top of page

When Data creates Competitive Advantage, or not?!

In this interview with Prof Julian Wright and A. Prof Andrei Hagiu, experts in Platform Strategy and Network Economics, we talked through the 6 questions to ask to evaluate data-enabled learning moats. It helps to understand if the learnings from your customer usage data are really helping you create a competitive advantage and sustainable growth. We also discussed briefly about data network effects vs regular network effects and platform business model.

We discussed the Havard Business Review article with the same title here.


(00:00) [Music] today I'm really excited to have two experts in platforms strategy and business model network effects and network economics to talk about using data especially data network effects to build competitive edge they co-authored the have a business review article when does data create competitive edge

(00:24) and when it doesn't we will explore the six questions that you can ask yourself to evaluate if your data enabled learning will actually build modes and create sustainable growth so let's start with a brief introduction about what you're working on and what led the two of you to collaborate on this so I can start I'm based in Singapore

(00:48) the national university of Singapore and we've been working on platforms and network effects and businesses that build our network effects for many many years maybe starting around 2000 so you know 21 years and of course these were not so hot when we started or at least when I started working on it but over the last few years you

(01:12) know there's been this massive increase in interest in platforms and one of the interesting aspects of them is their use of data right and data to give them a competitive advantage so Andrei and I have been working together for quite a long time we met at a conference in France and this you know this is one of our most

(01:33) interesting projects that we're still working on which is you know many different aspects of how not just how data gives you competitive advantage but also some of the policy implications surrounding the use of data by incumbents I've been working mostly with Julian I would say we're close to a monogamous relationship

(01:52) research relationship for quite a few years now so I'm based in Boston actually so yeah we met at a conference in france and you know because we had very similar research interests we started working together about 10 years ago on platforms so at this point I think what's really exciting like what drives both of

(02:11) us is I think we do essentially cover the entire value chain of ideas that's why I like to think about it so we do economics research so you know we publish like economics articles we do some consulting for large companies we also work on public policy issues and then more recently we've also started angel investing in

(02:33) startups which are platforms which have elements of data-enabled learning so I mean it's a pretty I guess it's a pretty unique and exciting position to be in in the sense of like you know cover being able to cover all aspects of these ideas I'm more practitioner so I work with startup scale ups and growth

(02:53) companies to help them with strategy and execution and what I'm seeing is more and more ais and data startups claiming how data is giving them the edge and your article really gives me a frame to question if their claims has any legs so let's dig in by setting the context first of what has changed the part that's most

(03:17) significant in terms of changes with data enabled learning is simply the speed and the scope with which we use data enable learning today so in itself learning so using data to learn from customers is not a very new phenomenon I mean companies have been learning from their customers for decades the issue is that that process used to

(03:39) be very slow and actually pretty limited so they would do customers you know they would sell products and they would try to do customer surveys from a pre-limited set of customers and they'd say what do you like about the product what do you not like about it and then eventually the feedback they got from

(03:54) customers the data would make its way into future versions of the products but that happened much later I think once again what's really significant in the past and past decades especially in past few years is that process has sped up significantly especially of course with software products to the point that now these products are

(04:15) learning almost in real time so customers use the software products the software product provider gets the data from customers and they're able to improve the product based on the data they obtained from customers almost instantaneously again using machine learning using artificial intelligence so that

(04:33) again that's why it's become such a huge part of how companies compete now because a lot of these products are digital there's a scope to personalise the products so previously you know whatever you were learning well you're sort of learning for the whole group so your next version of the car

(04:52) would have some improvement and everyone who bought that car will get that same improvement but now with these insights that you can get that are customer specific you can create a version of the product for every different type of customer and so that sort of personalization was not a feature of

(05:09) learning from customers in the past but has become much more important nowadays maybe let's clarify what is data enabled learning and what it encompasses and what it does not because with ai and machine learning companies there is a natural assumption that they will have the data enabled

(05:27) learning or data net effects when we thought when we think of data enabled learning what we have in our mind is actually somewhat specific so first of all it has to be learning from customer data and obviously you can learn a lot of stuff through machine learning which comes from other data sources

(05:46) right and we wouldn't call that data enabled learning because we have something very specific in mind and that is that you learn from your customer data you improve your product and that generates more customers and more data and therefore more learning and so on and so forth so we have this virtuous

(06:04) cycle and so when we talk about data learning we have in mind that virtual cycle something like you'd see on say using google maps you know more people use google maps when they're driving so that gives you better traffic information better predictions that makes a better product more people use it

