mother or you could be a working professional, or you could be
• 27:59
  something like me. I mean, I'm, I'm health freak
• 28:03
  for a couple of months and then I completely go nuts.
• 28:06
  And then I'm so it's like versatile personality.
• 28:10
  But if that agent or that personal shopper still
• 28:13
  understands me and you know, tailor those those
• 28:17
  recommendations, I will be like glued to it forever you.
• 28:21
  Know what I just realized and I would love to see data on this.
• 28:23
  You probably have it because we're, we're talking to ifood,
• 28:28
  we're talking to all these different companies in the
• 28:31
  process group that are, are doing and trying this innovation
• 28:34
  stuff and with the ifood like magic chatbot that they created.
• 28:42
  One thing that is difficult with chat is getting more context.
• 28:48
  But with voice, I just naturally am ready to give up more
• 28:53
  information. Have you tried both of those
• 28:56
  modes and you see a big difference?
• 28:58
  I I can imagine like I explain much more when I'm speaking
• 29:02
  through voice than I do when I'm chatting.
• 29:04
  Because chatting, first of all, just like holding and looking at
• 29:07
  my phone, nobody wants to do more of that during the day.
• 29:11
  Yeah, I fully agree. Absolutely.
• 29:13
  So, you know, a concrete example of that might be, hey, I need a
• 29:17
  burger for lunch. And you know, the assistant
• 29:20
  starts pulling those options up for you if you're on, if you're
• 29:23
  on chat and you, you want to just make an adjustment and say,
• 29:27
  hey, I'm trying to lose weight. Like I said earlier, you might
• 29:31
  not want to do that interrupt the flow of the chat because it
• 29:34
  just doesn't feel natural. Right?
• 29:35
  On voice, you can just interrupt the assistant.
• 29:38
  She'll stop or he'll stop, listen to you and quickly adjust
• 29:42
  what she's recommending. So, yeah, voice is it is about
• 29:47
  natural language, right? So chatting with your e-commerce
• 29:52
  provider via text is a big move away from navigating menus and
• 29:56
  Guis. And that's where we've been for
• 29:57
  the last 20 years. Voice is even more natural.
• 30:00
  So I'd agree with that. But you know, there's there's
• 30:02
  space for both. And I can share my personal
• 30:05
  experience. So when we were testing voice
• 30:08
  internally, so within incubation, when we are done
• 30:11
  with any of our ideas, we just test internally.
• 30:16
  So I was, I had a party at home and you know, when you test your
• 30:22
  internal product, you just want to be very careful that you
• 30:27
  don't, you don't screw up the product itself, right?
• 30:30
  So firstly, I was giving very simple instruction or can I have
• 30:33
  a vegetarian burger or very simple and easy instruction
• 30:37
  which voice can help? And this was giving me a, you
• 30:40
  know, kind of a satisfaction and happiness that Oh yes, this
• 30:42
  product is working. This is the cool product.
• 30:44
  But I actually had a party on Sunday.
• 30:47
  And then I said, can you actually arrange a party for 10
• 30:51
  people, these many adults and these many children's?
• 30:54
  And don't tell me the recommendation.
• 30:55
  This is my budget. Just add things into my basket.
• 30:57
  I was just testing it and it actually happened, which was
• 31:01
  really great. And that gives me, you know, a
• 31:04
  clarity around when you talk, how much you express yourself
• 31:09
  exactly. Because to your point, if the
• 31:11
  same scenario happened with the text or typing, I might not type
• 31:16
  this much like, you know, Yep, I might have said something more
• 31:20
  like I'm feeling anxious or I'm feeling this and that, but.
• 31:23
  You're so much more direct when you type because it is so much
• 31:26
  more cumbersome to actually type, especially if you're on
• 31:29
  the phone. And I've noticed this just using
• 31:31
  a lot of these speech to text tools that I I am way more
• 31:39
  verbose when I talk. So usually on a lot of this
• 31:43
  speech to text tools, I'll have to pare it down or I will make
• 31:47
  sure that I know what I want to say before I actually hit the
• 31:50
  start dictating button. But when I'm I'm talking, I'm
• 31:54
  giving all this context and I find that I'm much more lazy to
• 32:00
  type now because I've gotten used to the mode of talking.
• 32:05
  Yeah. Yeah, but this voice agent, the
• 32:09
  moment you want to see like, you know, once it says that, OK,
• 32:12
  these are the things in your basket you can switch easily
• 32:16
  because you can see the cards you can swap and make decision
• 32:20
  whether you want to go for that dish or not.
• 32:22
  So it's quite. Intuitive I imagine too.
• 32:25
  Recipes not in the details, but if you ask for recipe it can.
• 32:30
  Yeah, it will explain. Order me or yeah, plan my party
• 32:33
  for me. Basically, yeah.
• 32:35
  Order everything I need. Does it also say, Oh well, how
• 32:38
  about some flour arrangements? Yes, yes, it gives those
• 32:41
  recommendations, yes. And because we do maintain your
• 32:44
  order history and your dietary information.
• 32:48
  So it will also say that last time you tried the fries with
• 32:52
  this burger, would you like to add that in the basket, Things
• 32:56
  like that. So it's quite proactive to allow
• 32:59
  to help you to make decisions and react.
• 33:01
  And where does it stop? Because as you're talking about
• 33:04
  this, food is almost like the core, right?
• 33:08
  But then you start going out and you have tech and you have
• 33:11
  groceries, which is food. So that's not that big of a
• 33:14
  leap. But is there a world where I am
• 33:18
  travelling, you know, that I like exercising and there's a
• 33:22
  gym nearby. And so it's like, oh, well, we
• 33:26
  got your day pass at the gym if you want it.
• 33:28
  Boom. I can now go to that.
• 33:31
  So where do you have a line for what you want to do or is that
• 33:35
  kind of your job as the innovation lab to continue to
• 33:38
  expand the footprint? With the acquisition of Process,
• 33:41
  we now have such a bigger playing field and we can really
• 33:45
  start to think about the partnerships that this company
• 33:48
  has and the companies that Process owns and we can start to
• 33:52
  serve. So we set everyday convenience
• 33:54
  for food and retail. It can be way bigger than that,
• 33:56
  right. So like you mentioned, it can be
• 33:58
  travel, it can be hotels, it can be second hand goods, it can be
• 34:05
  you name it. So absolutely we see a world
• 34:07
  where I think there will be, like I said, there'll be a
• 34:10
  spectrum. You know, some people will
• 34:11
  prefer to use individual apps for those use cases, but the
• 34:15
  Super app concept, we see this $0.10 and WeChat is huge on this
• 34:19
  and they have been for many years.
