Latest / Inside Commerce: Ecommerce Strategy, CX and Technology Podcast / EP321: Reframing Ecommerce's Build vs. Buy Debate - Practical Uses Of AI To Clean & Optimise Product Data
Transcript
- 0:00This is where the where the technology is really empowering.
- 0:02You know, we work with various consultants and clients.
- 0:05I can see people within those businesses who are who are
- 0:07really embracing this. Until now they've they've lacked
- 0:10the technical skills to actually execute on that vision.
- 0:12You can see those people being fasting more productive as a
- 0:15result. You listen to Insidecommerce, a
- 0:19weekly ecom strategy, technology and CX focus podcast hosted by
- 0:23experienced digital consultants and re platform specialist James
- 0:25Get and Paul Rogers. Get insights into ecom strategy,
- 0:28leading brands, hear from industry thought leaders and
- 0:30leading tech vendors with practical advice to help you
- 0:32improve your online CX. This episode is sponsored by
- 0:38Worldpay. Got another exciting debate
- 0:41discussion rant going on today. So we're talking about I'm
- 0:45actually stealing this terminology from our guest.
- 0:47I'll introduce in a minute, bring in the build versus buy
- 0:50debate back into the light. I think it's a really nice way
- 0:52of framing it. So we're looking at practical
- 0:55ways to use AI to clean and optimise data, to help with data
- 1:00management, data migrations without having to spend insane
- 1:03amounts of money on licences and an ongoing commitments to
- 1:06tooling. And the reason for talking about
- 1:08this and using practical use cases and talking about what
- 1:10we're seeing and what we're doing is there's a lot of talk
- 1:12about what AI is for and how it's revolutionising the world
- 1:16and how amazing it is. But there's less like kind of in
- 1:19depth discussion about, OK, So what am I using it for and how
- 1:22am I going to ban it? And what are the pros and the
- 1:23cons? And that's what I want to do
- 1:24today with our guest Chris, who is the Co founder on stage.
- 1:28So welcome back to the pod, Chris.
- 1:31Thanks, James. Nice to be back.
- 1:32Thanks for inviting us back. Yeah.
- 1:34This is something that, you know, we've been very much
- 1:37engaged with over the past. Yeah, couple of years, really.
- 1:40So I know we met the other week and you were, I think you were
- 1:44you, you were, you were mired in the middle of a, a migration at
- 1:47the time. So this was very much kind of
- 1:49top of your mind, wasn't it? Yeah, exactly.
- 1:51And you know what I like is that you've got additional skills I
- 1:54have because you are running an agency that is building and
- 1:57delivering, whereas I work within projects of Ising and
- 2:00guiding. So I do less active work in the
- 2:03AI than you do, although I am using it to to fulfil some use
- 2:06cases of projects. So though this would be a really
- 2:08interesting one for our listeners about how we can how
- 2:11you blend the use of tools with, I think you framed in the DIY
- 2:16approaches, which I think is quite a nice way of talking
- 2:19about this. But before we get into it,
- 2:21because there'll be people who listened to you on the podcast
- 2:23previously, people who know you and people who don't.
- 2:24So do you want to give people a bit of a taste for who you are
- 2:27and what onstate is? Yeah.
- 2:28So I'm Co founder of Onstate. We are a Shopify Premier partner
- 2:35based up, up or down depending on where you are there in Leeds.
- 2:40There's eighteen of us. We yeah, we delivered.
- 2:45We've been a Shopify partner for five or six years now.
- 2:48Prior to that we were Magento Salesforce partner and my
- 2:52background was in was in multi channel retail.
- 2:55I was ecom manager. I have ecom for a High Street
- 2:58fashion retailer before setting up one state.
- 3:02And yeah, so and we've been as a business, we've been very
- 3:06interested in a, I know ever since the advent of Chat 3.5 I
- 3:10think. And very much came at it from a
- 3:13point of view that, you know, from existential crisis really
- 3:16all of a sudden were presented with a technology which knows
- 3:19things and can write code, you know, which are two things that
- 3:22agencies do a lot of. So we've sort of been down the
- 3:25rabbit hole and, and back out really on it.
- 3:27And you know, we've the past year, I say we've fairly
- 3:30successfully integrated it into, you know, lots of, lots of the
- 3:34work that we do as a business. Yeah.
- 3:36And I think so one of the areas we're going to private into more
- 3:39is around kind of data management, product data
- 3:41management specifically. And this came out a few chats
- 3:44we've had about how AI can be used to help with data
- 3:49transformations for product data when you've got product data in
- 3:52a source. And I've been working a project
- 3:53received Magento to Shopify in, you know, the legacy data has
- 3:57always got mess in it. It's people create hacks to get
- 4:01data into and a developer finds a way of getting that onto the
- 4:05front end. But the data might not be
- 4:06structured in the right way, might not be set up in the way
- 4:08you ideally want it in a clean new system, especially in the
- 4:11light of AI product discovery and better structured data.
- 4:14And then you realise that trying to get it from A to B without
- 4:18paying somebody loads and loads of money to do manual
- 4:20transformations is actually really quite challenging.
- 4:22And this is where the AI comes in.
- 4:25So I think, I think framing it and then I want to get into like
- 4:30some use cases you worked on. But the framing for me is that
- 4:32there are three, three main approaches that you now get
- 4:36taking out the equation, paying a data specialist to manually
- 4:40grind through this in Excel using form and stuff.