(06:21) and so on so that when we talk about data numbering that's exactly what we have in mind so it may be with or without machine learning it's possible to think of examples where you don't use machine learning to do that of course machine learning makes that much more powerful the insights you get can be personalised they can be obtained

(06:36) very quickly and also the predictions are more accurate right but it's that's just sort of a matter of degree yeah so just again to make the contrast very clear so like Julian defined data and I think we really need to emphasise this it's a very specific process so it's truly data-enabled learning we have to have in mind this

(06:56) virtual cycle and we would we would contrast this would say in some cases you can buy data so instead of getting it from your customers from serving customers it is possible to just buy data sets right from other sources and of course you can use machine learning and artificial intelligence to learn

(07:13) from the from that data now that is not data enable learning in the sense that we use the term because obviously anyone irrespective of how many customers they have in principle can have access to these external data sets so that doesn't lead to this feedback loop that we are primarily focusing on the reason

(07:32) we're so interested in this feedback loop and the concept of data enabled learning is because that's where the long term sustainable advantage comes from actually the accumulating advantage that a firm would get because if you can just go out and buy the data and you know do your machine learning on it

(07:51) and come up with some improvement then obviously anyone with resources with money can go and do that and you're not getting a long-term advantage from doing that although it still may be something you want to do it's just not going to give you a long term and accumulating advantage I think one of the premise in the

(08:07) article was that there is a difference between the whole virtual cycle of data network effects versus the regular network effects virtual cycle so could you please elaborate more on that sure so I can start on this so the virtual cycle associated with data enable learning like Julian described very simply puts

(08:27) is a firm gets more customers it learns from those customers therefore it's able to improve the product and by improving the product it attracts more customers and so on and so forth right that's how the the self-reinforcing virtual cycle works now that looks very similar to the traditional virtual cycle associated

(08:47) with network effects so think businesses like facebook or airbnb or uber and the network effect there is a bit more direct in the following sense so if more people use facebook well the value to users of facebook increases right because directly because I want to interact with these people and again more people gets on facebook

(09:07) the more the more valuable it is and therefore more people go to facebook and so on so that's another that's a very in sense it's a very similar virtual cycle and the point that we're making in the article is that these two these two virtual cycles the one that's that is due to data enabled learning and

(09:22) the one that's due to network effects they're very similar and in some cases they're combined for example facebook does also have data and some data enable learning but it is very useful to distinguish between them because in some cases if you only say you can have either only data enable learning or only

(09:37) network effects well our one of the points we make in the article is that typically comparative advantage tends to be stronger with traditional network effects relative to data-enabled learning and we can talk about why okay let's get into the six key points or questions that you use to ask if the data enabled learning

(09:58) create modes let's have the first one first one is probably the most important one which is just to sort of ask the question does the data this customer data really add a lot of value compared to the product itself right so obviously if you're talking about a product where the data's the value you get from learning from

(10:20) customer data is quite minor compared to the value of the product then you know you're just not going to get that much advantage from this virtual cycle so we have in mind for instance the example of a smart tv you know a smart tv could of course look at your use of the tv to learn like which types of programs you'd

(10:40) like to watch and make recommendations a bit like netflix does but you know most people when they're buying a smart tv they're not thinking about well is it going to help me you know find which shows I want to watch they're more interested in you know is it does it have high resolution is the colour good

(10:55) is it big enough the right size you know and so on and so forth so you can contrast that with an example like autonomous vehicles right like if autonomous vehicles don't learn well how to you know react to various obstacles on the road and what to do you just don't want to get in that car right like

(11:14) you're not gonna when you're considering which autonomous vehicle you're gonna get your primary concern is how well has it learned from the data and able to predict to different circumstances so there the whole value is pretty much in the quality of the learning from the data and not so much about the

(11:31) actual car design and car features and of course most examples will fall somewhere in between these two extremes that Julian described right so again as angel investors we've actually seen this ourselves you know pretty much every other company these days claims that they have they use ai and machine learning to

(11:50) learn from their customers and it's actually quite amazing how many of those claims you can pretty much refuse just I mean by asking one of the six questions but even by asking this first fundamental question like how much value do you truly add with the data to the standalone value of the product it's important to keep in mind we're

(12:11) talking about the value that's added right because again a lot of people will say well you know with this data we've improved our accuracy you know and how much they've improved the accuracy of their predictions but if you think about the value created by autonomous vehicles you know if they improve the accuracy