• 34:22
  It works, right? And then if you think about, you
• 34:25
  know, what's happened in the last couple of weeks with
• 34:27
  Cloudbots, Maltbot, whatever it's called now, probably
• 34:30
  changed their name today. Open, Claude.
• 34:32
  Open, Claude. Yeah, yeah.
• 34:33
  So that was my heart take the other night.
• 34:35
  It's like, I don't know, it's a very techie sort of risky thing
• 34:38
  right now, proposition right now.
• 34:40
  But very quickly, you know, I believe that consumer
• 34:43
  expectations are going to move in this direction.
• 34:45
  So what does that mean? That means you should be able to
• 34:47
  either have an agent that can accomplish those tasks for you,
• 34:51
  objectives at a high level, plan my next week.
• 34:55
  You know what's going on, you have access to my calendar, you
• 34:57
  have access to my e-mail. You know what I like, you know
• 34:59
  my preferences and I can make really highly tailored
• 35:03
  personalized recommendations and actually do the things for you.
• 35:06
  Book the flights, book the hotel, contact your friends even
• 35:09
  and set up like a group chat so that you can meet up.
• 35:12
  All of those things are absolutely going to be possible.
• 35:14
  And I think it's a it's kind of like a chicken and egg thing,
• 35:17
  right? So we can start building that
• 35:18
  tomorrow, But are consumers going to be interested in that?
• 35:21
  Are they going to be worried about their privacy?
• 35:23
  It's really encouraging to see things like like Claude, I would
• 35:30
  say sort of paving the way because that's, you know, this
• 35:33
  is, it seems like a little sides sideshow right now.
• 35:37
  But once consumers get wind of that and and again, like AI
• 35:41
  seems like so much more commonplace now.
• 35:43
  You know, my, my folks are more like they know what's going on
• 35:46
  in AI. It's not, it's not backroom
• 35:48
  stuff anymore. Once somebody figures out a way
• 35:50
  to wrap that in a very convenient consumer packaging,
• 35:55
  people will start to think, Hey, why don't I have my own
• 35:58
  personalized agent? Can I just make my own?
• 36:00
  What are the other options out there that can do these things
• 36:02
  for me that can really simplify my life?
• 36:06
  And the beauty is that technology allows us to achieve
• 36:09
  all these goals. Yeah, we have everything in
• 36:13
  place to ensure that we can actually do these things.
• 36:17
  So just to answer your question again like the vision can be
• 36:22
  broader and broader based on you know the there is so much
• 36:28
  potential out there and with this new era like identity
• 36:34
  commerce and billing agents and like couple of steps it's all
• 36:38
  achievable. With infinite options now and
• 36:43
  kind of all this blue space to run in, how do you weigh what to
• 36:51
  tackle next? It's about combining a couple of
• 36:54
  things, right? So we need to understand what
• 36:58
  consumers want now. We need to have opinionated bets
• 37:02
  on what consumers are going to want in the future.
• 37:05
  And we combine that with what the company strategy is, what
• 37:07
  the company vision is and what our business goals are as well,
• 37:10
  right? So how do you balance building
• 37:13
  all the things with concrete growth expectations in the short
• 37:18
  term? So how we do that in the
• 37:19
  incubator is we invest heavily in what we call scanning the
• 37:23
  horizon, right? So we have teams that are really
• 37:26
  thinking into the future and talking to customers, looking at
• 37:31
  all of the tech innovations that are happening in our markets and
• 37:34
  in states as well and in China, just so we can really have those
• 37:38
  strong opinionated views on what we believe is coming next.
• 37:42
  And then we work in, you know, it's, it's an old paradigm now,
• 37:46
  but it still really works. The lean startup, right?
• 37:49
  So we really think about desirability, feasibility and
• 37:53
  viability and depending on the ideas and the hypothesis.
• 37:57
  Hypothesis are very, very important to how we think,
• 37:59
  right? So from scanning the horizon,
• 38:02
  we'll have a backlog of hypothesis.
• 38:04
  The outcome of that we'll think about, right?
• 38:07
  So we'll look at these and we'll decide whether we want to start
• 38:09
  with desirability or feasibility of viability in order to get to
• 38:14
  a stage gate, right? And in some cases, it'll be
• 38:19
  we'll want to look at feasibility first.
• 38:20
  So we'll work with a partner and we would think about what's the
• 38:25
  leanest thing we can do in terms of a tech POC, just to see what
• 38:29
  we need to do in order to validate the hypothesis of can
• 38:33
  we integrate, right? So can we get the data that we
• 38:36
  want into our systems and then decide what to do with that
• 38:41
  next? So that would be our first stage
• 38:43
  gate. That'll be an example of a tech
• 38:44
  POC. In other cases, we might want to
• 38:46
  validate desirability on the consumer side.
• 38:49
  So do customers even want this? Might they want it in the
• 38:51
  future? And there's a few ways we go
• 38:53
  about that. The good old painted door still
• 38:55
  works, right? So you put something in front of
• 38:57
  customers that could just be a question, are you interested in
• 39:03
  this thing? You can create like a button
• 39:06
  that just reveals a sign up For more information.
• 39:08
  We're working on this, something like that.
• 39:10
  And of course, like user research is super important as
• 39:13
  well, just to get that really to understand like what the
• 39:15
  appetite is. So we work in really tight stage
• 39:18
  gates and we try to validate and invalidate at the shortest
• 39:21
  possible cycles possible. So we're talking like one week
• 39:24
  turnarounds right during that process.
• 39:26
  So that's the classic way that many startups work.
• 39:30
  We're in a huge global company, so we need to operate a little
• 39:33
  bit differently. We operate in that way, but it's
• 39:35
  very important for us to leverage our really the
• 39:38
  resources that we have, which is we have scale, we have market
• 39:42
  teams, we have very strong commercial teams, we have
• 39:45
  stakeholders, we have deep, you know, design resources, data,
• 39:49
  etcetera, etcetera. So along that journey, we're
• 39:52
  taking everybody along for the journey, making sure that we
• 39:54
  have strong alignment on the storyline, why we're doing what
• 39:58
  and we need to drive alignment from the beginning.
• 40:00
  The ultimate goal of any of these tests or hypothesis or
• 40:05
  POC's is to get through the stage gates that are unique for
• 40:09
  each of the ideas and then ultimately to what we call
• 40:12
  graduated, right? And that means effectively
• 40:15
  moving it from incubation into scaling, early scaling or maybe
• 40:20
  an MVP. And at that point, usually
• 40:22
  there's a lot of. Alignments and collaboration
• 40:26
  with other parts of the organization.
• 40:27
  You know, we're a 25 year old company.
• 40:30
  There is a lot of complexity under the hood.