- 4:42But it is you invest in a specialist and product
- 4:46management tool, PIM and there are more modern AI focus ones
- 4:49like M fast, for example. But there's also existing and
- 4:52Pims in the market like Pimbley that have, you know, rolled in
- 4:55AI capabilities on top of the core platform.
- 4:58You outsource to a tech partner to do the entire thing, who will
- 5:01use AI tool in, in order to write and, and you know, or
- 5:07create custom apps to do this. Or you do the DIY approach,
- 5:11which is use your own toolkit and that might do 80% of the
- 5:15work and then you need to do the other 20% another way.
- 5:18So there's, there's this mix of approaches and to frame this, I
- 5:22guess it would be useful for you to talk people through coming
- 5:25from a, a systems integrator agency, you know, developer's
- 5:28point of view, what, what is the evolution and the use of AI as
- 5:32you've seen and like, how has it changed and what impact are you
- 5:34now seeing it have on project delivery?
- 5:38Yeah. OK.
- 5:38Yeah. So not to unpack there.
- 5:40So we, you know, like a good chunk of the projects we work on
- 5:44our migrations. So this invariably comes up.
- 5:47And obviously over the last five years we've seen, you know, a, a
- 5:49decent shift towards Shopify. So there's been a lot of focus
- 5:53on taking data from legacy systems and convert them into
- 5:56Shopify. The way, I mean the way we've
- 5:59have always approached that as a business, I mean most lots of
- 6:02people who work on Shopify, the thingy with Matrixify, formerly
- 6:06Excelify, which is I think Shopify actually recommend it
- 6:10now as the way to import and export data into the platform.
- 6:14So we're generally taking or have been taking data like
- 6:19product export, WordPress exports, whatever from that,
- 6:26from the old system and converting into Matrixify.
- 6:29And then in just then that way Matrixify is, you know, very
- 6:32mature products, very capable. It's got some nice little
- 6:34workflow tools in there like having a dry run before you
- 6:37actually import the data. And we found that, yeah.
- 6:40So that that's been the process and the way we've integrated AI
- 6:43into that is so for example, if I think back a year, 18 months
- 6:47to a project that we did where we migrated from another
- 6:49platform, we'd get the data into Google Sheets.
- 6:53It's a CSV, you get a product export.
- 6:56And then in terms of translating that, we're trying to get that
- 7:00into Matrix if I format. So previously it'd have been
- 7:04spreadsheet wrangling. We might have written some
- 7:06scripts. Yeah, as of the oven of, you
- 7:09know, ChatGPT 3.5 and onwards, you can now you can get AI
- 7:13systems to help you transform that data.
- 7:16There's a really good anyone who knows Excel, you'll know about
- 7:19macros, which are basically small programs you could run
- 7:22within Excel. We use Google Sheets and the
- 7:24Google Sheets equivalent is called Apps Script, which is
- 7:27basically JavaScript that allows you to add new functionality.
- 7:30And it is, you know, if if you use Google Sheets a lot and you
- 7:32don't know about Apps Script and you need to do complex things in
- 7:35it. And it's well worth looking at
- 7:37because it is a superpower and it can really extend what the
- 7:41platform what, what, what Google Sheets does.
- 7:43So I'll give you an example. Google Sheets, like Excel, has
- 7:48got lots of built in functions like song and average and things
- 7:50like that. Using Apps Script, you can
- 7:52create your own functions. So let's say in one cell you've
- 7:57got a bunch of data. Let's say it's a load of tags
- 8:00from product tags from the the legacy platform you're importing
- 8:03from. And you need to extract the
- 8:04value or all those values, you can fire up GB Team Claude or
- 8:12even Gemini, I think Gemini's built in now and give it the
- 8:15source data and say write me an app script that will allow me to
- 8:19take that data and turn it into the structured, the structured
- 8:22format. And that might be taking all
- 8:24those tags, for example, separating them out, put them in
- 8:26separate columns as meta fields. And it can do that.
- 8:29And you don't need to know JavaScript because the LLM will
- 8:33write that for you and it can tell you how to install it if
- 8:35you're not that comfortable as well.
- 8:36So straight away, you know, we, we were using that to sort of
- 8:40augment the, the velocity which we could, you know, which we
- 8:46could manipulate data within specifically Google Sheets.
- 8:49Ready to grow your business online?
- 8:51With World Pay e-commerce, you can deliver fast, secure
- 8:54payments your customers love, backed by global reach and real
- 8:59people powered support. Scale with confidence and
- 9:02enhance your customers experience with World Pay
- 9:05E-commerce. Yeah, I've been doing this a lot
- 9:09recently. The use case I had, I had a
- 9:10client who has old blog content in WordPress and Magenta and
- 9:14needn't get to Shopify's blog. But you've got a dump of like
- 9:19really unclean HTML. And actually, yeah, Chat GPTII
- 9:24got the subscribe the the professional license rather than
- 9:29just the the free one, because you get more computational power
- 9:32and you can do deeper scripting in it.
- 9:34And I'm not a developer. I can't write Python scripts, I
- 9:36can't create this stuff myself. It just wouldn't work.
- 9:40And I play around for age. I used, actually, I used Gemini
- 9:42in Sheets and I found that Gemini doesn't go nearly as far
- 9:45as GPT does in terms of ISP to execute more complex Python.