(12:29) from 90 to 95 or even 95 to 98 there's still not much value for a consumer because they don't want to get in the vehicle until it gets to 99.99 accuracy right so it's not about like how much incremental accuracy improvement and accuracy you can provide is how much sort of value are you providing to the end user

(12:49) and there's a you know we've seen other analysts look at this and they talk about well there's not much improvement in accuracy it peters out at some level and so on and just to keep in mind when we talk about this question we focus on the value that's actually being contributed would you say that it's about if the data is actually

(13:10) adding value to the value proposition of the product itself then yes that is ultimately that is the main thing that matters so it's not so much I mean again this is something you probably see a lot of companies they say oh look we've improved our accuracy or some other measure by using you know by this much well

(13:32) ultimately I think the right metric should be how much more willingness to pay or how much utility have we added to our customers that's what you care about and like Julian described there are cases in which you know you may see like the accuracy maybe just improved from 99 points nine percent to 99.92

(13:53) you may say that's tiny in terms of accuracy but in terms of value to customers it may be enormous right because we're talking about say saving human lives so you know if the autonomous vehicle is point zero zero two percent more accurate that's actually very significant value yes which is what

(14:12) I think is your initial points on the article which is what everybody is saying data is the new oil right but then it is like they grossly really overestimate the competitive edge that data actually brings that's exactly right let's talk about the next question to ask then the next question we ask is how quickly does the sort of marginal

(14:32) value drop off or depreciate so and this is again where it's important to think about value and not accuracy so we have in mind situations where it could be that you get quite a bit of learning from the data but you get that learning very quickly right like you just need a few hundred customers and you can get most of the

(14:57) insights and at that point you've improved your product and you've got everything out of it and you can't do any further improvement okay and that would be example where the so marginal value learning drops off very quickly and an example we have in mind like that might be something like a smart device like a nest thermostat

(15:13) right where it learns your temperature setting after you've used it a few times and really there's not a lot of additional learning beyond that and that's important because it means this virtual cycle kind of runs out very quickly you know I've learned a little bit I've improved the product but now

(15:30) there's no further gains to be made and that's in contrast to applications like the autonomous vehicle but there are other many others where the learning continues for a long time because actually there are many edge cases you have to deal with and so you really need a lot of customer data and you need

(15:48) that cycle to go on for a long time until you really have exhausted all the learning from the data so you know one example I'd use in my teaching on this topic is credit scoring so you know when you do credit scoring you actually need a lot of data because there are many edge cases and you only find out about these edge cases

(16:09) when the customer goes broke right like there's a default and those defaults don't happen so often they may only happen under extreme circumstances or in different macro environments and so really the learning is a long process and that means if you if you have an advantage and you're getting more data you know that

(16:27) accumulates over a long time it's hard for another firm that comes in that's new to compete with that because they have a very long period of learning ahead of them ultimately we care about value to customers and also what we care about is whether data-enabled learning in the particular context at hand can

(16:42) actually lead to defensible position or competitive advantage for the firm that benefits from that data-enabled learning so if we focus on this idea that's well in order to have competitive advantage like the value the marginal value of data should actually remain high in order for you to actually still benefit from

(17:03) getting more customers a good heuristic to say to figure that out would be exactly what Julian described which is are there a lot of edge cases in that particular application so credit scoring is a good one the other examples that I had in mind was for example search like you can ask like why has been why is google search

(17:20) so dominant I mean why have they been able to sustain their advantage for like for decades now why is it so hard to compete well it's a lot of it has to do with it's it basically there's like billions of searches every day and there's lots of long tail esoteric searches and then google is the only one

(17:37) I think they're the best that basically figure that you know they have enough data to actually cover all these edge cases I mean microsoft has spent billions of dollars with bing and I'm sure bing is actually okay on on very common searches but I think the difference between the two is going to be at these long tail edge cases where

(17:55) again google benefits from a lot more data well android just described this illustrates the discussion we're having before about the difference between value and accuracy because you know google search and bing search in terms of accuracy they may not they may differ only you know by a few percentage points right in some sense

(18:16) for all these normal searches they they're both quite reasonable in terms of giving results but all these edge cases which may only account for a small you know a few percent in accuracy they can actually derive or they can make a big difference in terms of value right for a user because what they really care about is

(18:33) you know that they get some results to these more difficult questions when they're doing their search so that and similarly with the credit scoring or the autonomous vehicle again like that difference in accuracy you know it may be that you reach that point where you're not getting much gain and accuracy but you're still getting

(18:51) a lot of gain marginal gain and value right so when we ask this question we're focused on does the marginal value of learning continue you know as you get more data not whether the accuracy goes up by much because that may actually getting the asymptotic right so what he's trying to say is that the additional

