Multichannel Success Podcast Season 3 Episode 5 - Transcript
Why you will die without AI
David Worby [00:00:09 - 00:00:46]
This week, we're talking about AI, both in the context of retail and maybe a little bit about the wider world of AI. I'm delighted this week to welcome Alex Dean from Snowplough. Great to be here. Thanks. And Mark Pinkerton from Prospero. Hello. I think we should start, Alex, because I know that AI is a significant plank of what your business does. We should start with a bit of an intro in terms of what the company does, how it does it and the kind of clients it engages with to give our listeners a sense of where you are positioning your business. Sure thing.
Alex Dean [00:00:46 - 00:01:18]
Thanks, David. So in a nutshell, I'm the co-founder and CEO of a company called Snowplow. And we've been around since 2012. And we do one specific thing, which is we help brands to generate very rich customer behavioral signal of their digital estate, which means their mobile apps, their websites, their digital storefronts, and so on. And yeah, retail is a big and growing industry for us, especially since COVID and the rise of the digital storefront is kind of the premier storefront for many businesses.
David Worby [00:01:18 - 00:01:37]
Fabulous. Now I know off-air we talked a little bit about how retail is not the biggest part of your of your client portfolio so it'd be great to learn some of the lessons and experiences you've got from places outside of retail that we can bring in to our listeners in terms of things that they should be thinking about.
Alex Dean [00:01:35 - 00:02:08]
Yeah, so retail is about 30% of our business and growing and at the heart of Snowplow really is adoption by data teams. So there are these modern data teams that have data engineers, SQL folk, data scientists in them and when organizations start to build that data team they find Snowplow and they start to use our data and that's been a really growing trend in retail since really since COVID and the rise of the digital storefront and more and more retailers realizing they need to invest in data platforms and data teams to really get a handle on this and this is accelerating now with AI.
David Worby [00:02:06 - 00:02:21]
And this is a Certainly a subject we're hearing more about. I know, Mark, you and I have spoken to a number of retailers in the past weeks and months. And AI seems to be on the radar, but it seems to be a kind of patchy story. How do you position it?
Mark Pinkerton [00:02:21 - 00:02:34]
From the retail perspective, I think there's still a lot of confusion about AI and the best use of AI that retailers can make of it, and we're seeing a very mixed bag in terms of the conversations that we're having with people.
David Worby [00:02:34 - 00:02:36]
They're making baby steps.
Mark Pinkerton [00:02:34 - 00:02:56]
They're making baby steps, and we'll cover those in terms of the rest of the conversation, but I still don't think very many have got to the level of... really engage with AI. And I think part of what we want to cover today is to say to people that, you know, without AI, you're going to be left behind. So you've actually got to
David Worby [00:02:53 - 00:02:53]
without A.I. think about it, but actually, practically, what can they do? And we'll cover some of
David Worby [00:03:01 - 00:03:26]
that later. Great. I think also it would be good, just before we get into it, just to kind of do our usual explainer of what AI is, because I think there'll be people listening who are thinking, where do I draw the line between kind of what we've had for a long time, which is kind of algorithmic driven things, and AI. So Alex, in your world, how do you help people define what AI is, and maybe as importantly, kind of what it isn't?
Alex Dean [00:03:26 - 00:03:33]
Right, yeah. So I asked ChatGBT to define AI and I didn't quite like the definition, so I'll try one.
David Worby [00:03:32 - 00:03:34]
So, I'll try one.
Alex Dean [00:03:33 - 00:04:55]
I'll try one myself. So I think you're spot on, David. So organisations, including retailers, are very used to, you know, a long history of algorithmic approaches, rule-based systems, whether those are encoded into ERPs or they're written in SQL or whatever, they're fundamentally built by programmers and packages. In the AI world, it's very different. So in the AI world, we're using large data sets, including data sets from our own organisations, to train models. So no one has to write any code, but we have models that are then trained on our data, and we can then give them a new scenario, a new piece of data, and ask them to make a decision based off that. So if you will, there are still those rules, but they're buried inside the model, and we don't have to formally write them. We just give the AI data, we train a model, and then we can use it to do all sorts of things. If I wanted to break down AI a bit further, a lot of the work in the last few years, a lot of the work that we've hired data scientists to do, has been predictive modelling, a lot of things like propensity scoring, for example, in retail. The new wave of GenAI and ChatGBT and others is this kind of wider concept of large language models and generative AI, and those LLMs, as they're called, have been trained pretty much on the open internet. So it's just a different mode of AI.