- 9:49So I was, I got to the point where I was getting it to do
- 9:52things like replace CDN strings to get the correct new one for
- 9:55Shopify, removing specific HTML classes, formatting things,
- 10:03deleting certain parts. Yeah, just there's probably
- 10:06about 20 different commands that I needed.
- 10:09And I've found that the iterative process in ChatGPT far
- 10:12quicker, more accurate. And it's amazing actually that
- 10:16that would have 588 blogs reclaim and reformat.
- 10:19And if you import all that data and then have to redo it, I
- 10:21mean, Christ, you've imported loads of crap HTML in the 1st
- 10:24place, which is just bloating the sigh.
- 10:26But it takes ages manually to go through that.
- 10:29And it by doing that alone, I estimate maybe 70 to 75% of the
- 10:33issues were stripped out by about half a day's work in total
- 10:37over all the iterations, which then stood up the ability of the
- 10:41content team to then go through and just sense check the
- 10:43formatting. And and then the last step is
- 10:46somebody wants all of the content and collections and
- 10:49products are in Shopify doing a internal link analysis of
- 10:53finding the broken links, so we can update those and then re
- 10:56import those. Yeah, that that was for me.
- 10:59I know that I would have been less efficient than other people
- 11:02because it's the first time I did that.
- 11:03But there's an eye out there to actually how quickly and it's
- 11:06you can learn on the go with it. Yeah, it is.
- 11:10You've raised a good point there about internal linking as well,
- 11:13because we've done that with importing WordPress content
- 11:17because it, you know, embedded within that are these valuable
- 11:20internal links like within within the body copy.
- 11:22Now if you're LLM that you're using to clean the HTML also
- 11:26knows the new URLs of of of all those posts, you can potentially
- 11:30update that at the same time as well.
- 11:32The temptation is right to just. In the ideal world, you would
- 11:36take all your product data and you would dump it in the LLM and
- 11:39it would sort it all out. And it can do that to a degree.
- 11:42But the limiting factor is the LLMS context window and the
- 11:47amount of the amount of data that it can handle in it's own
- 11:51memory at any one time. But the way that the modern
- 11:55tools deal with this is they don't actually deal with all
- 11:57that. What they will do is they will
- 11:59write their own Python scripts in the background.
- 12:01And if you look at the chain of thought and, and the reasoning
- 12:04that the, the sort of the output, you can see them doing
- 12:06this. They'll be writing Python
- 12:07scripts, they'll be importing it, they'll be analysing the
- 12:10data and then making recommendations and, and writing
- 12:14scripts to manipulate it off the back of that.
- 12:16And that, that avoids this situation where they need to
- 12:19cram all this data, you know, and if you're moving thousands
- 12:21of products, you're going to comfortably exceed any, any
- 12:24context window. Even Gemini, which is, you know,
- 12:27has a million tokens. It's probably the largest of
- 12:29the, the kind of the main models at the moment.
- 12:31It's still far too big for that. Yeah, the the other thing that I
- 12:35found it really useful is I, I know you can use advanced
- 12:37formulas and functions in Google Sheets or Excel to normalise,
- 12:42you know, and restructure like, you know, take take a product
- 12:45title, lower case it, remove spaces, put in hyphens.
- 12:49You can do that through like like structured formulas, but it
- 12:54it takes time if you, if you don't know all the formulas
- 12:56available, it takes time to research and learn them.
- 12:58Learn how to concatenate different formulas into a master
- 13:01one to make it work so much quicker.
- 13:03If you're not an advanced user to go into an AI model and just
- 13:08go right, I need I need you to do step one, do this step 2345.
- 13:12And then you're right. You can see it thinking and you
- 13:14can see it formulating the Python responsive.
- 13:16Then they say, this is what I think you want and this is what
- 13:18I'm going to do is that's correct.
- 13:19And then you can, what I found really helpful is you give it
- 13:23examples and go right in row 4 column DD4 should the output
- 13:28should now be X because the input was and you give it, you
- 13:31give it 3 or 4 tangible examples of what the correct responses
- 13:34should be so that it can learn on the go.
- 13:37Which really helps when you get into, I guess some bit more
- 13:39complexity about when you're trying to get it to match things
- 13:41like name value pairs like colour and for it to know silver
- 13:45is a colour blue, all those different ones.
- 13:47So yeah, I, I, I find it quite powerful that way.
- 13:51Yeah, it's when going through this process, it does force you
- 13:55to engage with the underlying content structure on your site
- 13:59and trying to think about the primitives, the the, the, the
- 14:02underlying kind of concepts that that drive your business from a,
- 14:06from a content and product perspective.
- 14:08So we're working with a Wellness client at the moment and we're,
- 14:11you know, we're engaging with that.
- 14:12So we're thinking about, right, things like ingredients and
- 14:14conditions and products and brands and how they all need to
- 14:16relate because if we crack that, if we get that, then get that
- 14:21right, we can build really interesting educational
- 14:23discovery experiences and stuff on top of that.
- 14:26So I've described just now like the process we went through a
- 14:30year 18 months ago. And then recently I took that
- 14:33same product export, which is just a list of product
- 14:36information with tags. And there was some
- 14:38categorisation implied in there and I gave it to I think Claude,
- 14:42it needs to be said. I think about the last, the last
- 14:45iterations of particularly Claude, like Opus 4.5 and 4.6.
- 14:50I was getting a bit, it seemed like that, you know, the, the
- 14:52gains with every new version were marginal.
