Podcast: Play in new window | Download
Subscribe: Apple Podcasts | Spotify | Android | TuneIn | RSS
“Leading-edge technology” may not be the first thing that comes to mind when you walk into an IKEA store, but maybe it should be.
IKEA is using AI technologies across its vast collection of businesses to deliver better content experiences to its customers.
Tane Piper leads an engineering team at Inter IKEA – the business unit that owns the IKEA brand – that is building their next generation of content and artificial intelligence tooling.
We talked about:
- his role at Inter IKEA
- the scope of AI activities at IKEA
- how their knowledge graph provides a “ground reality” for the info they share
- enterprise uses of AI at IKEA
- how narrowing the scope of models to your own enterprise improves quality and reduces costs
- the importance of testing implementations of AI technology
- how their knowledge graph helps connect content across the enterprise – and offers new content metrics and analytics benefits
- how their systems facilitate content discovery and reuse
- how he uses ChatGPT to accelerate his business research
- his thoughts on AI technologies can add a qualitative dimension to content metrics
- how AI and machine learning practices may reduce the amount of data that enterprises need to collect and store
- how they are developing prompt engineering skills at IKEA
- the importance of taking a pragmatic approach to AI adoption
Tane’s bio
Tane Piper is a self-taught software developer with over 22 years of experience. He has worked across a diverse set of environments, from startups and creative agencies to his current role as a Software Engineering Leader at IKEA. Here, Tane focuses on projects that blend content strategy, knowledge graphs, and artificial intelligence. His approach to leadership is centered on teamwork, innovation, and nurturing growth within his team.
He enjoys experimenting with a wide range of technologies. He is involved in the open-source community, releasing various libraries over the year, and writing technical articles sharing his findings.
When not engaged in software development, Tane can often be found in his garden, a hobby that provides him with a peaceful counterbalance to his professional life. Alongside his wife, he is also dedicated to the ethical breeding of Polish Hunting Spaniels, reflecting their shared passion for animal welfare.
Connect with Tane online
Video
Here’s the video version of our conversation:
Podcast intro transcript
This is the Content and AI podcast, episode number 5. When you think of the iconic furniture retailer IKEA, leading-edge technology may not be the first thing that pops into your mind. But it should. Like most enterprises now, IKEA is exploring the many ways that LLMs, machine learning, knowledge graphs, and other AI technologies can help them sell more furniture and understand their business better. Tane Piper leads an engineering team at IKEA that is building their next generation of content and artificial intelligence tooling.
Interview transcript
Larry:
Hey, everyone. Welcome to episode number 5 of the Content + AI podcast. I’m really delighted today to welcome to the show Tane Piper. Tane is a software engineering leader at Inter IKEA. And that’s the first thing I want to ask you about Tane, is IKEA is this big sprawling complex organization. Tell me about Inter IKEA and how that fits in with the overall IKEA brand.
Tane:
Yeah. Thanks, Larry. So Inter IKEA, as many people know, you go to IKEA, you go to a shop, but what a lot of people don’t know is it is actually a franchise system. So, Inter IKEA is the owner of the IKEA concept, so it owns the furniture side, the range we call it, the supply, and also the retail concept, which is where I work. So we come up with the ideas behind IKEA, how the store works, what the concept is when you go to an IKEA store, this kind of Swedishness of it all. And when you as a customer go to the store, you’re mostly dealing with franchisees, so somebody who is working with us to build the IKEA brand in a new market.
Larry:
That’s it. And having responsibility for the retail concept of one of the most iconic brands in the world, there’s absolutely no pressure in this job, I’m going to guess.
Tane:
Oh, no pressure. No pressure at all. No. In some way, yes and no. I mean, yes, it’s a big task. We are a very big, well-known brand around the world, but in insight when we’re working on things, I think we’re pretty normal. We talk about pretty normal things day-to-day. I mean, at the moment why I’m here with you today, we talk about things like content and AI and how we can use these to leverage them to improve not only the lives of our customers, they’re many people, but also what we do day-to-day.
Larry:
Great. And that’s the thing about AI is it’s so sprawling and especially in a big enterprise like Ikea, it’s going to be everywhere. Actually, I want to start because where we met, we met a couple of years ago at The Knowledge Graph Conference in New York City and we were talking then about knowledge graphs and knowledge representation and that. And we’ve subsequently talked about the work that Adam and Katariina are doing up in Sweden with the knowledge graph and all that stuff, but there’s so much more going on. Can you talk a little bit about the scope, like there I know that they’re doing the recommendation system that’s driven by the knowledge graph, but there’s so much else going on. Can you give a quick overview of how AI technologies are manifesting at IKEA?