(19:10) usage data of your data that comes from the usage itself from customer does not lead to further insights or improvement that's right so I mean the bottom line is if you only need a small number of customers to basically get most of the learning that you need to improve the product then that doesn't lead to a very sustainable competitive

(19:30) advantage because it means another competitor can come in obtain you know the very small required number of customers to get you know to get most of the learning so versus a situation in which you need a lot of customers a lot of past sales in order to learn well someone that has a head start actually will continue to in principle

(19:48) ignorance they don't screw up in principle their advantage will keep accumulating and it'll be very hard for competitors to catch up this of course is good news if you're the incumbent right and you're an industry where you have this you know situation where marginal value does not drop for a long time

(20:06) right but it's bad news if you're an entrant right you come into a market where there is an incumbent and you face the situation where you know it's autonomous vehicles or something like that and everyone else is way ahead and it's going to take you you know millions and millions of customers to catch up

(20:20) in terms of getting the data inside so what you're saying is that in this scenario actually having a first mover advantage is important for them to have basically the head start and what you say about the edge cases is actually quite interesting because I think as you put it what it does it actually allows them to gather the broad

(20:37) and the breath and cover the huge volume of possibilities and scenarios and capturing the long tail as you said you know that well that majority would not be able to capture because they will only be catering to the common 80 and not the 20 that is just as important and valuable yeah that's exactly right so the third

(20:56) one what will be the third question to ask them the third question is how fast does the relevance of the data that you get from customers depreciates actually we can use the same example that we just used I think of google search I mean part of the reason that like another way to think about the the reason that google search has been

(21:14) dominant is the fact that searches that were done on google say 10 or 15 or 20 years ago the information the data that google gets from those searches is actually most of it is still relevant today now I'm sure you know there's if you look at the data I'm sure there's certain types of searches there are new that come

(21:33) there's new types of searches that you know come on come all the time so we can make a lot of jokes about like what what are the common searches today versus 10 years ago and things like that but overall it accumulates right so there's still relevance all the data that google acquired even like 15 years ago is still

(21:48) relevant today so that's very good for competitive advantage if the value from the data does not depreciate fast that's good like it's good for the incumbent for someone that has a head start on the other hand if the relevance of the data depreciates very quickly then data enabled learning cannot

(22:06) sustain a very defensible positions I think an example we had in mind I can't remember what we talked about in the article I think it was social gaming so this was like an example of zynga and others I mean you know if with social gaming for example you can have a hit so it's quite difficult I mean it's very

(22:24) difficult to have a hit social game but the point is once you have that hit you're basically trying to replicate it it's actually very difficult because consumer fads change very quickly so whatever was relevant whatever you learned say five years ago and you know you had massive success with one social game

(22:42) may become completely irrelevant you know just a few years later because consumer preferences have shifted so you have you basically started from scratch another example would be google maps there's an example where the data depreciates very very quickly it's an interesting example because obviously google maps does have a strong

(23:02) competitive advantage and we can bring up the reasons for that later and with respect to another question but at least in terms of this dimension in terms of data depreciation you know the data that you collect on traffic obviously depreciates very quickly and is not so relevant the next day of course

(23:21) it is still useful for predicting on average the traffic conditions at a certain hour but if I want to know the current conditions on a particular road right knowing what people knew yesterday is not that relevant so the data's always depreciating and of course google is also google maps is also continuously

(23:40) collecting new data so it's sort of able to learn continuously but it would mean that if someone else came along and had just as good traffic data right like they can they can enter this market they don't need to they don't need past traffic data they just need current traffic data and they can compete

(24:00) of course it's difficult to get that data since they both sounds a bit similar as any additional usage does not lead to any further insights or improvement why do you keep them as two separate questions I was just going to point out one obvious difference and then maybe Andrei can can add on to that which is the

(24:20) previous one is about sort of more about future data right like as I collect more and more data what happens to my learning and this question more about past data do I get value from my past data that I've collected do I continue to get value from or does it just that value disappear so in the google

(24:39) maps example that value just disappears tomorrow pretty much right whereas in google search it lasts for a long time and in fact we can use you can use the google maps example so it does in some sense it's very it does very poorly on this last point because the data depreciates very quickly but I would

(24:58) argue it does very well on the previous point in which the additional I mean you know it actually they do need quite a lot of users to figure out you know traffic conditions so they are similar but actually they can be like in you know in some cases you may have one but not the other they're definitely not