- 14:54But I think, and I have seen other people saying this, that
- 14:57it does seem like these last versions have have represented
- 14:59quite a big step on. I just wanted to see with what I
- 15:02could do. So bought, you know, one hour's
- 15:05worth of bought a bunch of cloud code credits, got that CSV, gave
- 15:10myself an hour and see how far I could get.
- 15:12So I gave her the CSV. It analysed it and then you know
- 15:17you can you can what they call you can zero shot this right and
- 15:20you can say right write me an import tools take this
- 15:22information and put it into Shopify and it will do a decent
- 15:26enough this stuff it like or if you know what you're asking for
- 15:30and this is like I think for an operator like you, James, like
- 15:33you know, you say you've got no technical background, but you
- 15:35know you know you're technical enough and you know you know
- 15:38what to ask for basically and you know how things like product
- 15:42data work. I think this is where the where
- 15:44the technology is really empowering for people who know
- 15:48that. And you know, we work with
- 15:49various different consultants and clients.
- 15:51And I can see these people within those businesses who are,
- 15:54who are really embracing this where they know what they want.
- 15:57But hitherto, you know, until now, they've they've lacked the
- 16:00technical skills to actually execute on that vision.
- 16:02And you can see those people doing that now and being vastly
- 16:05more productive as a result. But anyway, so rather than just,
- 16:11you know what, just giving it a very basic prompt, it took some
- 16:14time to like say, right, OK, well, I want you to go through
- 16:16this data. I want you to infer the category
- 16:19structure. I want you to figure out all the
- 16:21core fields and how they map from the old system into
- 16:24Shopify. I don't have to explain to it
- 16:26like what Shopify's core fields are.
- 16:28It's already in there. It knows, it knows this, it
- 16:30knows there's a price skew description, etcetera.
- 16:35I want you to convert all the tags into meta fields.
- 16:38I want you to find opportunities to turn meta fields into meta
- 16:42objects where we might want to extend that later, but you know,
- 16:45for like icons or whatever on, on the PDP also go through all
- 16:50the collections and figure out what those smart collections
- 16:52rule needs to be. So a smart collection in Shopify
- 16:56is, you know, rather than signing a product and saying,
- 16:58right, that goes in blue flowers, you're going to write a
- 17:03rule. There's going to be a rule that
- 17:04says right, colour equals blue and type equals flower.
- 17:08And I was, you know, within an hour, this is by no means of
- 17:11finished products, but within an hour it made like very
- 17:15impressive headway into analysing that data set,
- 17:18presenting it back to me in a way that, you know, I could very
- 17:22quickly evaluate sense check. And I did some spot checks and
- 17:25you know, it all checked out. I had pretty high confidence in,
- 17:29in, in, in the quality of the output.
- 17:31And I, I'd set off trying to make a tool like a A1 size fits
- 17:34all tool for managing migrations.
- 17:37But then in going through the process or rethought, this is
- 17:40like, well, I don't need to build a reusable tool, right?
- 17:42If I can build this thing in a matter of hours, then I don't
- 17:46need to think about longevity, right?
- 17:48We, we, we can whip up a tool for a one off job like this and
- 17:53get it to do the job and tear it down.
- 17:55You know, when the next opportunity comes around next
- 17:58week, and it might be from the same platform, but we'll write
- 18:01something again that is specific to that, you know, there'll be
- 18:03nuances and differences in in that for that client that we
- 18:06might approach differently. And this is the thing, you know,
- 18:09the AI part sort of ate my tool, you know, and, and the AI ended
- 18:13up doing the work and, and built me a nice sort of UI to layer on
- 18:16top of it. It's a difficult concept to
- 18:18describe, but it's something that, you know, as with
- 18:21everything, AII encourage people just to engage with it and just
- 18:25see, you know, see how far you get.
- 18:27If you if you're worried about spending too much money on it,
- 18:30get something like Google and anti gravity, which is a sort of
- 18:35coding environment, but you'd have to write a line of code.
- 18:37You can just prompt and you can ask it to build something and it
- 18:40can and you can see it, you know, see how far you get on.
- 18:42I think I would recommend that to anybody.
- 18:45Yeah, I'd. It brings me back to a point
- 18:47that I raised when Paul and I were doing our our like future
- 18:52gazing at the end of last year for about where things were
- 18:54headed this year. And one of the key things I've
- 18:56flagged about tactical, tactical middleware tactical solutions
- 19:00because we I think we've evolved from where it's
- 19:03one-size-fits-all to a more flexible tech sack.
- 19:06But you, you buy in a capability or you build something that's
- 19:09going to stand the tests, the test of time and that's your
- 19:12investment, that's your R&D investment.
- 19:15You know it's CapEx, it's great. OK.
- 19:17We've invested versus you're right conceptually for people to
- 19:21write off stuff for a one off project and go what I'm not
- 19:24reusing that. I've spent a couple of grand and
- 19:26I'm just going to throw it away. It's it's there's a real
- 19:30education piece now about understanding that tactical can
- 19:33sometimes be the right way of doing it, because actually the
- 19:35other way of doing it would be to spend a lot more money and
- 19:39gets the same outcome potentially, but get there far
- 19:43slower. Yeah, and to, you know, you buy
- 19:46something in which you then have to learn and then you have to
- 19:50bend to shape your specific need versus something that you know,
- 19:54you can, you could design to your, to your, to your precise
- 19:57needs. Now, I don't want to
- 19:59oversimplify this, right? It does.