• 40:31
  So once we've validated the right things in the incubator,
• 40:36
  we need to be really, really smart about alignment.
• 40:39
  And that goes for tech as well. Very, very important to have
• 40:42
  alignment on what the goals are, what the expectations are.
• 40:46
  As soon as we have that initial signal, we don't just, you know,
• 40:49
  announce it, you know, three months later or two months
• 40:51
  later, surprise, this is your product.
• 40:53
  We take the guys along for the ride with us.
• 40:56
  So there's an art to that. Effectively, it's corporate
• 40:58
  innovation. And I've done a lot of startup
• 41:01
  work. And this is the thing about this
• 41:03
  is a big difference about innovating in a big company like
• 41:06
  this. There are so many powerful
• 41:08
  levers that you can leverage. You need to look at them as
• 41:11
  opportunities. You know, some people might look
• 41:13
  at them as as hindrances or things that slow you down.
• 41:16
  You can use them to speed you up.
• 41:18
  Even though this process looks like it's like a, you know,
• 41:21
  sequential process and we have different steps to do it, it's
• 41:26
  all happens in parallel. So all the time we have many
• 41:32
  ideas coming from all over the places.
• 41:34
  We have a research team, they come up with the innovation
• 41:37
  ideas. We have engineering people who
• 41:41
  have all some, most of the time they have different type of
• 41:44
  problems and they come up with the ideas.
• 41:46
  And then our organizational strategy is the key.
• 41:50
  So we balance these two together and we just, you know, look at
• 41:55
  the best possible bets. And then most of the time we are
• 42:01
  playing with uncertainty and we are traversing with ambiguity.
• 42:05
  And we have to ensure that we just, you know, have an
• 42:10
  engineering capacity in place to ensure that we are at least
• 42:14
  trying all the ideas even though we fail.
• 42:18
  So failing is a learning and failing is an experience.
• 42:22
  So we don't wait for everything to become the most perfect, you
• 42:27
  know, thing. We just, we just, we have two
• 42:29
  types of, you know, innovation and ideas most of the time. 1 is
• 42:33
  where the technology is evolving.
• 42:35
  Every day there's a new protocol, every day there's a
• 42:37
  new tool and new technologies to integrate and to see out of the
• 42:41
  possibility. So we really want our engineers
• 42:44
  to jump on it and see what we can do what, what type of
• 42:48
  outcomes we can achieve for our business.
• 42:51
  But at the same time, what Guthrie has explained, we do
• 42:53
  follow that process in parallel as well.
• 42:55
  We ensure that, you know, our ideas, even though they are
• 42:59
  failing, at least one of them is becoming a, an impactful version
• 43:04
  within the organization. And we have so many examples
• 43:06
  like that. Last year has been really
• 43:08
  successful where our team was able to, you know, fail with
• 43:13
  some, but was also able to succeed with many ideas.
• 43:17
  And that's how this whole process work.
• 43:20
  So it's, it's like starting with the real problem and the real
• 43:23
  challenge because idea can be weak.
• 43:26
  You know, there should be some practicality, some tangible
• 43:31
  action, some real problem to be solved.
• 43:34
  And that's I think our intake process.
• 43:37
  That's where our previous, our past experience as well as our
• 43:42
  learning and as well as where the market is evolving.
• 43:45
  Blending all these things together, we decide like which
• 43:48
  one is the best one to pick and to move.
• 43:52
  Yeah, really defining the problem.
• 43:54
  Yeah, defining the problem is the key.
• 43:56
  Yeah, it can be just a gut feeling, like I think it's this,
• 43:59
  but then you shape it and you're creating that.
• 44:02
  Yeah, because we can't just run after all the shiny things in
• 44:05
  the market, right? And AI is also not a solution
• 44:10
  for every problem. So defining the problem is
• 44:13
  important. That's what I think we focus on.
• 44:16
  So once you define the problem, how you fix the problem is
• 44:20
  completely different thing. You can easily, you know,
• 44:23
  sometime you can, can use AI and most of the time it just, it's,
• 44:27
  it's just simple fix. But being an incubator, we are
• 44:32
  not there to fix the near term problems because those problems
• 44:36
  are already part of our strategic map and other product
• 44:39
  teams are already working towards that.
• 44:41
  It is in their own map. We are just looking ahead at
• 44:44
  least one or the two years down the line and and figuring out
• 44:48
  the moon shots and the kind of evolution we'll see in future.
• 44:54
  Yeah, that's cool. You get to kind of play out all
• 44:57
  these different scenarios, potential scenarios of where
• 44:59
  things are going and how to strategically position yourself.
• 45:03
  Yeah, a lot of hypothesis, a lot of uncertainty and ambiguity,
• 45:08
  like, you know, playing with different things.
• 45:11
  Sometimes I imagine like you were saying, like, yeah, we can
• 45:13
  be directionally correct on this, but the market isn't ready
• 45:16
  for it yet because of some reason or other.
• 45:19
  Yeah, exactly. So I've got an example actually
• 45:22
  what what to build a little bit of color and what what he's
• 45:24
  saying. So we, one of the hypothesis
• 45:26
  that we had last year was, so the traditional app search is
• 45:34
  one of the most important interfaces, right?
• 45:37
  There's a few different ways that people search.
• 45:40
  One of the ways that people search sometimes is
• 45:42
  conversationally. But at a certain point last year
• 45:46
  we didn't support conversational search.
• 45:48
  So we, the hypothesis was that, you know, a certain subset of
• 45:51
  customers would like to ask questions in search, much like,
• 45:55
  you know, they do in many other products, including all of their
• 46:00
  AI agency stays. So we looked at that and we had
• 46:03
  to work with, you know, incubator had to work with the
• 46:07
  customer team, which is the team that owns, you know, the
• 46:09
  customer interfaces, the search team within that team.
• 46:13
  And was really, you know, to my point earlier about building
• 46:17
  strong connections and a shared view on what the outcomes are
• 46:20
  going to be. The potential to the business
• 46:24
  potential and the product potential and the customer
• 46:26
  delight potential of adding that feature was the story that we
• 46:29
  had to tell the whole way to really bring everybody on board.
• 46:32
  What we found though, when we were looking at the data, this
• 46:35
  is really interesting is that, you know, so conversational
• 46:38
  search, you could argue is going to add value in the short term,
• 46:42
  but it's probably also, you know, a mid to longer term bets,
• 46:45
  right? What is the?
• 46:46
  Customers will expect it. Yes, exactly.
• 46:48
  They might not expect it now, right.
• 46:50
  So there's only a small amount of customers who are just
• 46:52
  naturally asking questions because we're not prompting them
• 46:55
  to do. That I was going to say like,
• 46:56
  how did you tell people that you now support that?