David Worby [00:04:55 - 00:05:17]
Okay, so you know we could probably fill hours, if not days, talking about some of the elements you've just unpacked there, and we clearly can't do that in our 20-25 minutes, but the very fact of training AIs on data, just unpack that a little bit so that our listeners get a sense of what you mean when you say that. Right.
Alex Dean [00:05:15 - 00:06:10]
Right, so let's take two very different categories. So if we data science workloads that have been ongoing in business, including a lot of retail and manufacturing for the last few years, that's where we take very rich data sets. So for example, it could be customer 360 data, so customer behavioral data, it could be supply chain, it could be manufacturing data, it could be anything like that. We take all that observed history. So all of that rich data that we've captured in our systems in our data platforms for years. And we give that to data scientists to build models with and to train machine learning models. Once we built those in a kind of offline way in the lab, so to speak, we then operationalize them, put them into production and use those to make decisions and come to verdicts on new live, it's called online data that they receive. Okay.
David Worby [00:06:10 - 00:06:50]
So, thank you for that. I guess, Mark, in our experience, we've worked with clients who have given us large data sets, and we've used freely available tools to determine insights from that data that have often been quite revelationary for those clients. Like, we had no idea that this was going on, is a quote we hear quite often. In your world, how does that differ from the world that Alex has just painted of asking an AI or a capability to kind of predict or look at patterns that we could never, in our kind of world, predict.
Mark Pinkerton [00:06:50 - 00:07:33]
I think that's the main thing. Even if one is a relatively experienced analyst of these sorts of things, and I think we would probably fall into that category. Well, I hope we would. We're only going to do so much, whereas by applying an AI, you can consider all the possibilities, not quite infinite, but a very large number of possibilities. And you can see trends and patterns that a human will create a hypothesis on, but you will only probably test a very small subset of those. Whereas with AI, you can pretty much test all of the hypotheses that you want. So you'll find the lesser trends and the sub- patterns. You'll probably notice the big
David Worby [00:07:26 - 00:07:27]
Okay.
Mark Pinkerton [00:07:33 - 00:07:35]
trends with a good analyst.
Alex Dean [00:07:35 - 00:07:52]
Right, so you can try many more scenarios with AI and you can also feed it much more data than human analysts would be able to deal with, a much richer signal, whether it's much more granular activity from your supply chain or much richer signal from your customers.
Mark Pinkerton [00:07:52 - 00:08:01]
And this effectively can do the regression analysis in the background and say yes, you know, weather did have an impact on your sales last week or not, whereas you go into your
David Worby [00:07:57 - 00:08:01]
Whether You go.
Mark Pinkerton [00:08:01 - 00:08:06]
Monday morning trading meeting, everybody will always blame the weather over the weekend for poor results.
Alex Dean [00:08:06 - 00:08:07]
Yeah, and this is what we saw at Snowplow.
Alex Dean [00:08:07 - 00:08:40]
So at Snowplow, we've got the advantage of having been around a long time since 2012. And for years, we were asking our more advanced customers, Hey, are you doing AI yet? And for years, they said, No, no, we're just focused on web analytics and business intelligence. And about three years ago, it started to change. And they looked at us like we were stupid. And we're like, of course, we are. The behavioral, you know, signal that we're capturing off the websites, storefronts, whatever, it's highly, highly predictive. And so when we feed that into models, we get really interesting outputs. So I would say the last few years, it's really changed.
Mark Pinkerton [00:08:38 - 00:08:44]
So. It's gone from being retro to being forward-looking.
David Worby [00:08:44 - 00:09:31]
So I think later in our chat, maybe if we've got time, we'll come to how reasonable it is for retailers to actually be able to implement change in that wide spectrum of things, because I suspect that in recent past, the insights that came from the top level, Mark, you described, they could act on those, they could do something about those. But if now you're multiplying that by 20, there may be a challenge to the operational execution of how you implement that. But let's come to that in a minute. Before we do, Alex, in the wider sense, our audience is retails and brands, but I know that there are other sectors out there and maybe other sectors who are slightly more ahead. Just give us a sense of what you're seeing from the adoption of AI in other sectors of the market relative to retail.
Alex Dean [00:09:31 - 00:10:22]
Yeah, so we have a lot of customers in media and they've been early adopters of rich capture of their digital estate. So they've always been tracking a lot of signal off the digital estate and they've been using that to power interesting use cases like recommendations, which can be quite AI infused. But I think the most interesting segment of our customer base are what we call the digital natives. And so these are businesses, they might be transactional, they might be in retail, they might be marketplaces, they might be travel search, those kinds of businesses. But they're businesses where an awful lot of people in the company have engineer in their job title. And those are the businesses that have been able to kind of get stuck into AI earlier. They've had data science teams, they've had very kind of fast loops from the front end teams into the data teams into the data science and beyond. And that's been that's been a big step. That's been a big.