(25:13) they're definitely not synonymous so the fourth one is asking how easy or how difficult is it for competitors to copy the product improvements that that are the results of data-enabled learning so obviously if let's say I learn from my customers data and then I translate that learning into

(25:32) some new features that I put into my product if those features are easily observed by my customers and easily copied by my competitors and easily copied by my competitors then I haven't obtained much advantage I mean the data is very valuable but in terms of competitive advantage I'm I haven't obtained anything right

(25:52) because my competitors can just sit back not even worry about learning from their customer data just look at what I'm doing and then just copy the features that I have versus there are other situations in which whatever you learn from the data it translates into improvements that are not easily observable from the outside

(26:11) in particular by competitors in that case obviously it's you know it's much better situation to be in so I think the example we gave for this latter better case would be something like suppose you provide software like you're so you're learning from customers how to operate your call centres

(26:29) now this is very like backhand process type of learning which obviously will translate into a better experience for your customers but it's impossible there's none of this stuff is observable from the outside it's impossible for a competitor to sit there and say like let me figure out how they've improved

(26:46) their their customer experience that's very different from say some software products which certainly learn a lot from customers but then it translates into features like which are everyone can observe and then those features are easily copied so they're like the data enable learning doesn't

(27:02) doesn't provide much advantage you could think of it I guess whether you can reverse engineer from the from the observable you know products or features and in the call centre you can't really work out what they did to create this better user experience you know better more friendly and more helpful call centre staff

(27:20) whereas the other example of a software we can just look at the nice features that it has and just copy them you know obviously you don't need to do any reverse engineering so so in fact actually just occurred to me there's probably like a nice way to sort of summarise this and have some like

(27:35) let's say a clear guideline for companies to think about so if the data enable learning translates into improvements that are either some sort of back end things or processes that's good that's very difficult to copy or to reverse engineer on the other hand if the data enable

(27:55) learning translates into improvements that are front ends i.e features into the product that's not as good because obviously that's easier to copy all other things equal I prefer that the data enable learning is something that actually improves the backend processes rather than the front-end

(28:12) you know user features yeah or basically what you're saying is that the improvement is maybe on the differentiated activities or the secret sauce that lead to the attributes that make a difference to the customer experience in terms of the value proposition okay so question here is how difficult is it to get some other

(28:32) data that you can use to get the same sort of insights and some in some cases that's going to be there's going to be some other data set that's going to allow you to get the same kind of learning that you got from your own customer data and in that case you know one company's painstakingly you know got all

(28:50) made gathered all this customer data got the insights built up improved product and if someone else can just go out and use an alternative data source that pretty much does the same thing and buy it you know then you're not really getting much competitive advantage as opposed to an example where your data

(29:06) is very unique that your customer data is very unique and there's no good alternative to that data you can't go and buy something that is going to give you similar training data that you can use for your machine learning to get of similar insights so you know an example where I guess where you can get a lot of the

(29:26) data sort of there's various versions that are available would be something like captions and subtitles you know like imagine you're trying to put on subtitles on your social media or your videos or whatever right nowadays there's just so many ways you can get sources of data that allow you to

(29:48) train models to create subtitles or caption voice and that's because there's just lots of publicly available data sources on both the spoken version and the you know the subtitles or captions for those I can just go on youtube I can go on netflix I can go on all these different services and I can just

(30:10) see the spoken version and then the subtitles for those and I can use that to train my models right so that that's just readily available okay let's go to number six because this is the one that I really got the best ins the most insight for right so the question to ask for number six is well let's hope that the way the one

(30:29) that we have in mind is the one that you want so for number six the one that we're thinking of is is the learning from user data within user or across users is that what you had in mind Josephine awesome broadly speaking the so the learn that you can get from customer data there's two extreme cases in one case

(30:52) the more Josephine uses my products the better the product becomes but only for Josephine so like I'm able to just figure out your preferences and give you exact the service of the product that can customise it the way you want it and but it doesn't help with anyone else like it doesn't help me improve the

(31:12) product or the service to anyone else so that's what we call within user learning so the data from one user only helps to improve the product or the service with that particular user the polar opposite situation is in which everything I learned from a customer is actually relevant to everyone else

(31:28) so here is I learned from Josephine I learned from Julian and I learned from a bunch of other people and then I can improve the service or the product for everyone so that's what we call across user learning and of course in many cases you can think about there's probably a combination of these two types of

(31:45) learning there's some within user learning and some across user learning but if you focus on these pure like these two pure cases you can ask the question which one's better is it better or other things equal to benefit from from data that helps mostly to you know to target within user learning so to try to improve the product for a