• 46:59
  Exactly. So it's that's customer
• 47:01
  education, right. So it's actually in play right
• 47:03
  now. It should be launched in the UK
• 47:04
  in the next couple of weeks. Is that real onboarding?
• 47:07
  Like hey, here are some suggestions on conversational
• 47:10
  search terms and by the way, you can type whatever you want in
• 47:14
  here. Ask a follow up question.
• 47:15
  Exactly right. So, so it's that when we were
• 47:18
  looking at the data though, we found that there's a good amount
• 47:22
  of searches that end in 0 results.
• 47:26
  Common problem across most consumer apps, right?
• 47:30
  So people just aren't finding what they want because of
• 47:33
  spelling mistakes or. You know, yeah, that's me.
• 47:36
  Fill in the blank exactly so true.
• 47:38
  So semantic search is one of the ways that the, you know, we're
• 47:41
  tackling that the conversational search managed well.
• 47:45
  It was actually almost like a side benefit of what what our
• 47:48
  initial hypothesis was, is that conversational search solved a
• 47:53
  big percentage of those dead ends you could call it, right.
• 47:56
  So which ended up having a revenue impact in the short term
• 48:00
  that was significant. Yeah.
• 48:02
  So by thinking about conversational search really
• 48:06
  driving you know the short term innovation and trying to and and
• 48:10
  getting that into the product, we managed to solve a real
• 48:13
  problem in the short term much more related to you know the big
• 48:17
  OK Rs which is really around revenue and growth.
• 48:20
  What are all the things that didn't make it through this
• 48:22
  great filter? What are things where you as the
• 48:25
  Innovation Lab thought, oh, of course this is the way the
• 48:28
  future is going to be, we should put this into play?
• 48:31
  And then you looked at it and you were like, actually, it's
• 48:34
  not going to work. One of the example is sending
• 48:38
  notification based on your location.
• 48:41
  We called it location by notification.
• 48:43
  So we wanted to notify customers for a collection order when they
• 48:52
  are near to a restaurant, let's say 1015 meters.
• 48:57
  And that the idea was that our partners will get some benefits
• 49:02
  because they are doing collection orders.
• 49:04
  And sometimes customer don't know that this particular
• 49:07
  restaurant had this preposition and other way helping the
• 49:13
  customers for customer convenience.
• 49:16
  And we started thinking about, you know, how to fix the
• 49:18
  problem, but the kind of challenges we had was first the
• 49:23
  legal, you know, constraints you it's very difficult to track
• 49:30
  customers live location all the time.
• 49:33
  And if we do that, there is some physical devices needs to be
• 49:39
  attached to you know, the stores or the partners.
• 49:44
  And that means there is operational overhead because you
• 49:48
  need to maintain those physical devices which keeps on taking
• 49:51
  signal through the your mobile device.
• 49:53
  And then we are app gets those signal and send it to customer.
• 49:56
  And in the end, if you look at the value, as we discussed
• 50:00
  earlier, it could actually become a customer grievance
• 50:04
  instead of customer, you know, delight.
• 50:07
  So all of a sudden you're looking at this and you're
• 50:08
  saying we have to put this huge amount of effort into putting in
• 50:14
  devices into all the merchants shops?
• 50:16
  Yeah. Only to send more notifications,
• 50:19
  which the end user probably isn't going to like.
• 50:23
  Yeah. So that.
• 50:24
  Yeah, I could see how that one died.
• 50:25
  Yeah. But then we also thought about,
• 50:28
  you know, experimenting, experimenting with the simplest
• 50:30
  thing. So we, we thought about not to
• 50:32
  put any physical device, not even ask our customer to share
• 50:36
  their live location with us. Just, you know, have some
• 50:41
  reasoning model in the background which will keep
• 50:44
  learning from the customer behavior.
• 50:45
  Like you order coffee every day at 9:00 AM or you order lunch at
• 50:49
  1:00 PM. So we just send some, you know,
• 50:53
  blind notification at that point of time based on your order
• 50:57
  history and your behavior. And we designed it and
• 51:02
  technically it was achievable. But then again, when we
• 51:05
  evaluated, you know, the impact versus value and as we
• 51:10
  mentioned, you know, how do we, what is our intake process?
• 51:13
  We have some parameters and based on those parameters and
• 51:17
  all the conversation we had, we just dropped this idea.
• 51:20
  Didn't look good. No, you know what it really
• 51:23
  reminds me of? Your jobs are very much in line
• 51:28
  with who's that marketer that's famous in the UK.
• 51:33
  He, I think he worked at Ogilvy. He wrote a great book.
• 51:36
  I'm literally was, I just finished it.
• 51:40
  I think it's Roy Sutherland. Do you know him?
• 51:43
  Rory. Rory Sutherland.
• 51:46
  Maybe Sullivan. I don't.
• 51:47
  I'll figure it out. What's the book called?
• 51:51
  I'll have to pull it up on my phone because I I was listening
• 51:53
  to it. And he gives all of these
• 51:57
  antidotes where you think through something and you think,
• 52:00
  oh, yeah, this is the logical thing.
• 52:03
  But then when you put it out there, customers and consumers
• 52:08
  do the opposite, depending on how it's framed.
• 52:11
  So the one that I think he had a tech talk or Ted Talk on back in
• 52:15
  the day that I first saw and then he adds in the book that
• 52:19
  really captured my imagination was he said the UK government is
• 52:24
  proposing to spend billions of dollars to make a train, a high
• 52:29
  speed train that will take a trip from London to Manchester
• 52:34
  or London to Liverpool and make it 30 minutes faster.
• 52:38
  So billions are getting sunk into this high speed train.
• 52:41
  And he said, what if we just made the current train more of a
• 52:48
  delightful experience so that the customers don't care if
• 52:52
  they're on a three hour train or a 2 1/2 hour train ride.
• 52:55
  Yeah, yeah, yeah. And that's the challenge, being
• 52:59
  an incubator like space, because it's really hard to find that
• 53:04
  balance between, you know, choosing a perfect idea versus
• 53:09
  something which is, which may not work or which is not logical
• 53:13
  at all. This is the real challenge.
• 53:15
  We sometime get emotional with what we are are doing because we
• 53:19
  are actually putting our energy and our focus into it and it's
• 53:23
  really hard to be rational about it and make a decision that this
• 53:27
  is the right time to stop or pivot or keep going.
• 53:31
  We do have those type of frameworks and you know, kind of
• 53:36
  ceremonies where we challenge each other, we make each other
• 53:39
  accountable. But it's still like, you know,
• 53:42
  very difficult that when you are spending a lot of your time into
• 53:46
  a product or an idea to execute it and see the art of
• 53:49
  possibility and you've been asked to stop it or pivot to a
• 53:55
  different direction. This is a challenge, I would
• 53:58
  say. And it's it's really hard to
• 54:00
  nail it down. What is it called?