David Worby [00:10:22 - 00:10:40]
So it might sound like a silly question but what do you think the kind of fingerprint, the DNA of those businesses is? Is it the fact they've got engineers or is it that they're early adopters to anything that's kind of on the forward momentum trend? What is it about them that means that they've become who they are?
Alex Dean [00:10:38 - 00:10:52]
Yeah, good question. I think willingness to take risk and pretty fast learning loops and a strong data culture or an emerging data culture, all of those things have helped them get on the AI bandwagon.
Mark Pinkerton [00:10:50 - 00:10:55]
And I would say that a lot of them are set up as product teams as well.
Alex Dean [00:10:54 - 00:10:55]
Right, great. As well.
Mark Pinkerton [00:10:55 - 00:11:00]
So not like a traditional retailer, but you look at, you know...
Alex Dean [00:11:00 - 00:11:02]
Yeah, the website is the product.
Mark Pinkerton [00:11:00 - 00:11:20]
The website is the product. Yeah, the website is the product or the service is the product. So the just eats of this world and people who are very much progressing on the basis of digital natives. They are set up as product teams, and they think in that way, and those product teams are driven by data. Oh, great.
Alex Dean [00:11:19 - 00:11:23]
A great example of a Snowplough user, Just Eat. I didn't know that.
David Worby [00:11:22 - 00:11:28]
So there you go, a virtuous circle. So I think what that means then for our So there you go.
David Worby [00:11:28 - 00:11:49]
There you go. Happy Christmas. listeners is there isn't a sector you should go to to kind of learn faster than anybody else but there are some characteristics of businesses utilising this. So maybe some of our listeners who want to think about this might need to reach out and ask who the early adopters are with a kind of entrepreneurial risk-taking kind of fingerprint but we can come back to that later.
Alex Dean [00:11:46 - 00:12:19]
One thing I'd add on that David is because of this explosion of almost prosumer or consumer gen AI with chatGBT even in an organization that's less data mature or has fewer engineers there's already quite a lot of interesting sort of shadow IT style experimentation with things like chatGBT in your departments and so you might think well we're not adopting AI strategically but you already are in multiple departments using it day to day. I like that shadow IT idea.
David Worby [00:12:13 - 00:12:25]
I like that shadow IT idea, it sounds like it's contradictory to the strategic thing but tactically we can't not be doing that.
Alex Dean [00:12:24 - 00:12:30]
People are learning how to use these tools in-house.
David Worby [00:12:29 - 00:12:34]
I think if the IT department found out about that, they might have a different opinion and view.
Mark Pinkerton [00:12:34 - 00:12:37]
That's because they may want to control things.
David Worby [00:12:34 - 00:13:25]
That's because they may want to control things. Yes, let's gloss over that. And I'm not sure you can. So moving back to our principal audience today, retail specific use cases. Let's just unpack maybe some of the likely use cases that we kind of see. I think one of the things we're hearing is an uncertainty about whether services retailers are buying that are packaged as AI, or products they're buying that are packaged as AI, may or may not be. I seem to remember three or four years ago when organic became the thing. Everyone slapped organic on everything. And it's now widely discredited. I wonder what's going to happen in that sense for the packaging and promotion of things that are only vaguely AI, but are seen to be able to be packaged as AI. Do you see that?
Alex Dean [00:13:25 - 00:13:34]
Yeah, so we see a lot of that and we sort of call it internally the kind of now comes with AI stick around on the package box. And joking aside,
David Worby [00:13:32 - 00:13:34]
On the package box.
Alex Dean [00:13:34 - 00:14:07]
there are some really good applications inside of these environments using AI. So you've got interest, you know, people like Zendesk are exploring how to do AI agents inside the support environment. You've got like Magento Sensei, Salesforce Einstein. So these are powerful tools. The challenge with them is that they're very bounded into the kind of packaged environment that you purchased. So, you know, they don't have the kind of breadth and depth of understanding of your like enterprise data that you can build, you know, in a kind of more DIY approach.
David Worby [00:14:08 - 00:14:37]
Yeah, so I guess probably some of our listeners who might have a kind of monolithic stack of Magento or Salesforce Commerce, you mentioned those two, or BigCommerce, or many, many others, there are likely to be AI components embedded within them now, I think is the message. So the question would be, which bits do I kind of go discover that might help me do a better job? I guess, what, do you go to your SI for that, or do you go to the vendor for that? How do you determine?