(32:03) specific user or does it mostly help with across user learning and they work actually both can provide some competitive advantage but they work quite differently so in the situation in which the learning is mostly within user it creates switching costs so if whatever product or service I'm selling really gets a lot better the more

(32:24) Josephine uses it then she's going to be very unlikely to switch to a new product just because my product has figured out her preferences gives the exact gives her exactly what she wants however it doesn't really help so it's good so that's kind of within user learning is very good to maintain to to keep

(32:41) existing customers you're not going to switch but it doesn't really help in attracting new customers because if Julian's considering my product versus someone else well I don't really have an advantage right the only if the only advantage I have is when Julian uses my product a lot well you know if he's like if he hasn't used

(32:57) my product then there's like zero value to him versus the case in which there's a cross user learning so when there's a cross user learning and the product truly learns across different users then a new customer actually does benefit from learnings from the existing customers so with the cross user learning

(33:16) there is a lot more competitive advantage when it comes to attracting new customers so I would think it basically you know there's a trade I mean there's a trade-off so if you're trying you know if you really if you're trying to have very high switching costs it means to keep your existing customers within user learning

(33:31) is great however if you're in a situation where it really matters is try to grow very fast and attract new customers across user learning is a lot better because that gives you a lot more you know it makes it a lot easier for you to attract new customers so can you talk about the example of spotify and pandora because I think that

(33:48) really connect the thoughts much better for a lot of people well thank you very much for bringing to my favourite topic because I think I'm probably only the remain the only remaining pandora user in the entire world so it's exactly what we're talking about so pandora it's quite interesting pandora I think as far

(34:06) as I know it's still only available in the us I mean I guess you can get it vpn in Australia and Singapore but it's pain so pandora has been great so they invented the music genome project which is basically trying to categorise music and figuring out like different interesting relationships between different songs

(34:22) and types of music and they develop this amazing recommender system which figures out the more you listen to their music and the more you say like I like this but I don't like this they're just very very good at figuring out your preferences and figuring out other things that you might like so you start listening to pandora radio

(34:38) they customise it to a degree I find that amazing I love it I absolutely love that product and I've been using it for years I'm unlikely to switch and in fact I have resisted to this day switching to spotify precisely because there's so much within user learning from pandora spotify is quite different I mean I

(34:55) think they have some of that but they're mostly they mostly have traditional network effects and across user learning so for example they they allow people to share playlists so they've been much better at growing internationally especially they're much larger than pandora now so I think it's a very you know it's very

(35:11) it's a very good illustration of the difference between the two so pandora has a you know an established lo a very loyal user base I guess I'm I'm one of those but spotify has grown a lot faster so I mean this is an example I guess across user learning is obviously is obviously better when we think about

(35:28) within user learning we have in mind often the smart devices right like we mentioned before the nest thermostat I mean that's just basically learning your preferences for temperature and that's pretty much mostly within user learning my favourite example is the smart bed by eight sleep which is it learns your

(35:48) preferences for temperatures throughout the night to give you the best possible sleep right and my preferences for temperatures throughout the night could be very different from yours and so it's you know mostly within user learning there's probably some across user learning but it's mostly within and you can contrast that to our example

(36:05) previously of autonomous vehicles right you would hope that the autonomous vehicle is learning when it comes to predicting you know determining how to react to an object that comes out in front of the car is not personalised who's driving the car right you want that to work the same for any user who

(36:22) gets into the car so that's pretty much across user learning and then I have an example of a tractor and a autonomous tractor an autonomous tractor is interesting because autonomous tractor combines both of these things right because a tractor works on a farm and the route that it's going to take is

(36:40) going to be the same you know every day pretty much so there's within user learning it learns the route of the farm and all the obstacles on the farm very well but it's also got a lot of across user learning because it has to work when you know maybe the farmer takes it to a different farm or to a different certain

(36:54) you know different situation or there's something in the environment that changes and those things it's learning from you know all the other autonomous tractors out there so it's you know there are when you think about each example you can think about how much within user learning there is how much across user learning

(37:08) and how much combination of these things together well if you can it's great if you have both like we I mean we prefer as angel investors I think we prefer companies that have both because obviously that you get the benefits of both types of competitive advantage we've alluded to the fact that you have

(37:23) this virtuous cycle right which is great and that you could think of as like a data network effect you know the more people use the product the better it gets and therefore you know the more people want to use it okay it looks like a network effect but there's some subtle differences okay and when we think about comparing