Tane:
Oh, absolutely. I think in many layers we’re looking at it. I think when you think about what AI is, I mean AI is kind of a catchall term for a lot of things. So I mean on one end you have machine learning. So where we’re really looking at existing things like lots of unstructured content for example, in terms of an organization, we have lots of things in PowerPoints and PDF files that you can be understood by using something like machine learning. And then, yeah, very much on the other side with generative AI, certainly the opportunities that are there, the future opportunities of doing things like home furnishing recommendations and doing that in such a way that the machine understands the customer context, like the size of the room or where the location where they are.
Tane:
And can tailor and help to give the right recommendations based on that with that knowledge graph at the bottom. I think that’s the thing that when we talk about AI and machine learning, especially with a lot of the LLM stuff that is coming along is we know that’s very unstructured, unatributed. So by having that kind of ground reality with knowledge graph, that’s really what we are looking to build upon, at least in the area where I work, with content as well. So yeah, I mean, we have a group inside where it’s an informal kind of group, but we chat across multiple areas in the business and I think we’re all kind of looking at how can we bring AI to different parts of the business because there are different use cases and what use case works for our area does not necessarily work for say supply.
Larry:
Oh, interesting. Yeah. Well I guess the one that comes to mind immediately for a lot of people, because it’s been all the fuss for the last year, is degenerative AI stuff. And the famous thing about that is its proneness to hallucinations and things like that. And when you’re running a brand like IKEA, it’s really important to not hallucinate, and in your communication, and you talked a minute ago about the knowledge graph as the grounding your information in reality. Can you talk a little bit about that, how you can benefit from, because I’m assuming you can benefit from technologies like generative AI but still maintain your brand and maintain the accuracy of the information you’re sharing?
Tane:
Absolutely. I think one of the things I like that what we’ve done is we’ve not necessarily rushed into building a lot of stuff with AI and really going all in with it in many ways. And I think what that’s allowing us to do is as an organization and as brand look at what is it we need to do to show going forward that there isn’t anything that can harm the brand perception, or harm the brand itself. I mean out there are already things like there’s somebody created a model that allows you to create IKEA style manuals for things and it’s a specific model to do that. That kind of stuff is very difficult to say what effect it would have on the brand, but I think as long as people understand it’s not coming from IKEA.
Tane:
I think that’s the thing is when you put a brand out there like that, the people have a bit more of that brand themselves. So I think that’s the thing we have to take away from it. But obviously, there is protection of the brand, and yes we have to look at when we bring these generative AI services in, especially where people will be having conversations with them, putting those guardrails in there that mean that if you’re talking to a home furnishing bot or agent, you’re only getting home furnishing information. It’s not going to give you information on philosophy, or science, or anything like that. It’s keeping within the bounds of what should be doing.
Larry:
Yeah. Well, you mentioned both a minute ago when you were talking about machine learning, the benefits of learning from what you already have, and then as you just mentioned that, that kind of understanding of the world you operate in. So tell me about the relationship between understanding your current … Because you have not only product information that we all see on the web, but there’s tons of information, I forget what you call them, but the folks working in the stores and other enterprise employees at IKEA use. Can you talk a little bit about that ecosystem and how AI and ML interact as you’re both understanding and generating content?
Tane:
Yeah. But I think when you bring up, say for example with the coworker, because like I mentioned before, we’re a franchise system, then we have to train people on that system and we have to train coworkers on how to work in a IKEA store and what the range is and what the products are. So you could think of a generative AI system that is connected to that information where the coworker can start having a conversation based on maybe a customer’s needs. A customer comes up and says, “I have a bathroom that has this certain style, this certain size, what can help me here?” The coworker no longer needs to in some way dig into their own knowledge, or keep up with everything in the product range, they should now be able to go to an agent and ask questions based on this. But again, it has to be based in reality, it has to be based on the information that’s available. And yeah, certainly there’s a lot of learning as part of what we would call the franchise offer. We have to not only bring the products, the brand but the learning as well.