Mark Pinkerton [00:14:37 - 00:14:57]
I think it's a bit of both. You have to get the best practice from Salesforce, if we're talking about Einstein, and the SI will have to configure some of that for you, but also it's about training your teams to make sure they can make the best use of that capability as well. Because it's got to ingest the right data before it can make any decisions about
David Worby [00:14:52 - 00:14:54]
Yeah, I mean, well, it's got to.
Mark Pinkerton [00:14:57 - 00:15:02]
things, and I'm not convinced that all of the implementations are set up to do that.
Alex Dean [00:15:02 - 00:15:28]
No, exactly. And I think the kind of countervailing trend in all this is that these organizations are also starting to build their own data science teams and they're also starting to build their own data platforms. And you have to bring the data to those places and you have to do your decisioning and your AI really in that platform. And so what you can do in those packaged environments is always going to be limited compared to what you can do in your central data platform.
David Worby [00:15:27 - 00:15:33]
Yeah, however I still suspect that there are a lot of listeners out there who kind of passionately
Alex Dean [00:15:28 - 00:15:29]
Yeah.
David Worby [00:15:33 - 00:15:52]
believe that they can be more efficient in how they operate. We did a podcast about efficiency as the new superpower not that long ago and are looking to use new tools like AI to help them become more efficient. Clearly that's possible but I'm not sure that necessarily there's an easy path to understand how you do that.
Alex Dean [00:15:52 - 00:16:42]
Well maybe let's talk about some specific use cases because I think that's an interesting place to look for efficiency. So I think within the retail life cycle we're seeing some really interesting areas where AI is being adopted. So a very traditional place has been around scoring your customers, understanding their predictive behavior, what's their propensity to buy again, what's their propensity to churn etc. So that's been a very interesting rigorous data science area. I think the new interesting area for efficiency is a lot around the kind of Gen AI and the shadow adoption. So a lot of retailers looking at how to create campaign copy or SKU copy and things like that for their merchandising side as well. So that's a really interesting area where there's a lot of cost in those businesses in those areas that potentially Gen AI can help with.
David Worby [00:16:41 - 00:17:13]
I think we'll come a bit later to some of those areas and maybe also some of the more progressive and forward-thinking options that might exist, but Mark, I just want to come back to the retail capability challenge here, namely retail, particularly in Europe, certainly in the UK and maybe even globally, has gone through a hell of a change. Retail margins are at the lowest they've kind of ever been. We talk about the brain drain in retail as if it's been going on for years because it quite frankly probably has.
Mark Pinkerton [00:17:11 - 00:17:13]
Yep. Yep.
David Worby [00:17:13 - 00:17:19]
The capability to actually adapt to this change, how do you see that? What are the challenges that you kind of think about?
Mark Pinkerton [00:17:19 - 00:18:00]
I think the capability for many retailers is quite limited and that is the challenge because, as you rightly say, the bright and intelligent people who you would want to be driving these sorts of things are to some extent migrating elsewhere and they are going to the new digital native based organisations where they can use data to do things because it's more cutting edge and they will learn. I mean in many cases e-commerce is now a mature industry and they are the hub of the digital natives within many retailers and in my experience many retailers are actually struggling to maintain their capability and they're trying to simplify
David Worby [00:17:57 - 00:17:58]
Thank you.
Mark Pinkerton [00:18:00 - 00:18:12]
costs at the same time because obviously, as you said, margins are under a lot of pressure so they're actually interested in streamlining, cutting costs and going through that process.
Alex Dean [00:18:12 - 00:18:43]
I definitely recognize that, and we're always interested in the fact that pound for pound, size for size, retail organizations have smaller data teams than we would see in other industries, including even traditional newspaper media and things like that. So I definitely recognize that. I think one thing that is helping counteract that is the tooling, the models, the data platforms are just getting better and better, and they are getting cheaper as well. So that is the thing that is helping.
Mark Pinkerton [00:18:43 - 00:18:51]
But there's undoubtedly a first-mover advantage from those players who did invest ten years ago and have had data science teams.
Alex Dean [00:18:48 - 00:18:48]
Yeah.
Mark Pinkerton [00:18:51 - 00:18:52]
Yes.
Alex Dean [00:18:51 - 00:18:52]
Yes.
Mark Pinkerton [00:18:52 - 00:19:02]
So their customer share will be less and their revenue per customer will be more. And you can probably see that they're still successful, even in their threatened environments. Yeah.