(37:44) regular network effects and data network effects we think data network effects are generally not as powerful for a few reasons okay and one is with network with regular network effects often there's sort of a coordination problem which is people have to think about well what are other people going to do are they going to use this new

(38:06) type of product so you know do I want to cut do I want to switch to this much better new social network that doesn't have all these privacy concerns okay Andrei has one in mind that he's looking at as opposed to using facebook right but would I switch well that depends on whether everyone else would switch

(38:24) because there's not much use joining a new social network where no one else is there right so there's that coordination problem that coordination problem can also exist in with data network effects okay but it often doesn't so an example where I think it does exist and which explains why google maps has

(38:43) a strong competitive advantage if you think about google maps why is it each morning that I go to work I use google maps well because I expect that everyone else is using google maps and therefore the traffic predictions will be accurate right so it's that it's that coordination of everyone using the same

(39:03) service if a new mapping service comes out that actually has superior features no one's going to use it unless they think other people are going to use it therein lies the coordination problem right and that's very similar to the same problem with that with regular network effects and that's a very powerful advantage to

(39:20) have right but it's just that when we think about many examples of data network effects they don't have that very strong coordination problem partly because users are not very they're not sort of consciously thinking about well is this product going to get better because as more other people use it they're not thinking like that they're

(39:38) just thinking is this product better today than some other product right they're not making that sort of forward-looking decision when they come to use the product so that's one difference the other of course difference is if you just have within user learning and you don't have across user learning

(39:53) then there is no network effect no data network effect right the only way you get data network effect is if the product gets better as more other people use it but if it's just within user learning it's just personalised to each person it's not getting better for you as more other people use it right so my nest thermostat doesn't

(40:12) get more accurate for me just because Andrei and Josephine are using it right so that's an example where you know you have data enabled learning but you don't have a network effect so basically what you're saying is that on the last point about within and across user learning if the learning is actually

(40:33) across user that is when the data enabled learning will have the regular network effects is that right yeah that's right like you really need across user learning to have any hope of having data network effect and ideally you have both that's even better but if you just have within user learning

(40:56) you're not going to get anything like a network effect so I would add so just again to drive this point to drive some of these points home like perhaps even more forcefully I mean one way to think about it is the difference between regular network effects and data network effects and regular network effects

(41:10) this particular thing about the social network example I really care directly immediately who else is on the social network like I that's that's the only thing that matters right like Julian described with data enabled learning I actually don't really care per se who else is using the product the only

(41:26) reason I care about it I may care about it is because while other people make the product better for me but again like Julian explain well you know most people probably don't understand this most people don't really think about like well I'm going to use google search because it has more users and therefore it

(41:41) learns faster they just probably use it because well it's just really good and presumably it's better than bing for their obscure questions and another way to think about I would say another way to think about it is linking to one of the points we made earlier with data enable learning often times not always but oftentimes

(42:00) you can kind of like find ways to get data to help you overcome the initial the initial shortage of you know of data from customers so like if you can find a data set that helps you overcome the fact that you don't have a lot of customers that actually helps so you can buy basically you can buy your way to some

(42:20) learning by buying data sets and you don't have a lot of customers with regular network effects you just can't do that I mean I guess you can buy customers but there's basically no way around it like you can't just go and buy data sets you just need those customers so again if I have a social network there's no way I can go and buy

(42:35) some data I just need to attract other customers so that's in this sense is what Julian explains I think it's very important is the part that there's a coordination problem and that coordination problem can create much stronger competitive advantage with regular network effects than with data network effect again you can have data

(42:54) network effects or you can have traditional network effects traditional network effects what we mean is the ones that we talked about for facebook I think one of the summary of the article is that the companies that has both data network effects and regular network effects tend to have the strongest competitive edge

(43:11) yes of course regular or traditional network effects in most cases tend to provide stronger competitive advantage than data network effects but you know in the words of tony stark I think it was in iron man one it's like is it too much to ask for both like you know if I could like why do I have to choose between the

(43:30) two right so ideally if I'm an investor I would prefer to invest in a company that is both that has both traditional network effects and data network effects so say you know and in fact the very big tech companies the ones that are very successful they do have both so think about facebook think about airbnb so they have regular

(43:48) network effects in the sense you know the more hosts they have the better is for travellers and vice versa and they have better network effects because obviously the more you know the more customers they serve the better the product gets for customers they can make better recommendations they they can learn how to how to