Tane:
So with that, yes, there’s a lot of potential for. I think we call it gen AI, I think even here we’re going to start seeing a distinction which is fully generative, which is things like DALL·E and Stability AI and all these kinds of things that are generating images and doing things like that. But then there’s this other layer which is the agent layer where it’s generative because the computer is having to do this stochastic work, but the output is more based on working with facts, if that makes sense. So it’s not so much the creative side of things, but actually working within the boundaries of the box that you have as your data set. So outside of IKEA, you can think of another industry, or another retailer, whatever, they’re going to have something very similar. How do you work within the boundaries of our brand, what we do as a company, how we interact with customers? I think we’re going to see a lot of businesses focusing more on that area, maybe more than say the generative AI that we’re seeing as a popular generative AI right now.
Larry:
Got it. You remind me of, I had a couple of episodes back, I had Paco Nathan on, I don’t know if you’ve met him at KGC, but he’s millennial, insanely knowledgeable about this stuff and he’s just been going to all the AI conferences, talking to everybody, and he shared some take homes that he’s learned over the last six months or so, one of which was the benefits of small specialized models over large models. Is that kind of what’s going on in IKEA, when you mentioned learning from your own corpus and stuff?
Tane:
So this is more going into my maybe opinions on this stuff now and certainly this is something I’ve talked about internally, which is yes, as a company, first of all, for us to have that brand protection, we should know what the models we’re using are. So I mean if you go with a ChatGPT model from OpenAI, as we know at the moment, we don’t really know where all the attribution is. We don’t know what material in there is potentially copyright versus out-of-copyright material. And I think this is where I call it a one-shot, but I guess this is small, very focused models that are trained on very specific things like home furnishing for example. So when you do that, you can create these agents that become very knowledgeable in that area, but not in a bigger context. In some sense, you have to still give them meaning, or I don’t want to say it’s not like because these things are not sentient, but you have to have that conversational style with them so they still have to understand conversation, not just the pure data of what they’re working with.
Tane:
So I think there is definitely work to be done to have these smaller models and not go for big investments in infrastructure that can be sitting around doing nothing for a very long time and then being very computationally expensive when you run them. I think it’s better that we look at having yet pipelines that generate these models that are much more focused and smaller. I mean for me, I guess thinking of where this could go is actually maybe it’s like technical side, but for example, if you think of Docker when you generate a Docker image, you have this base image and then you apply the layers on top of it. And I think it would be if we could get to something with large language models like that where you have this base model which is very small, understands conversation, but maybe doesn’t understand context, and then you build the context on top of it. And for a business like us who we release products several times a year, we have to keep building on that knowledge, but the knowledge that needs to be added on is not huge.
Tane:
It’s just that little delta from the last year for example, if you think about a year of information. So if we can get to a place where that kind of work is easier to apply, I think then these models become much more economical and much more trustworthy. I think that’s the other part as well is we need to look at the trustworthiness. I mean when I did a talk a couple of months ago with our GPT interest group as we call it, there I think my focus was that as a business, we are in EU and we have a lot of EU regulation coming up around the Digital Services Act, AI Act, but also on the industry side, lots of things around sustainability, addressability of products, for example.
Tane:
And this is where both a knowledge graph and the AI systems can work together I think quite well, but to do that we should be very pragmatic. We shouldn’t spend a lot of money in terms of infrastructure and also we need to make sure that the testing is good as well, that we test these things regularly for drift, and that we make sure that there’s a way for us to make sure they’re not going to have any effect on the brand, or on the business at all.
Larry:
Yeah. And that, grounding in facts is one that I always hear. One thing I wanted to back up a little bit about that because we had a conversation several months ago about your work and I think we talked at KGC about this, about how the knowledge graph that they’re working on, the folks up in Malmö were working on interacts with one of the CMSs that you’re running here in the Netherlands. And that I think this is, especially for my technical content people, the more content engineering folks, they’re going to be really interested in that because I think that’s just sort of a specific technical implementation of a knowledge graph that injects facts into your CMS. Can you talk a little bit about that?
Tane:
Yeah. Yeah. I mean so the team I’m specifically running is modular content delivery. I mean that’s what we have been looking at in terms of, I think this is where things are changing in the industry. I mean we want to reduce cost, the amount of data we’re transferring over networks and things. But also just get to a better way to deliver the factual content that we need to be able to deliver to our coworkers, to our customers. And working with the knowledge graph, what we’re doing is connecting those taxonomy values to content that is being produced. So when you produce a piece of content, you could say if you were thinking of a customer-facing piece of content, maybe you tag it with a room. So we have an instance of the room in the knowledge graph, like a living room. We have an instance of a style and we have an instance of an activity maybe, that you want to do. And based on that, then the content that’s tied with that now becomes searchable through the knowledge graph.