• 54:02
  Like some cost, Some costs. Yeah, exactly.
• 54:06
  And you're, you feel ownership and you don't want to Kill Your
• 54:09
  Darlings, but you have to have that like ruthlessness to you in
• 54:14
  a way. Or park it for some, even
• 54:16
  parking for some time is is also challenging.
• 54:20
  It's it's difficult to make that decision.
• 54:21
  OK, let's pause it. Let's do something else.
• 54:24
  Yeah, well, we find ourselves in that position quite often
• 54:26
  actually, right? So sometimes we're just too
• 54:28
  early, so we'll put something on ice.
• 54:31
  We did that with our integrations with Mercedes-Benz.
• 54:34
  You know, so we have an in car ordering app with Mercedes-Benz
• 54:37
  right now. We worked on that for, you know,
• 54:42
  probably a year, maybe six months before really landing the
• 54:47
  contracts and the commercials etcetera, etcetera.
• 54:50
  Working with a company that is not natively a tech company,
• 54:53
  something we've learned it's it's certainly interesting.
• 54:57
  Yeah, because you know, the difference between working with
• 54:59
  a a Meta or a Google versus, you know, automotive OEM night and
• 55:05
  day. But to go back to our vision, if
• 55:08
  we want to be ubiquitous, this is the playing field we need to
• 55:11
  operate in, right. So, so in some cases you need to
• 55:13
  put something on ice for three months, six months, which in our
• 55:16
  world is like a lifetime, but in another world might be, you
• 55:20
  know, a very short amount of time.
• 55:22
  So it's interesting to balance those two things.
• 55:24
  And for us, not just experimentation is, you know,
• 55:30
  something which we are, we are we are focused on, but also
• 55:36
  measuring our successes. Experimental is experimental as
• 55:40
  well. So when it's very hard to
• 55:43
  measure your success based on an incubated idea, it's like, you
• 55:49
  know, you put something out there like this NLS, we can
• 55:54
  market it when it is not fully ready to be, to be shared and to
• 55:59
  be talked about. But at the same time, we want to
• 56:02
  learn from it and we want to improve it.
• 56:04
  So our success is also like, you know, quite, quite, quite
• 56:08
  iterative and experimental. So we have, I found out our way
• 56:14
  to to, you know, measure our success with different
• 56:16
  parameters. We just have, we just say when
• 56:20
  we define the problem, just think about what's the outcome
• 56:24
  or what's the clear success. We are looking here.
• 56:26
  Hypothesis is one thing, but what will be the real parameters
• 56:30
  and measuring points? Sometime it could be as simple
• 56:34
  as just customer engagement. Most of the time it could be
• 56:38
  completely like, you know end goal which is order conversion
• 56:42
  and that few time it is just the exploration or setting up the
• 56:46
  foundation so that we can enable something else in future that
• 56:51
  also happens. And in identity Ki world, before
• 56:54
  we even start building new agents, we wanted to think about
• 56:58
  how the data will be perceived because we need structured data
• 57:03
  as a prerequisite to build the potential agents.
• 57:07
  And that's the foundational work.
• 57:09
  And that's quite, quite, quite a tech, you know, experiment
• 57:12
  rather than a clear business objective or a business goal.
• 57:16
  So finding out what kind of success we are looking for is
• 57:20
  the key thing when we define the problem.
• 57:23
  And then making sure that we have our internal, you know,
• 57:26
  testing people or the group of people who can test and
• 57:31
  genuinely give us the feedback so that we can improve, iterate
• 57:35
  and have that feedback loop in place.
• 57:37
  And once it goes to market, then also we go with the, you know,
• 57:40
  the the lowest risk possible. We just roll it out to 5% people
• 57:44
  to 1 country and sometime to even one region so that customer
• 57:49
  can share the feedback. And then we just go into that
• 57:52
  feedback loop. And once we feel like, yes, it
• 57:55
  is now ready to be talked about and to market, then we launch it
• 57:59
  fully. And Voice Agent has gone through
• 58:01
  this phase for a longer period of time.
• 58:04
  The product which we see today wasn't like this.
• 58:07
  It was. It has gone through that
• 58:09
  iterative cycle of improvement before we have put it across the
• 58:14
  consumers. What were some of the hard
• 58:16
  challenges creating voice agent? I think the first challenge was
• 58:23
  the idea itself because when we started it was not in something
• 58:28
  which others, you know, competitors or other vendors
• 58:33
  were using. So finding out the right
• 58:36
  technology, because there are so many nuances when it comes to
• 58:40
  voice, we need to make sure that latency is accurate, the quality
• 58:45
  is there, the conversation is intent based because we were not
• 58:49
  aiming for a voice mode, We were aiming for a voice companion, a
• 58:54
  personal voice agent to really understand you and give you an
• 58:59
  option of being, you know, switching between the visuals
• 59:02
  and the audio. And that's where the real
• 59:05
  challenge was to actually technically deciding which is
• 59:08
  the right path to go for it. And then it has gone through a
• 59:12
  lot of back and forth. There had been challenges
• 59:15
  between the UX and the audio quality because it has both the
• 59:20
  versions, right? So you can see the cards you can
• 59:22
  swap across and you can choose the right dish for you.
• 59:26
  But at the same time you are on the audio.
• 59:28
  So you have to make sure that the customer experience is
• 59:31
  balanced. And it's it's basically to the
• 59:34
  point for both audio as well as your visuals.
• 59:38
  And that's where the UX part plus the core engineering, the
• 59:41
  server based or the back end based engineering have been
• 59:46
  through some challenges. Yeah.
• 59:48
  I imagine you can't take for granted that somebody's going to
• 59:51
  be looking at their phone if they're on voice mode.
• 59:53
  Yeah, you have to have the experience be flawless only on
• 59:56
  voice, and then you sprinkle on some delight with the.
• 1:00:01
  Visuals and yet the then you can go back to voice, you can switch
• 1:00:05
  between. So which means you're not just
• 1:00:08
  building a voice agent, you're also building an intuitive
• 1:00:12
  experience for your consumers. How did you deal with latency
• 1:00:16
  constraints? How did you deal with all these
• 1:00:18
  challenges? So we are using OpenAI Real Time
• 1:00:23
  API and with that integration we have we have server based
• 1:00:29
  approach. So we just as soon as you start
• 1:00:33
  the conversation, even before you start the conversation, we
• 1:00:35
  load the whole context in the background so that you have the
• 1:00:40
  smoother experience and the latency is being reduced at that
• 1:00:45
  point only because when you when in the in the background, you
• 1:00:49
  have the context. So by that time agent actually
• 1:00:52
  knows most of the things about you.