- 20:01You know the, the thing about AI and anyone who's engaged with it
- 20:03will know this, that to get to that point I got to the other
- 20:06day, the 1st 80% is like magic in the final 20% and getting it
- 20:11to actually work. There's a lot of back and forth
- 20:13and quite, you know, and quite menial work where you're
- 20:16basically pleading with the machines to try and get it to do
- 20:18what you want. But.
- 20:19And there's a distinction to be made here right between people
- 20:22talk about it's the death of SAS.
- 20:23Yeah. And the idea is that this this
- 20:25this is going to become the norm.
- 20:27Whereas rather than buying in a product, I'm just going to
- 20:30describe my own and I'm going to build it specifically for me.
- 20:32And I think there's a distinction to be made between
- 20:35like these one off use cases where I think it is valid,
- 20:37right. If you're confident enough
- 20:39versus, you know, somebody I saw on LinkedIn the other day, I've,
- 20:42I've coded a version of Shopify. Well, you know, good luck with
- 20:46that because you know, you might have a shopping cart and it
- 20:48might have an admin on, but there's all that other stuff,
- 20:51you know, there's the whole ecosystem.
- 20:53If you built an API onto it, how are you going to extend it?
- 20:55And you know, as it, as it's like, you know, think back to my
- 20:58former role as an e-commerce manager, head of e-commerce, I
- 21:01could, I could build these tools, right?
- 21:04But don't want to build something that the business then
- 21:07relies on because I'm already, I've already got a full time job
- 21:10doing what I do. I don't want them to become a
- 21:13product manager and also the head of engineering as well and
- 21:17be responsible for this software which is now running the
- 21:19business. I think yeah, we we should be
- 21:21careful to distinguish the two. Yeah, 100%.
- 21:24I think that's really, really sage advice.
- 21:26And but for me, it comes to time to value when I work on these
- 21:28projects, I look at things and and we all know, right, whether
- 21:31it's you're doing a one off piece of do you know what we're
- 21:34updating? We're we're fixing our product
- 21:36data because, you know, it's not good enough, but we're not
- 21:37moving platform or you're moving platform and therefore product
- 21:40data migration is part of it. These the this work can become
- 21:44tedious and it can become stressful because it's not
- 21:47foolproof right there are you come up against barriers where
- 21:51you suddenly go shit that those attributes are wrong.
- 21:54The mapping in the current system is wrong and therefore
- 21:57it's given us all the values are wrong or that it's not set up
- 22:00correctly. There's all these continuous
- 22:02things or there's missing values.
- 22:03And for me, it's like it's not AI want this approach of AI to
- 22:08be perfect so I don't have to vias.
- 22:10I want to speed up how quickly I can get to the value part, which
- 22:14is right now it's in a decent enough place where someone can
- 22:18take it in a matrix price sheet and Polish it, or they can use
- 22:21bulk editing in a Shopify. Or actually the the core data
- 22:25now is robust enough to put into the PIM that you've already got
- 22:28so that you can then use the PIM data cleansing capabilities to,
- 22:32to Polish it how you want. It's that's where I'm coming
- 22:35from. It's let me speed up so that I
- 22:37don't have to spend 5 grand and another expert over here to get
- 22:42the the foundational work done. Yeah, precisely.
- 22:45And again, you know, if it is part of your role as well, it
- 22:50does afford you the ability to, as I say, engage with the data.
- 22:54And, and, you know, it's an opportunity to rethink how you
- 22:58approach things like content as well, especially in light of of
- 23:01of LLM based discovery and getting that, you know, the one,
- 23:07no one really knows what's happening there.
- 23:09Yeah, but there's a lot of people spending a lot of time
- 23:11thinking about it. And, you know, and it in large
- 23:15part the main recommendation at the moment is to sort out your
- 23:19data. And it's a it's a good
- 23:20opportunity, it's a good way into it to actually engage with
- 23:24this. And it does.
- 23:27It does force you to think in these sort of foundational terms
- 23:30about the data that runs you up in the content that is important
- 23:34to your business. Let's let's give people some,
- 23:37some like some practical advice on this as well.
- 23:39Because if you don't know what you're doing or if you come to
- 23:43the point you made earlier about, you need to understand
- 23:46what the outputs, what good outputs are like.
- 23:48Because if you can't define what a successful output is, you're
- 23:50going to be put in invalid or inaccurate prompts in and you're
- 23:54going to get something that you can't interrogate.
- 23:56So how can people use these tools intelligently so they
- 24:00don't end up, you know, faffing away hours of their time and
- 24:04getting nothing out of it? I think you know you will fuff
- 24:07away like a significant amount of your type to begin with.
- 24:10You know, as you explore it, you made a good point earlier on
- 24:14about being very specific about inputs and outputs.
- 24:21So a good, you know, if you're going to ask a system to do a
- 24:26transformation, say for between one product and and that same
- 24:31product for a different system and give it examples of both.
- 24:35So take the time to, you know, find your most complicated
- 24:38product, yeah, in your in your current export and then invest a
- 24:42bit of time in like, you know, figuring out what that looks
- 24:45like in the new one and give it those two pieces to work from so
- 24:48we can understand the transformations.
- 24:52That said, like the ability of the systems to work almost more
- 24:58autonomously on that is getting is getting better and better.