David Worby [00:19:01 - 00:19:18]
Yeah, I think if you're an enterprise-sized business that set out to build a data science capability three or four or five years ago, you're going to embrace the AI world, probably in-house, but that is the minority in the UK, and I think therefore what you've got
Mark Pinkerton [00:19:14 - 00:19:16]
Yeah, yeah, yeah.
David Worby [00:19:18 - 00:19:40]
is possibly the need to use third-party services when you want to dip your toe into whether it's efficiency AI-driven things or whether it's other sales growth-driven AI capabilities. You're probably not going to have the capability in-house. You're probably not going to buy it in-house. You might need to partner with someone who can take you on that journey. Would that seem fair?
Mark Pinkerton [00:19:40 - 00:19:47]
Yeah that seems fair and certainly we've worked with a number of retailers where they have had, you know, they've started setting up data lakes but they haven't
Alex Dean [00:19:44 - 00:19:45]
Hmm.
Mark Pinkerton [00:19:47 - 00:20:09]
really understood what they wanted to do with the data. Right. And therefore without that sort of strategic mindset at the start, great, you're collecting all the data, what are you going to do with it? So I guess they were hoping that at some point the technology and the suppliers could help them get to a point where they can do something with the data.
David Worby [00:20:07 - 00:20:22]
Yeah, I'm still always staggered how inefficient that process is, and often the insight capability, which is rarely very good, but when it is very good, doesn't have an operational, executional capability attached to it. They're kind of two separate things.
Mark Pinkerton [00:20:22 - 00:20:24]
Yeah, it's the linking of those two, which is the...
David Worby [00:20:22 - 00:20:31]
Yeah, it's the linking of those two, which is the... The business gets on with its job it's always been doing, and we've got this data science team over there. I'm not sure 100% what they're doing.
Mark Pinkerton [00:20:30 - 00:20:34]
And they're very retrospective in the way that they're looking at the data.
David Worby [00:20:31 - 00:20:32]
And there's very retrospective.
Mark Pinkerton [00:20:34 - 00:20:38]
It's to analyse what has happened as opposed to inform what is about to happen.
Alex Dean [00:20:35 - 00:21:14]
is to analyse what Yeah, 100%. I mean, we saw this in other industries, not so much in retail, but in other industries, we saw this kind of field of dreams, build it and they will come phenomenon, especially sort of 2019 to 2022, where there was a lot of investment in these data platforms, build out of data teams and data science teams. And they hadn't connected through to a really big like commercial reason to do this, how to make more money, how to save money. And that really got ironed out a lot last year. And I think retail actually sort of missed that wave a little bit and is now catching up and gets to actually be more concrete on the use cases.
David Worby [00:21:14 - 00:21:39]
Yeah, that makes sense to me. OK, in a moment, Alex, I'm just going to turn to you to talk to us a little bit about where the next set of use cases might come from. Before I do that, Mark, Alex mentioned a little while ago about some of the AI capability being used by retail. I think you've got an idea about the four big areas of where we see AI being used. Do you want to run through those? Yeah.
Mark Pinkerton [00:21:39 - 00:21:55]
Yeah, I think at a practical level, where we are seeing retailers use AI is obviously around CRM, trying to understand what the customers are doing, the behavioural data, which kind of is getting into your sort of world, Alex, but not quite there yet.
David Worby [00:21:51 - 00:21:51]
Hmm.
Mark Pinkerton [00:21:55 - 00:22:16]
And certainly the CRM packages are all doing with AI. Then the second one would be around content, and there's just been an explosion of content production, which we now know Google is trying to suppress a little bit. All of it AI driven, and to some extent that's a race to the bottom, as we'll come back to.
David Worby [00:22:15 - 00:22:16]
The
Mark Pinkerton [00:22:16 - 00:23:03]
Then on imagery, where there's very much a production gap between needing to take lots of images. Being able to create 360 degree videos from static images, all of which AI can now do. So that's about improving the process capability for retailers. And then from a user perspective, the whole search and merge piece, where we have always been advocates of tools being actually better than human beings at curating the list, A, because you can personalize it, and B, because if you've got a big volume, you can't do it very effectively. And AI has very definitely taken the lead in that space.
David Worby [00:23:03 - 00:23:07]
Okay, alright. I think at that point, we'll take a short break.
Mark Pinkerton [00:23:10 - 00:23:35]
Maybe we should take a break here, for a message from our sponsor, SWEFT, Retail's leading solution for product workflow automation. SWEFT is designed to replace the in-house tools and spreadsheets retailers use to manage product workflow, by bringing buying, planning, marketing, studio and e-commerce teams together on a single powerful platform, to launch products faster and more efficiently. Find out more at gosweft.com
David Worby [00:23:47 - 00:24:04]
I'm sure there are other areas that some of our listeners are using, and we'd love to hear your story, so do let us know where you're using AI. But Alex, let's come to the next tranche of how retailers or other sectors could be using it. Where do you think it could be in three to five years' time?