Tane:
So your content intelligence is improved because you understand more about your landscape of your content, but now you have that that can now also be fed into with the language models so that you get that contextual information from your conversation, you go to the knowledge graph, you try find your semantic reasoning from that, and then from there you can go and fetch the content around that. So I think we’re seeing there that by putting the knowledge graph at the heart, it connects to so many services and systems. You can now by having an instance of something in a knowledge graph connected not only to the content but to the analytics of the content, or the metrics of the content so that now you can actually have a full view of that content lifecycle, where’s it being used, how’s it being used?
Tane:
But also later if you want to find content based on say, products being refreshed, maybe you want to find some old content and remix it. That content findability I think is one of the key things of using the knowledge graph. And then I think this is where LMS people are using them as a search engine, they’re not a search engine, but if you give it the right information and the right function it can help with searching for information. For example, I’m doing a talk later this week in Brussels and I used ChatGPT to find information for me, but I switched the Bing mode on and I said, “You’re a speech writer in Europe. I need to find some information on some European policies around sustainability, around using customer data.” And then I said, “Okay, now we have a client called IKEA who are coming in. We want to also find some information related to what IKEA do in this area.” And it went out and found the stuff for me, it gave me references.
Tane:
So maybe in about two hours I compiled the outline of the talk with references. If I had to do that myself and had to use Google to search for this to find the right things, I’d have to read the whole paper to make sure that it’s relevant. Maybe I can read the extract and maybe understand it. Then maybe when I dive into it, there’s nothing relevant in there. But by using the ChatGPT to do that work for me, I think that’s something that a lot of business users will benefit from, especially. I mean developers, we can put code into all day, it’ll find bugs, we can help it get it to right tests and things, but we are on that early part of the bell curve anyway. We know about this stuff, we understand it, or think we understand it, we’re playing with it and we get it. But for business users at the moment, they’re maybe looking at it and having some fun with the generative AI stuff, but they don’t yet know the power of what can come with this when you connect it to your data.
Larry:
Yeah. That’s my hope in this podcast is to get more people more attuned to the full range of benefits from AI technologies in content practice. You just mentioned a couple of things that fall out of what you’re doing, like reuse and discovery, that’s huge in a lot of content operations, especially in technical stuff. But you also mentioned the analytics, and I’m really intrigued by that because it just frustrates me to know most of what we talk about when we talk about content analytics is page-level web analytics. Are you doing fancier stuff than that now, or?
Tane:
I think when we think about the business of IKEA, it’s only successful because of the people. If you go to a store and there’s no people, then there’s no success because we’re not being able to sell anything. So knowing how for example, learning is being consumed, how learning is being engaged with. There’s quantitative stuff that you can get from systems there. You can plug them in and create data metrics from that, and then you can provide it as dashboards to people who need them. I think where it becomes interesting for me is where up until now, especially a lot of analytics has been thought of as very quantitative, it’s just numbers. The interesting future with this is how you bring the qualitative part back into this and actually how you work. I mean, you could imagine, for example, one of my things about learning is it’s not only just contextual, but it’s contextual to the person that day. Imagine if, before every training course that somebody does, you just say, “How are you feeling today? Are you in a good mood, a bad mood? How was your commute?” Just a couple of questions.
Tane:
And then based on that, you can maybe get some sentiment from people as well. You can understand that if people have a bad commute and a bad day, then they learn terribly, but you can verify that now. I think that’s the thing is where we’re seeing bringing in some of the more quantitative elements. And we have a big learning organization, but there’s always more to learn. There’s always better ways to optimize things, but that’s just one area. I mean, when we talk about bringing metrics in, it’s across how quickly can we optimize supply chain. How quickly can we improve getting stuff out to the customer? How quickly can we improve customer service? There’s lots of areas there where I think in my opinion of metrics in the last 20 years is that we’ve gone down this route of gather everything almost like an NSA-type thing where it’s like have all the information and then we’re just going to run all this stuff of this information. But then we have so much information that it doesn’t actually give you that much that you can get out of it. I’m trying to kill it.
Larry:
No, it’s funny you’re reminding me of why I feel somehow safe online is because I’m aware that I’m just lost in this morass of information that all the big orgs are trying to figure out.
Tane:
But it’s also, it is costly for a business. It’s not only costly in terms of storage, moving data around, but costly in terms of brand effect. If something happened and some of that data got leaked out and then suddenly somebody was releasing information about customers, then that’s not great. So we need to, I think actually going forward, one of the things that AI and or at least machine learning and LLMs might help with is actually reducing the volume of data that we need to store and collect. I think there’s routes there that will help with that. I don’t know if it will happen right away. I think, now you don’t have Google Analytics actually a lot of places are like, “We don’t have this data coming in now. We have no visibility on this stuff.”