- 25:02People talk about prompt engineering as a, as a skill
- 25:04and, and, and the, and the ability to ask exactly the right
- 25:09question. That's definitely the case,
- 25:10right? Like my, my example earlier on
- 25:13where, you know, I brought my experience with Shopify to ask
- 25:17the system to create exactly what I knew I wanted versus what
- 25:20it came up with when left to its own devices.
- 25:22The version that, you know, thankfully for my own, my
- 25:25career, you know, it was still right, right, better than what
- 25:28it came up on its own. But the gap is narrowing.
- 25:31And you know, you might want to say, well, OK, well, have I got
- 25:34transformed it yourself to begin with?
- 25:36And it might, you know, it might make a decent fist of that.
- 25:38And it does give you a way into it again, start engaging with
- 25:42the the data. But that I think that's that's a
- 25:45good that's a good way to do it. Like if you, if you could
- 25:47specify the inputs and the outputs, as I said, the one that
- 25:51I built the other day built, I mean, you know, whisked up in an
- 25:55hour, I the system was doing all the work behind the scenes.
- 26:00What I, what I essentially built was AUI to layer on top of that.
- 26:04That allowed me to interrogate what it was going out with.
- 26:07And one thing these systems are great at is like building
- 26:10dashboards, you know, like that. And you'll see any, any anybody
- 26:12who's spent in time sort of like coding has produced something
- 26:15with a beautiful dashboard. And yeah, you could.
- 26:20You can certainly ask it to do that so that you can see the
- 26:22data that it's out or if you can validate it yourself.
- 26:25And if anything that is changing, you go back into the
- 26:28into the model and you ask it to make those changes and represent
- 26:31the data. Yeah, I think for me, the, the
- 26:35biggest learning I've had going through this entire process is,
- 26:39is accepting the, that it's not going to be perfect.
- 26:42And because in the same way that it's hard to write code so
- 26:45absolutely perfectly, nothing ever goes wrong with it.
- 26:47And it's absolutely like perfect on every level.
- 26:52But the same with these it's, it's to know that you're not
- 26:55going to get it 100%. Not even after several
- 26:58iterations. There's still a few things where
- 27:00it's not quite getting the accuracy of the transformations
- 27:02that I wanted in the product data.
- 27:04But it also the, the, the big caveat for me is, is being able
- 27:09to spot where it's clearly getting something wrong when
- 27:13it's told you it's done something and pointing outs and
- 27:15pinpointing because I've had this where I, the blog one where
- 27:18it said, yes, I've done this. I've replaced all of these with,
- 27:21of course, X with Y and I go and look and Y is, is not being
- 27:24replaced. And there's still loads of it's,
- 27:26you know, situation where X is in the, in the cells and it's
- 27:30been able to go in and go, but that hasn't worked.
- 27:32You haven't done it. And initially it's frustrating
- 27:35because it's like having a conversation with somebody.
- 27:37He says, yes, I've done it. And then you look at you, but
- 27:38you haven't done it. Yeah, I have done it.
- 27:40So you have to be able to unwind that and not get really wound up
- 27:43and start smashing your computer around and have the patience to
- 27:45to try and educate it, but also accept that you will probably at
- 27:50some point exhaust the ability of those models to do exactly
- 27:54what you want. Yeah, yeah.
- 27:57And like things like hallucination are are a problem,
- 28:00you know an unsolved problem. So I mean I I I was interested
- 28:04yesterday to see I took a desire a desktop and a mobile screen
- 28:09grab of home pages of one of our current one of our recent builds
- 28:13and threw that into like the latest version of Claude.
- 28:16You know the best. Yeah probably the best model out
- 28:18there at the moment said right recreate that pixel.
- 28:21Yeah pixel perfect. And you know, it doesn't all
- 28:24right, job it was, you know, it's nothing that you'd actually
- 28:27it was fine. You know, it's a decent, decent
- 28:29end of layout, but it was missing all the design niceties
- 28:31and everything that our developers have built in.
- 28:34But also it chucked in a completely new promotional
- 28:37banner right in the middle of the page, something really weird
- 28:40like a discount for Shopify Plus users or something, which was
- 28:42not implied anywhere in any of the designs, right.
- 28:46So when it's producing something visual like that or a piece of
- 28:49written word like your example, it's really easy to spot these
- 28:53things going wrong because it's just, you know, just
- 28:56self-evident. It's right there in the middle
- 28:57of the night. He's a bit built.
- 28:59If you're a non coder and it's producing code, you can't,
- 29:03you've no idea. We validate the output.
- 29:06I think the thing, you know, the thing, the approach I was taking
- 29:09the other day when putting that tool together is I'm not asking
- 29:13BLM necessary to do all the transformations itself.
- 29:16I'm asking it to build a tool to handle those trucks.
- 29:19There's no scope for it to hallucinate values, right?
- 29:22All it's doing is taking one form and putting the other.
- 29:26Yeah, yeah. And I think this is for me
- 29:28where, so the project I'm working on where I was doing
- 29:31some of this foundational work, they've the client has engaged
- 29:36with the team Mfas and Mfas have a tool that has lots of
- 29:41different AI capabilities with it around data enrichment, data
- 29:43cleansing, where you can create the AI logic in order to do
- 29:48mappings between two different data sets where attribute values
- 29:52exist, don't exist, or the attribute names are totally
- 29:54different and the values could be mixed up across multiple
- 29:57columns and you need to get into a standardised one.