(44:04) adjust the prices and so on so you know again if I can if I can combine the two that actually has a chance of providing very strong competitive events is this like the network effect as well as data enabled learning something there is a key characteristic that belongs that only a platform business model actually

(44:20) enjoy to be very to do it very briefly the way and there's lots of companies lots of people have very vague definitions or notions of platforms the way we think about it platform business is a business that connects customers it could be multiple types of customers so almost by definition platform businesses

(44:44) have traditional network effects so something like uber is a platform business it connects drivers to passengers airbnb is a platform facebook is a platform so the only I mean I guess the requirement there is that there's you know I'm creating connections so I'm enabling interactions between these customers

(45:05) so network effects traditional network effects are the quintessential characteristic of platform businesses data network effects are not required but they can again if they're present they make things even better okay so what you're saying is that for the platform model itself you know the network effects are usually

(45:22) part of the characteristic of why it makes platform you know business models so interesting and actually difficult to build right and therefore it is the secret results that give platform model the competitive edge that give them the sustainable growth but even after they have successfully built it it's really different

(45:42) if it's really difficult to unseat them because they don't even have to do anything much because the never effect itself you know allow the edge to continue and to grow and to retain the customer and to retain its advantage that's exactly right so as one I think it was the founder of intuits I had a

(46:04) conversation with him the way he likes to talk about this is basically to say the reason I like network effects so platform businesses with network effects is because they basically get better when I sleep and I think that pretty much summarises it right so unlike a traditional product business if you want to get more customers or if

(46:22) you're on the product to get better there's you actually have to do the work right I mean you have to improve put you put on new features with platforms and network effects I mean by the way no one says you shouldn't actually try to have new features of course you should do that but even if you stopped right even if tomorrow i

(46:39) stop improving the platform it can still get better just by virtue of there's customers participating and because more customers participate that makes the product better for all the other customers precisely because they care about the presence of everyone else I think one of the inside that was that was there

(46:55) in your article that platform model really gets a few profit margins by doing very little which is where it's which is something that's so attractive so they grow very fast and they have relative very low cost structures precisely because I'm in the business so if I'm a platform business I'm just enabling interaction so think

(47:13) about something like airbnb compare airbnb to hotels airbnb is just pure software right to me it's amazing like they don't own any they don't have to own houses or hotels or anything they're just allowing people that have houses or whatever like apartments to rent them out to others versus a hotel chain which

(47:32) actually has to own inventory same thing compare uber with taxi companies right uber is just pure software they don't have to own cars okay so is this the reason why you're so fascinated with platform yeah I think like Julian I came to the world of platforms by recognizing pretty early on that platform businesses

(47:50) raise some very interesting issues so you know there's some very specific characteristics that are associated with platforms and yes they do create this kind I mean they tend to get very large they're very important and you know we can see I mean there's something intrinsically interesting about it

(48:05) but I would also say you know I've also noticed that they tend to be very how to say when they succeed they tend to be financially very successful so obviously as an angel investor I would very much like to be able to identify the next big platform businesses so yeah I think if you just say like I'm just

(48:23) going to invest in platforms that's probably a decent you know it's a decent strategy and you have you have a chance at outsized returns again given how powerful platform business models can be so well I have one client who actually is just listed last year on a stock exchange here and they are a platform

(48:41) business and one of the fastest growing startups as well as one of the bigger platform marketplace or platform business actually in Australia I have a startup who's currently in the process of building a platform business as well however you know as you know and it's written in the article as well

(48:56) it's not easy to build one right if you manage to be successful with it you'll be very successful but it's a lot harder I mean I think of course this is like you know they're in life they're trade-offs right I mean yes if successful it's much more successful than the average product business like there's no

(49:12) question but obviously it's a lot more difficult to build than traditional product businesses precisely because you have network effects which like you have to get past the initial chicken and egg problem it's very and there's lots of by the way there's lots of other things that can go wrong but yeah I mean it's it's actually very

(49:27) very difficult so what is the best way for readers and an audience to each either get your content reach you or actually find your content so I think the best way for both Julian and I is to go see our sub stack newsletter so can can subscribe to our newsletter we put

(49:48) out an article every two weeks hopefully we'll have a further discussion more in detail with regards to moving from a product to a platform business as well as you know what defines a platform what makes a platform you know how to make a platform and build one actually that is successful so yes we do cover in great depth what

(50:06) you just mentioned so how to turn traditional products or services into platforms we'll talk about how to you know the main challenges when you're starting platform businesses and so on the goal here was we're basically trying to cover every single interesting aspect about the business of platforms


bottom of page