Alex Dean [00:24:04 - 00:25:24]
Yeah, great question, great area to explore. So first of all, I recognize all of the different areas for retailers that Mark talked about. I think there's a spot on. I'll come back to the personalization one for a moment and then I'll talk about two net new ones that we're excited about. So just on the personalization, these AI approaches are really powerful. So for a few years we've been able to create rich propensity scores. What we can now do with GenAI is create much more tailored decisions on next best action and we can get into like hyper personalization. So we can get to point in time messages, creatives that are literally just for one shopper or one potential new customer. So that whole personalization area is going to be very, very exciting over the next couple of years. I think the second is AI is an input into this, but it's a trend in retail that I wanted to call out, which is this rise of kind of retail media and the retail media network. So we've got a lot of customers in media itself that have been building very rich audience profiles for years and they use those for, you know, advertising and monetization. We're now seeing that coming to retail. Some of the larger retailers are now building their own audience segments, understanding their own shoppers, and then figuring out how to kind of monetize that, often working with the brands that they're retailing. So that's a really interesting growth area for retail.
Mark Pinkerton [00:25:24 - 00:25:43]
The challenge around cookies has probably driven from my perspective, so you've got to go to first party data because cookies won't allow you to use third party data, or the legislation won't allow you to do that, and therefore you've got to build your own data capability to offset the drop in data that you're getting.
Alex Dean [00:25:44 - 00:26:00]
Which is actually working out really well for retailers because retailers are now actually building and investing in their own first party data asset using technologies like Snowplow and they're then able to monetize that and they've got some of the best audience or shopper profiles out there so that's super exciting. I'll give you guys one more wacky
David Worby [00:26:00 - 00:26:02]
Yeah, go.
Alex Dean [00:26:00 - 00:27:22]
idea. We're here for the wacky. So I think a lot of the narrative around Gen AI and LLMs and chat GPT and all of this at the moment is around the idea of making existing processes more efficient. Maybe you need fewer people working on your product copy, maybe you need fewer support desk agents. And I recognize that and that is a trend and I recognize that retailers are always trying to work on their margins. But I'm excited about this other area which is more like well what can these AI models, these LLMs do that we could never have afforded to do in the past. So imagine a really large scale grocer with 10 million customers in a new LLM mode every single one of those 10 million customers could have their own personal concierge or their own personal customer success manager which is really exciting that just wouldn't have been able to be possible before. Or each SKU could have its own personal merchandiser which sounds totally nuts but you know once you stop thinking in terms of almost these agentic approaches are just going to kind of supplement or augment existing humans. Think about what new roles a retailer could open up powered by AI and that gets me excited. Yeah that gets a bit mind-blowing doesn't it. I guess
David Worby [00:27:19 - 00:27:38]
Yeah, that's gets a bit mind blowing, doesn't it? Yeah, I guess you're kind of saying that we no longer need to exist using averages of averages. And we can take unique transactions or unique pieces of data and manage those as individual pieces of data rather than as collective groups of data.
Alex Dean [00:27:28 - 00:27:43]
Exactly. Yeah, and it's the process all technology goes through. So when video games finally
David Worby [00:27:38 - 00:27:39]
Yeah, and it
Alex Dean [00:27:43 - 00:28:02]
were able to do 3D engines, everyone wanted to replicate the real world. Then you're like, well actually no, let's create completely new worlds that you can't visualise in the real world. It's the same with this. The current agentic approaches are looking at current roles inside a retailer, inside a newspaper group or whatever, but where's it going to go in a couple of years?
Mark Pinkerton [00:28:01 - 00:28:05]
But their current approach is rely on a human to make a decision. Yes.
David Worby [00:28:05 - 00:28:06]
Yes.
Mark Pinkerton [00:28:05 - 00:28:14]
And effectively what you're saying here is that the agent... Yes. The virtual agent will make a decision. 100%. Because it has to, because you can't scare it otherwise.
David Worby [00:28:14 - 00:28:49]
Well, I think that's fabulous. But I'm going to bring us right back down to earth. Why are our chatbots so bad? Now, I know this is kind of a proxy for the fact that there's a massive gulf between the theory of AI and the stuff we read about in the books and stuff and the practical experience a lot of customers will be having right now of how some things work. And we kind of picked out before we started this podcast, one of those areas where we all get very frustrated, and that is chatbots being terrible. We try to use them and we
Mark Pinkerton [00:28:48 - 00:28:50]
Yeah, I tried to use the...