- 29:59But actually doing the precursor work in the AI to clean up and
- 30:02have better data then puts the data in in a cleaner state so
- 30:06that the rules within that engine are more efficient.
- 30:10And therefore you go through fewer iterations.
- 30:12And yet you still have to go through iterations.
- 30:14But actually, that then means that you're focusing your
- 30:16license fee on getting more valuable outputs, rather than
- 30:19spending loads and loads of money on credit just to get what
- 30:22was crap data in the 1st place into a decent, usable state.
- 30:26Yeah, yeah. And there is like I say that
- 30:31there's, there is a cost aspect to all of this, like there's a
- 30:34monetary gun. There's also there's an
- 30:36ecological cost as well. There's an energy water cost as
- 30:39well, which that you know, I think is an industry we need to
- 30:41be mindful of. If you're, you know, if you're
- 30:44interested in engaging with going back to your sort of
- 30:50question around how do you validate things?
- 30:55I think 111 important concept that people don't often
- 30:59understand is that an AI driven application that's making use of
- 31:04an LLM, all it's doing in the background, right is making is,
- 31:08is generating prompts in the same way that you do when you
- 31:11sit down and you use track GBT or core or whatever.
- 31:14It's just doing that at scale and it's doing it
- 31:16programmatically. If you want to understand, I
- 31:20mean, unlike Google have a product called AI Studio which
- 31:24allows you to and there's a decent 3 tiers.
- 31:27It's not as good at the three tiers as it used to be.
- 31:29I mean they've recently reduced that.
- 31:31And prior to that, this was a very, very valuable tool for
- 31:34doing large data manipulation. But if you want to understand
- 31:38how it works, then you can, you can use that.
- 31:42It's sort of reasonably like a, a chat interface, but I'll give
- 31:47you an example. We used it last week.
- 31:50So we've got a client, we're doing a migword Shopify upgrade
- 31:52at the moment and we're looking to help them get the most out of
- 31:56the content they've already got. They've got Avastarkarp content,
- 31:58images, blog posts, everything. So I think it right.
- 32:01Can we add metadata to the blog posts to the, to the images?
- 32:05So fired up AI Studio pick a model and like we can pick a
- 32:09cheap one, but Gemini's got several different models.
- 32:12It's got pro at the top, flash and then flash like which is the
- 32:15fastest and most basic. But even using the most basic
- 32:17one, upload an image into it and say right, extract meaning from
- 32:21that, you know, bring it back in a in a structured format.
- 32:24Give me some, you know, something about the setting, the
- 32:26colour product. This these are photos of shoes.
- 32:29So I'm really interested in the clothes that people are wearing
- 32:32and you know the colour, the sort of ambience, you know, is
- 32:35it formal? Is it usual?
- 32:37And you can ask it to do that and it will output, it will
- 32:41output the, you know, a block of code which basically describes
- 32:45the image. And then you think, right, OK,
- 32:48well, if I could, if I could build a program that does that
- 32:52at scale, I can. As long as I'm very specific
- 32:54about, you know, these are the values, you know, it's going to
- 32:57be formal or, or or casual. They're the two that I'm going
- 33:01to allow you like as as long as you can be specific about that,
- 33:06you know, you could potentially then, you know, build a tool
- 33:10that will go through an archive of images and add fairly
- 33:13significant and useful metadata to all of them.
- 33:16Which again, you know, from if you're thinking of them building
- 33:19a front end experience and you've got a customer on the
- 33:22site who's looking for, I don't know, like wedding shoes, for
- 33:25example, you can suddenly call on all of those images and, and
- 33:28generate something, something, you know, really, really
- 33:33engaging. It's it's another area where you
- 33:35know if things aren't just right in code, they can do visual
- 33:38analysis as well in that way. I'm, I'm intrigued to see where
- 33:43this, this goes and how many more teams use this, use this DI
- 33:47approach as a de facto part of their day-to-day work.
- 33:50Because we've seen a lot in, in content generation.
- 33:52It's it's been established in econ teams, you know, content
- 33:56ideation, content creation, people using it to whether it's
- 34:01creating like blog post templates, which they can then,
- 34:03you know, on brand and put into their own phrase and etcetera.
- 34:08But it seems definitely more nascent in product data
- 34:11manipulation outside of traditional, I guess vendors and
- 34:15and agencies doing it. Or are you are you starting to
- 34:18see more client teams actually do more product data work with
- 34:21AI Toilet? I think like with all AI
- 34:25adoption, it's sporadic. There will be, and this has been
- 34:28the case for the last couple of years, there'll be certain
- 34:29individuals within certain businesses.
- 34:31So you can see it really engaged with it and, and doing things,
- 34:36but you know, very, very few businesses have got like a like
- 34:41a, a wide-ranging sort of policy around AI and, and, and you
- 34:46know, sort of enforcing its adoption in that way.
- 34:49So still, it's still early days, which I guess, you know, this is
- 34:54the thing as well. This is, this is all very
- 34:55emergent. And this is, you know, what we
- 34:57talk about today will change in six months time as these as
- 35:01these models continue to evolve. But yeah, it's there's no, yeah,
- 35:07it's some fits and starts. I want to suggest the adoption
- 35:09at the moment. Yeah, Yeah.
- 35:11I think that's what I see as well.
- 35:13I'm definitely seeing more and more interest, but I'm yet to
- 35:16see it embedded in in BAU practices.