• 1:00:54
  And when you say context, I imagine there that is a very
• 1:00:59
  difficult thing at context. Engineering is huge, right?
• 1:01:01
  And it's huge because it is difficult.
• 1:01:04
  So how do you know what to weigh?
• 1:01:06
  Are you just giving them the last 10 orders?
• 1:01:08
  Are you just like? How do you explain that prompt
• 1:01:11
  so that it is contextually relevant?
• 1:01:14
  So as of now it is quite simple and standard.
• 1:01:18
  So it is based on the order history and some level of
• 1:01:22
  personalization for your dietary requirements.
• 1:01:25
  So context is the scope of the agent is very simple and concise
• 1:01:31
  so that it do not hallucinate and it you know, sometime prompt
• 1:01:35
  engineering and context engineering is being heavily
• 1:01:39
  loaded for one engineer, sorry, one agent and that becomes
• 1:01:43
  actually cumbersome for the agents to respond appropriately.
• 1:01:46
  So we wanted to make it a lightweight easy to communicate
• 1:01:50
  and easy to easy to be consumed. That information which the agent
• 1:01:55
  is giving you should be, you know, very much tailored for
• 1:01:58
  you, but at the same time, not every possible information you
• 1:02:02
  have shared with us is going to be in the context.
• 1:02:05
  So it's very simple context and it's even based.
• 1:02:09
  So there are multiple events in parallel.
• 1:02:11
  So if you are asking a complex question like I asked that I
• 1:02:14
  want to arrange a party for these many people.
• 1:02:16
  So based on my my query, it just, it just surface multiple
• 1:02:21
  events in parallel so that there's a speed, there is
• 1:02:24
  accuracy and the even the UI structure is being ready in the
• 1:02:28
  background even before you see it as a visual.
• 1:02:32
  So a lot of processing is happening in the background for
• 1:02:35
  voice agent. And that's why once you interact
• 1:02:38
  with it, it's quite fast, it quite accurate, and it gives you
• 1:02:43
  that personalized recommendation because it will not just tell
• 1:02:46
  you that try this pizza, it will also tell you because you're a
• 1:02:49
  vegetarian, you might like this vegetarian pizza.
• 1:02:53
  And you can start the conversation as a simple thing
• 1:02:56
  like, you know, I'm looking for something spicy.
• 1:02:58
  Or you can also give a complex query like I'm looking for
• 1:03:01
  something spicy, which is vegetarian, but I am hungry, so
• 1:03:04
  I can't wait more than 30 minutes.
• 1:03:07
  And all these processing is server based processing in the
• 1:03:10
  background, which is quite fast. And a lot of information is
• 1:03:14
  cached as well, because your conversation is cached the whole
• 1:03:19
  conversation and that also gives the speed and accuracy.
• 1:03:23
  It's interesting you talk about how you almost preload things
• 1:03:27
  just in case, because that is a very common design pattern I've
• 1:03:30
  been seeing a few folks use, especially when it comes to
• 1:03:34
  voice. Because you'd almost rather have
• 1:03:37
  it and not need it than need it and not have it and then make
• 1:03:40
  the user wait since it is so high stakes, the user, if it has
• 1:03:44
  to wait 30 seconds, it might think something's wrong,
• 1:03:48
  especially if you're not getting any voice cues that oh, sorry,
• 1:03:51
  I'm still searching. Oh wait, I've got something.
• 1:03:53
  Let me check. You could do a bunch of fun
• 1:03:55
  stuff with the voice cues, but 30 seconds is still for 30
• 1:03:58
  seconds. And so if you can shave some
• 1:04:01
  time off of that because you already have the latency of the
• 1:04:05
  model which is going to be. Yeah, which you can't avoid.
• 1:04:10
  So if you can have all of the data and you can have all of the
• 1:04:13
  stuff that it needs there, and then you decide whether or not
• 1:04:18
  you show it to the user or you decide whether or not it's
• 1:04:21
  relevant. I've seen that.
• 1:04:23
  And then I've also seen constellation of models, not
• 1:04:26
  just one. It's not just hit one model and
• 1:04:29
  then you come back and you've got it.
• 1:04:32
  But they'll have many different agents hanging out and they're
• 1:04:37
  ready in case there is a moment they need to go and grab
• 1:04:41
  something. And they're just observers of
• 1:04:43
  the conversation and they know, oh, if I hear that keyword, I
• 1:04:48
  should go do this. Or if I but they're not jumping
• 1:04:50
  in, they're not the agent that the user is interacting with.
• 1:04:54
  So the similar concept here. So when I say parallel
• 1:04:57
  processing, so we have tools in this case.
• 1:05:02
  So while you're asking for a pizza.
• 1:05:05
  So there will be 3 parallel tools being called, one to load
• 1:05:10
  the menu, another to load the partner information and these
• 1:05:14
  are cached. So because partner information
• 1:05:17
  is something which we are not going to change, we can easily
• 1:05:19
  cache it. And that also allows some level
• 1:05:22
  of you know, faster speed and ability to, to use it in the
• 1:05:26
  most intuitive way. And this processing makes it
• 1:05:30
  unique and better. And also the biggest learning is
• 1:05:36
  being fine tuning it. So starting with very basic
• 1:05:40
  structure, which is just solving the basic purpose, which is just
• 1:05:46
  to understand the intent of the customer and giving the
• 1:05:50
  recommendation. But then fine tuning the prompt
• 1:05:53
  after every iteration and based on the conversations we have
• 1:05:58
  that prompt. System prompting has had been a
• 1:06:01
  key here because we wrote a lot of integration tests with so
• 1:06:07
  many hypothetical scenarios based on different kind of
• 1:06:10
  conversation customer can have. And based on those integration
• 1:06:14
  tests, we fine tune our prompting system.
• 1:06:17
  Our text assistant has been a market for longer than our Voice
• 1:06:20
  Assistant. Right.
• 1:06:21
  Voice is brand new and there are some subtle ways that those two
• 1:06:25
  products differ. So there's a threshold on the
• 1:06:27
  text assistant where if you speak to it, I believe it's if
• 1:06:30
  you have 5 conversations within the space of two weeks, it will
• 1:06:36
  retain more memory than if you have not hit that threshold,
• 1:06:40
  right? And how that plays out for the
• 1:06:41
  consumer is you come back after that 6th conversation and it
• 1:06:45
  will actually tell you a little bit about your preferences in
• 1:06:49
  text form, and it'll make a subtle recommendation and maybe
• 1:06:52
  what you want to try next. So that's something we haven't
• 1:06:54
  taken into our voice product yet, but you can imagine that
• 1:06:57
  we're a great conversation starter, right?