David Worby [00:28:49 - 00:28:51]
give up. How do you deal with that?
Mark Pinkerton [00:28:51 - 00:28:53]
How many have you used this week?
David Worby [00:28:51 - 00:28:53]
Well, I think that's a really good question. I think that's a really good question. I think that's a really good question. I think that's a really good question. I think that's a really
Mark Pinkerton [00:28:53 - 00:28:55]
Well, probably six.
David Worby [00:28:53 - 00:28:55]
good question. I think that's a really good question. I think that's a really good question. I think that's a really good question. I think that's a really good question. I think that's
Mark Pinkerton [00:28:55 - 00:28:57]
Probably used about the same number.
David Worby [00:28:55 - 00:28:57]
a really good question. I think that's a really good question. I think that's a really good question. I think that's a really good question. I think that's a really good question. I think
Mark Pinkerton [00:28:57 - 00:29:21]
And all they're doing is actually re-representing the FAQ information from somebody's supplier help file. And that's really not helpful in terms of answering questions because probably somebody's read those before they asked the question, which they expect to have a degree of intelligence responded to. They're so stupid.
David Worby [00:29:20 - 00:29:26]
They're so wrong and so bad. Yeah they're frustrating aren't they? But I guess the
Mark Pinkerton [00:29:21 - 00:29:23]
They're so wrong and so bad.
David Worby [00:29:26 - 00:29:49]
point, the wider point, is there's a gulf between the theory and the practice. There's a gulf between how we all get very creative and very kind of enticed by the future and sometimes can't quite connect that with the reality of where we are today. And I think that's true of any trend.
Alex Dean [00:29:47 - 00:32:12]
It is, but I think the chatbots are a nice specific example to call out, you know, how the roadmap of AI is going and where we may be, you know, sometimes tripping up a little bit along the way. So I think the first thing to say is there's of course a set of almost like legacy chatbots that were built and designed before this whole like LLM gen AI wave. And those chatbots will often have a fair amount of information about your business to your point, but they're very dumb, like they just do not know how to have a rich conversation with you. So they're basically just menu, you know, sort of obfuscated menuing systems. If we then jump to new LLM powered chatbots, so imagine something that's got almost like the power of chat GBT underneath it, we see two separate different problems that can, we can have either them or both of them. So the first is hallucinations. And it's the idea that these LLMs have been trained on an awful lot of stuff out there, a lot of public internet data, a lot of junk as well. And it's hard to constrain them to your specific business domain. And there was a big, there was an example with an airline recently where the chatbot just went rogue, talking about like refund policies, it was just making them up. So we've got a challenge around, we've got these almost like, you know, super smart LLMs now, but they're very scatty brained, and it's hard to constrain them and get them to not make stuff up. The separate other challenge with the LLM approach is whether those LLMs are being given enough data about your business. And there, there's a really, there's different techniques people are trying, but most of the industry seems to be consolidating on something called RAG, R-A-G, which is retrieval augmented generation. And the easiest way to describe that is think of this as the next generation of enterprise search. So in the old world, you had a search engine, it would look at your SharePoint and answer questions. In the new world with RAG, what you do is you take data, you take insights about your customers or your SKUs or your supply chain or your FAQs, and you process it in a way that's really fast and efficient for an LLM to look at. And basically what you then do is when you get the LLM to answer a question, you tell the LLM a couple of things. You say, go and look in the database to retrieve important information and do not make stuff up.
David Worby [00:32:12 - 00:32:15]
You just explicitly tell it not to invent things.
Alex Dean [00:32:12 - 00:32:19]
And you literally, you literally tell it, yeah, it literally do that. And then you get a decent chatbot out the other end.
David Worby [00:32:19 - 00:32:35]
But I suppose, just listening to you answer that first part of the question, why would we expect anything that's trained on the internet to be perfect? Because the internet isn't, the world isn't, we aren't, nothing is, and therefore, dump stuff in, dump stuff out.
Alex Dean [00:32:31 - 00:32:42]
Right. Right, exactly. Think of that training as like us putting the LLM through primary school, and then how do we like go and take it to specialised?
David Worby [00:32:35 - 00:32:35]
Right, exactly.
Alex Dean [00:32:42 - 00:32:43]
Oh, Microsoft!
Mark Pinkerton [00:32:43 - 00:32:50]
Microsoft was the classic example, wasn't it? Within 24 hours it was a racist...