- 35:20I think it's it's, it's tactically involved in some big
- 35:23projects, but it's not on a regular basis.
- 35:25Now I think this is where we'll evolve to in ECOM teams is these
- 35:28tools will be used as a regular part of product upload
- 35:31processes. There will be, there'll be less
- 35:34than moaning our we don't have these 4 attributes.
- 35:37And it will be using AI tooling to help plug gaps, but also
- 35:41helping to clean and restructure data that's coming from
- 35:43suppliers. That's not particularly lovely
- 35:45for a front end website. Yeah, I think.
- 35:48I think the enriching existing data, like with the image
- 35:52example I just gave is a really, you know, a really a really good
- 35:56use case as well, given the fact that, you know, given the
- 35:58emphasis at the moment on creating structured data, yeah,
- 36:02making all websites, the e-commerce stores included, more
- 36:06semantically understandable to machine more machine, yeah.
- 36:09And actually it's an, that's an interesting one because I've
- 36:12been, I've worked on loads of projects over the years where,
- 36:14where people have mate drawing a line in the sand and said, we're
- 36:17not going to fix historical products because we just don't
- 36:21have the time. We're going to port them as they
- 36:22are. But all new products are going
- 36:24to go through in the new data structure, better data, more
- 36:26actually it's enriched. And then over time, the, the,
- 36:29the debt drops off as the products get retired.
- 36:32But actually the, the speed of which you can do that in a, like
- 36:36focus on your call range, but then go back to older products.
- 36:39It's far easier now to do that at scale.
- 36:43Yeah, you've reminded me actually like of an actual use
- 36:45code we used on project last year.
- 36:47So we're doing a another Shopify upgrade or one of our clients,
- 36:51but Lion. And part of the brief was to
- 36:56update the sort of brand, not the brand identity, but the sort
- 36:59of brand voice. And there was a very much like
- 37:01redesigned part of the the project.
- 37:03And as part of that, that required them to revisit a lot
- 37:06of product photography, which, you know, across thousands of
- 37:10products was going to be, you know, just reviewing the
- 37:12products was going to be a pretty tedious exercise and a
- 37:15long desk job for somebody. So at the time we were looking
- 37:19at loveable, which was sort of emergent at the time.
- 37:22And we did, we built one of these one off applications.
- 37:25So we, we using loveable, we built something which synced
- 37:29back all of the product data from the Shopify store,
- 37:33including the images, and then presented everything.
- 37:38Well, it, it stored all the, the product data, it allowed the
- 37:42user to filter by product type or whatever, or newness or
- 37:46various other things. So you could slice and dice
- 37:48these thousands of products into manageable chunks.
- 37:51It then presented AUI where you could see the 1st 4 images.
- 37:55You could click on it, you could see the rest of the images.
- 37:57And then you could very quickly took a few box and say, right,
- 37:59that needs reshooting. It's the packed shot, it's the
- 38:01lifestyle shot, etcetera that needs reshooting.
- 38:04Save that and then download that as a, as a CSV that someone
- 38:08could put into a a spreadsheet. And all of a sudden you've got a
- 38:11list of all the products that needs that need reshooting.
- 38:13And we, you know, an afternoon spent building A1 off tool like
- 38:17that probably spent, you know, probably saved several days or
- 38:20somebody wading through lots of imagery.
- 38:22Yeah. Yeah, nice example.
- 38:25And product and imagery generally image audits always
- 38:28fun. Well, that's, you know, I'm
- 38:32going to stop there because we could just go down rabbit holes
- 38:34and all these sub use cases. I guess for anyone listening who
- 38:38who's got strong thoughts on this or wants wants to chat
- 38:41about a particular challenge or issue they've got and how you
- 38:45would approach it, How do they reach out?
- 38:50Just on LinkedIn, Chris Marshall on State, my e-mail address,
- 38:55Chris at on state.co.uk. I don't think I'm open myself up
- 38:58to anymore spam by revealing that I think that's 15 years.
- 39:02That's a matter of public record.
- 39:04Is now there's now 400 AI app developers who are now sending
- 39:09you their biz dev credentials for for buying their tools Mate,
- 39:12that's what's going to happen. More than happy, like, you know,
- 39:17really enjoy talking about this and listening and understanding
- 39:20what other people are doing. So yeah, if anyone's got any
- 39:24kind of real world use cases then please.
- 39:28Thanks everyone and listen as always.
- 39:30Chris's contact details are on the landing page and do connect
- 39:33with him on LinkedIn. He shares lots of really
- 39:35interesting posts about explorations and joys and trials
- 39:39of of working with this tech. So it's well worth connecting
- 39:43and and reading that content. And do keep your ears peeled for
- 39:46our next episode. Take care everyone.
- 39:50You're listening to the Inside Commons podcast, hosted by
- 39:52experienced consultants James Gerd and Paul Rogers.
- 39:55Inside Commerce is your fast track to digital knowledge,
- 39:58featuring interviews with leading e-commerce retailers,
- 40:00industry thought leads and the C-Suite, key tech companies to
- 40:03explore the latest digital strategy, customer experience
- 40:06and tech. Discover what the latest trends
- 40:08really mean and get practical advice based on the experience
- 40:11to help fine tune your e-commerce strategy and techs.
- 40:14Follow us on LinkedIn and reach out to James and Paul if you'd
- 40:17like to discuss an e-commerce project.
- 40:19We hope you enjoy the latest episode.