• 1:06:59
  Yeah. Yeah, and and text assistant has
• 1:07:03
  been actually very much useful for voice agent because those
• 1:07:07
  conversation helped us to shape up, you know the integration
• 1:07:11
  testing and prompting for the voice.
• 1:07:13
  Yeah, what people are asking what the happy paths are, how
• 1:07:17
  quickly they need things that that makes sense.
• 1:07:19
  And also I can see going back to I worked out today, the Voice
• 1:07:25
  Assistant knows that when I open it, it says, hey, how was the
• 1:07:28
  workout? You want me to order some?
• 1:07:30
  Yeah, protein. Yeah, exactly.
• 1:07:32
  Exactly. Yeah.
• 1:07:33
  So sometimes we start with, you know, a very limited feature in
• 1:07:38
  a very limited region. Other times we just put a single
• 1:07:42
  robot on the road. Segway to drones and robotics.
• 1:07:46
  Yeah, if you're ready for that. Yeah, yeah.
• 1:07:48
  When we think about AI, we have a virtual or a software based AI
• 1:07:53
  solution and the physical AI solution.
• 1:07:56
  And for physical AI solution we have drone delivery and robotics
• 1:07:59
  in place. For the virtual or the software
• 1:08:01
  solution, we need to think about our consumers, our partners and
• 1:08:07
  logistics and that's where you know the solution which we are
• 1:08:11
  building currently, they're primarily focused on consumer,
• 1:08:14
  but in parallel we have automation and agent take AI
• 1:08:18
  going on to serve our logistic teams and to serve our logistic
• 1:08:24
  partners as well. And we have sales agents in
• 1:08:27
  place to onboard our partners in the most automatic way.
• 1:08:33
  Now, when it comes to the physical AI, we have already
• 1:08:38
  launched our first partner in Switzerland.
• 1:08:43
  And as we talked about, you know how in incubation we start in
• 1:08:47
  the most experimental way, we go very lean.
• 1:08:51
  So I think that we can add more, but we started with the one
• 1:08:55
  robot just. 1. Just one robot just to see that
• 1:08:59
  of the possibility, just to see how it get that end to end
• 1:09:04
  delivery done because last mile delivery is something we want to
• 1:09:08
  achieve. But at this point of time our
• 1:09:10
  our robots are actually doing the end to end delivery.
• 1:09:13
  It was the leanest possible robot pilot you can imagine in
• 1:09:17
  that there was a human walking next to and it was a robot dog
• 1:09:20
  by the way. Right?
• 1:09:22
  So the partner there is RIVR and their form factor is like your
• 1:09:27
  classic 4 legged quad robot dog has wheels as well.
• 1:09:33
  So they're unique in that way. So it can ride and it can climb
• 1:09:36
  upstairs, etcetera. For the launch, we had a human
• 1:09:40
  walking next to the robot just to make sure that no, there were
• 1:09:43
  no issues. And it's an interesting thing,
• 1:09:46
  right? Public impression of a robot
• 1:09:51
  walking through the streets. Turn some heads.
• 1:09:53
  Indeed, indeed I've. Seen some videos of people
• 1:09:57
  attacking the autonomous vehicles that are coming through
• 1:10:01
  and delivering things. Or they try, they attack it just
• 1:10:05
  because they're taking. Yeah, and they're taking out
• 1:10:08
  their aggression on it. Or it is, it's like, hey, let's
• 1:10:11
  see if it can handle a little adversity.
• 1:10:13
  Or they attack it because they try and get the food that's
• 1:10:15
  inside. It's like, hey, there's some
• 1:10:17
  free food here and let's get that.
• 1:10:20
  Exactly. There's a lot to say about
• 1:10:22
  rights, job replacement and those kinds of things.
• 1:10:25
  Our vision with this is really that we're not going to be
• 1:10:28
  replacing our human couriers. We're augmenting them, we're
• 1:10:32
  making them safer. So the vision is to have an
• 1:10:36
  orchestrated last mile delivery system orchestrated between
• 1:10:40
  humans and robots with seamless handoff, right, Depending on
• 1:10:45
  things like weather conditions, traffic in.
• 1:10:48
  I feel they have this interesting position, a
• 1:10:51
  situation where they have a huge river, right?
• 1:10:54
  And it takes, although it's quite close to get to the other
• 1:10:56
  side of the river as as the crow flies, it takes an hour to drive
• 1:11:00
  because of the traffic. So that's that's that's their
• 1:11:04
  use case for the drone, right. So the drone takes the delivery
• 1:11:07
  across and then hands it to a human.
• 1:11:09
  Yeah. So it's all about orchestration,
• 1:11:11
  augmentation and then obviously expanding our reach.
• 1:11:14
  That's where the growth piece comes in for the business.
• 1:11:16
  There's two parts to that. First is expanding our reach so
• 1:11:19
  you can deliver to rural areas or areas that would take much
• 1:11:22
  longer for, you know, a driver to get to.
• 1:11:25
  And then the other one is that the IT reduces the costs of a
• 1:11:29
  delivery significantly. So that's, you know, as many
• 1:11:32
  mentioned, we're a three side marketplace consumers, partners
• 1:11:36
  meeting restaurants and grocery stores and then our logistics
• 1:11:39
  arm as well. So innovation happens across
• 1:11:41
  those 3 user types. And the most promising thing
• 1:11:45
  we're doing right now in in the third one is drones and
• 1:11:48
  robotics. We're delivering via drone in
• 1:11:51
  Ireland as well. We've partnered with a startup
• 1:11:53
  called Manna and people love it. Yeah, it's really interesting to
• 1:11:58
  see, right, because we again, back to how we work.
• 1:12:02
  So very iteratively, very hypothesis driven and we really
• 1:12:07
  kind of soft launched into it and we wanted to understand
• 1:12:10
  customers perception in the app if we gave them the choice.
• 1:12:13
  Do you want to deliver this via drone?
• 1:12:15
  It's a new thing, it's a novel thing, right?
• 1:12:16
  And that was what we some of the metrics we were looking at as
• 1:12:19
  well. It's like, is this just going to
• 1:12:20
  be an initial burst of novelty and then it'll tail off?
• 1:12:24
  But it hasn't. No, no.
• 1:12:25
  So is it? Because it's faster.
• 1:12:27
  People would want it. It's faster.
• 1:12:29
  Yeah, it's faster. Why?
• 1:12:30
  Because if you can fly, it's going to be faster.