David Worby [00:32:49 - 00:32:53]
Racist. Yeah, yeah. Because the world is.
Mark Pinkerton [00:32:50 - 00:32:55]
Because the world is. Or parts of it are. Elements of it are. And we found those elements.
David Worby [00:32:55 - 00:33:29]
Okay, look, we're rapidly running out of time because this is a subject that we could probably spend the whole day talking about, but I know we can't. So let's just try and keep our kind of responses to this kind of as short as we can. I say that for myself. Where do we go to get started? So we're listening to this podcast, we're a retailer, we've kind of dabbled with some of it, but how do I really get my head around the way to get started? What three things would we suggest to our audience they begin to think about or do?
Alex Dean [00:33:30 - 00:33:57]
So my first recommendation, if you haven't done it already, is go and poke AI with a stick and see what it does. So get a chat GPT account or get a mid- journey account, which is an image generator, and just play around with it with prompts. You don't have to be a data scientist, a data engineer, a SQL whiz, you don't have to be any of those things to actually interact with prosumer gen AI, these LLMs, but it is a learning curve.
Mark Pinkerton [00:33:57 - 00:34:00]
Oh, it's an amazing learning curve.
Alex Dean [00:33:58 - 00:35:00]
Oh, it's an amazing learning. Yes. So that would be my first port of call. I think the other interesting area, once you start thinking about inside your business, is try and understand what you've got as kind of raw ingredients. So the only approaches that work here are these kind of crawl, walk, run approaches. So you're not going to build a whole shiny AI infrastructure in one go, but you can pick a use case you think is important, and you can figure out how to go through those steps. To the earlier point, you might be working with agencies, SIs, you might be working with your contacts at Snowflake or Databricks or Azure or whatever. But that's really important. I think the good news is that you are going to need some kind of data platform to build on top of, but you probably have some of those ingredients inside your retailer already. Yeah, that sounds good.
David Worby [00:34:59 - 00:35:11]
Yeah, that sounds right. I think that the crawl, walk, run things are good things. I think the temptation is to kind of try to achieve a huge amount in a very short space of time, whether you tell your IT department or whether you don't tell your IT department.
Mark Pinkerton [00:35:10 - 00:35:45]
Yeah, and I just want to give an example, I think, at a very practical level, because I know people who've been doing it recently, is actually applying AI to your customer research and your analytics side, which is a data set that you have that is often quite rich. So, as a base level is something that is yours that you can learn from, and it's not easy necessarily to draw conclusions from all of the data of random things that are going on, whereas AI will pull out the trains for you much more easily. Thanks for tuning in.
David Worby [00:35:44 - 00:35:45]
It's time to get back to work.
Mark Pinkerton [00:35:45 - 00:35:48]
And that's very much in the cruel world.
David Worby [00:35:45 - 00:35:48]
We'll be right back. And that's very much in the... We'll be right back. We'll be right back.
Mark Pinkerton [00:35:48 - 00:35:49]
Yeah.
David Worby [00:35:48 - 00:35:49]
Yeah. We'll be right back.
Mark Pinkerton [00:35:49 - 00:35:54]
And I think some of the merchandising tools which have AI applied to them, because you could... And I think
Mark Pinkerton [00:35:56 - 00:36:02]
We'll be right back. We'll be right back. things infinitesimally through the right database if it's set up in the right way. That, again, is an example of crawl.
Alex Dean [00:36:02 - 00:36:15]
Yeah, that's a great point. Find where you've got rich historical data, whether it's, you know, supply chains, use, customer support, customer 360, and that's an incredible place to start. Fabulous.
David Worby [00:36:15 - 00:37:03]
Okay look guys, we've run out of time I think because that's been a fabulous subject and maybe one we'll come back to at another point because I think there is still so much more to unpack. I think my final point on the three things would be if you're starting your journey and you're listening to this great advice but you're still unsure, reach out. Alex would be very happy to take a phone call, as would we and we'll pass it on, but don't hesitate to reach out to the network of third-party services that do exist and we know which one we would call because it could save you an awful lot of time and get you off the ground in a much better way than you might otherwise do without some third-party advice. Okay right I'm gonna say that's it. Thank you to Alex. Thank you very much. And thanks to Mark. Thank you. And we'll see you again on the next one
Other similar Episodes in Season 2:
Episode 3 - Understanding headless and Compossable
Episode 2 - Strategy and Planning
Episode 6 - Optimising physical and Digital
Articles you may be interested in from Our Blog:
Why do many Headless/eCommerce strategic initiatives fail - 8 key reasons
Man vs. the Machine: The Future of AI in Retail and Digital Commerce