• 1:12:32
  Right, exactly. And I think the, the novelty
• 1:12:34
  thing is part of it as well. It's just fun to get your order
• 1:12:37
  via Drawn, right? So we we see that a huge
• 1:12:40
  percentage of our Burger King orders in our in the town that
• 1:12:44
  we've launched in Blanchett's town in Ireland are delivered
• 1:12:47
  via Drawn. But it's not like I open my
• 1:12:50
  window on the 5th floor and the drone is hovering outside of my
• 1:12:54
  window. Not just like.
• 1:12:55
  No, there are compliances. We have to follow those
• 1:12:59
  compliances. So we have to maintain that
• 1:13:01
  distance from, you know, the buildings, the buildings and we
• 1:13:05
  have to make it safe. But I think the biggest
• 1:13:08
  advantage is to handle situations like pandemic are
• 1:13:13
  natural disasters and reaching where human cannot reach.
• 1:13:17
  Machines can do what human cannot do.
• 1:13:19
  And when it comes to delivering last mile and that is another
• 1:13:25
  aim that we still you know, manage the customer convenience
• 1:13:29
  even when there is the weather conditions are not very
• 1:13:33
  promising or when the things are not as natural and normal as
• 1:13:38
  they are and there are some, you know, casual calamities and all.
• 1:13:42
  Things get delivered sooner, they can get delivered further
• 1:13:45
  away. Me living in a rural area, I'm
• 1:13:47
  very happy to hear that because, yeah, I don't have the ability
• 1:13:51
  to use a lot of these food delivery services.
• 1:13:54
  Yeah, Since there's just not restaurants that are very picky
• 1:13:59
  about my food. And also, I live near a city,
• 1:14:02
  but it's a 40 minute drive. Yeah.
• 1:14:04
  So if I could get the food from that city delivered via drone,
• 1:14:08
  that would be amazing. Yeah.
• 1:14:09
  Yeah, and there's an operational efficiency as well using drones
• 1:14:13
  and robotics for last mile delivery.
• 1:14:16
  How so? Because of the speed, because of
• 1:14:20
  the amount of, you know, deliveries you can do parallely.
• 1:14:26
  So this is this is actually going to save cost as well as
• 1:14:32
  going to provide the better efficiencies.
• 1:14:35
  Yeah, you can scale up. Yeah.
• 1:14:38
  A lot. But by no means it is going to
• 1:14:40
  replace human or our couriers. We are still going to maintain
• 1:14:44
  that it is just an exit. Kathy mentioned that it is just
• 1:14:48
  an augmentation in our existing logistics.
• 1:14:51
  Yes. So concretely, you can imagine
• 1:14:53
  that a Courier is stuck in traffic jam, unexpected traffic
• 1:14:57
  jam, right? So he sees his ETA is going to
• 1:14:59
  be an hour. Now, this is not a great
• 1:15:00
  experience for the customer, obviously.
• 1:15:03
  So what we want to do is have the ability for him or her to
• 1:15:08
  say, let me hand this off to a robot or a drone to finish the
• 1:15:13
  delivery or to pick it up from the driver and hand it to
• 1:15:17
  another driver on the other side of the traffic jam, for example.
• 1:15:20
  Yeah, who gets the tip? The robot, of course.
• 1:15:23
  Right? And it's paid impact.
• 1:15:26
  I'm kidding. I'm kidding.
• 1:15:27
  There's one point we haven't touched upon is about what does
• 1:15:32
  AI innovation means to jet in general, because the innovation
• 1:15:37
  is not limited to just one team, it is actually across the whole
• 1:15:41
  organization. We are AI first organization, so
• 1:15:47
  we have our AI champions, the group of people who proactively
• 1:15:52
  work on the, you know, the tools in the market, play around with
• 1:15:55
  them. And then once we see the value,
• 1:15:59
  we just allow our engineers to use those tools.
• 1:16:02
  So we can see the efficiency and the velocity is increasing.
• 1:16:06
  Even though we don't have the real data, the concrete data
• 1:16:10
  which can actually show the real difference.
• 1:16:13
  But we know that, you know, whatever you were doing in five
• 1:16:17
  days, now with these tools you can do in one day or two.
• 1:16:20
  So there is definitely a scope and opportunity to improve and
• 1:16:24
  we are already kind of ahead or with base in the market.
• 1:16:29
  So we have AI engineering innovation to use the right
• 1:16:33
  tools and the right technologies beyond just the incubation team.
• 1:16:39
  And the same is happening in products.
• 1:16:40
  So the product managers in my team are using AI to work faster
• 1:16:44
  in the right areas, right? So the crafted product is still
• 1:16:47
  super important. And I think it's important to do
• 1:16:50
  deep thinking, but we now have effectively like Co pilots to
• 1:16:54
  help with that to speed up the right.
• 1:16:55
  Are they submitting PRS and shipping code yet?
• 1:16:59
  They do. I mean, my yeah, for some
• 1:17:01
  prototype we do. Yeah.
• 1:17:03
  Yeah. Lucky enough to have PMS that
• 1:17:05
  are, are technical, right. So they do a lot of vibe coding
• 1:17:07
  and increasingly it's, it's making things way more
• 1:17:11
  efficient. We're also using evals.
• 1:17:14
  It's so, so important to understand what's actually going
• 1:17:16
  on in your conversations. And the PMS are setting those
• 1:17:19
  frameworks up, right. And something else I'm really
• 1:17:21
  interested in is is scaling usability testing using AI,
• 1:17:26
  right? So I can't remember what it's
• 1:17:28
  what the term is, what the industry term is.
• 1:17:30
  But basically there's some start-ups out there that can
• 1:17:33
  spin up 1000 different personas for you and provide real
• 1:17:37
  feedback on your product. So you don't have to, you know,
• 1:17:39
  get actual humans. Super important to have humans
• 1:17:42
  in the process. But it's a great augmentation,
• 1:17:45
  yeah. Yeah, actually I've seen some
• 1:17:48
  project that will help you do that.
• 1:17:50
  Just to check your current app and interface.
• 1:17:55
  All the links go to the right places, and because otherwise
• 1:17:58
  you're sitting there and you're clicking on the buttons and
• 1:18:01
  you're making sure that everything works like you want
• 1:18:04
  it to work. But why can't you just simulate
• 1:18:08
  that? Yeah, indeed.
• 1:18:09
  But you can also get feedback like behavior type feedback,
• 1:18:12
  like is this a good experience? Where'd you get, where'd you get
• 1:18:15
  stuck? Right?
• 1:18:16
  The things that historically had to sit with the user and spend
• 1:18:20
  the time. You know, you can scale that
• 1:18:23
  using AI. Think that's really, really
• 1:18:24
  interesting. Wow.
• 1:18:37
  Wearable, wearable.