Tane:
But I would bet if you spoke to them and asked them when you had all this vast amount of data, what were you really getting out of it? What were you really able to understand and target and improve in the business from what you had there? And if it’s more just a case that they’ve gone from having visibility of something to no visibility, what’s the key things you’re missing? And then think about that and take that forward. Because actually, you probably don’t need to know every piece of sensor data from a customer. You just need to know are they completing a certain journey, or something like that.
Larry:
As you say that, I’m really interested in, I’m thinking of that data, information, knowledge, wisdom pyramid. And I think what you’re saying is a lot of that data at the bottom is just kind of data points that support a piece of information-
Tane:
It’s nice.
Larry:
… Or knowledge that we’ve discerned. And once you’ve codified it as information or knowledge, you can just get rid of the data that supported that. Is that?
Tane:
Exactly. It’s just static noise that sits around. And yeah, I think unless you, for archival and historical reasons maybe need to keep some data around, there’s some cases there, but for most stuff you don’t. And when it comes to Germany here in the EU, I mean that’s usually a big black box of data because you’re not allowed to get the data out of Germany, for example. So that hits every business across, and especially in retail across Europe is a lot of the time you don’t know what’s going on there.
Larry:
Yeah. Everybody loves it. Hey, one thing I quickly want to ask about. You’ve mentioned a few times which things that remind me that one of the benefits of these LLMs and the generative AIs, you almost always interact them with a conversational interface, like a chatbot kind of thing.
Tane:
Yes.
Larry:
Has that changed content practice at IKEA? Are there more like conversation designers, or folks around, or how is that manifesting?
Tane:
Not yet. I think that’s something that as we take the time to do this is learning what is it we need. I mean, I’m involved in a small work group around the generative AI stuff, and I think we’ll have to come up with our prompts. Just the other day, I think the open AI prompt came out, it showed you what they feed to ChatGPT before you even put anything in there. Don’t create any art from anyone past 1912, I think it is. So if you put an artist’s name who’s been producing art after 1912, it will say, “I can’t do that, but here’s something like it.” Or, “Here’s some suggestive type way of doing it.” And I think it’s the same for us as we all have to have our prompts for certain areas that we have to work out. So I think you’re right, it’s become conversationalist.
Tane:
Yeah. I think we call it prompt engineering right now, but I think in the long run, yeah, it’s a very crass name for just basically making sure that when you deal with this system, and when you’re interacting with it that it doesn’t go, yeah, you’re trying to get it to not go off the rails. I think there’s no pretend that this thing is in any way sentient. I think there’s an argument to say consciousness. I think what we mean by consciousness is more like it’s able to take sensory information, which you’re providing it and make a reasoning based on that. So I mean, you could call that an element of consciousness, but it’s not a sentient creature, but it understands context, it understands meaning. It can get confused, of course it can. And you can definitely take it around in circles. But what human is not like that? What human has never just said random things in the middle of a conversation, or misunderstood what you’ve said and gone on a completely different tangent.
Tane:
These things, I think that’s the thing for me is LLMs are just a mirror, part of society in general because we feed that information in there and that’s where unfortunately, biases and misinformation can creep into the system. But then the second half of it is you have to have a human to interact with it. I mean, back last year I did the stochastic parrot, I think it is still online, stochasticparrot.long, where I just gave GPT empty prompts and it just generated stuff. From empty prompts, just kept generating and generating. And that I think made me realize there’s a technique, like a Markov chain. It’s where you feed a lot of text into a program and then the program picks a word and then it does the stochastic choosing. OpenAI ChatGPT is nothing more fancy than that except it has a lot more information and then it has some additional things behind it to help with context and understanding.
Tane:
But the general system is just, it has to have a starting point. So the starting point is the human asking the question, and then from there it builds on that context. And I do see from my wife using it, she does a lot of stuff around behavioral stuff with this, and she’s playing with it, and it’s about being nice with it. It’s about having a conversation with it, and it’s an iteration. So it’s almost like a client relationship in some way, you’re working with someday. And it’s almost like a digital whiteboarding of an idea or a concept and see how far you can take it.
Larry:
You’re reminding me of, I can’t remember who it was, somebody in the knowledge graph community, we were talking about the human in the loop. They go, “No, it’s the machine in the loop.” And I’m kind of taking home a similar lesson from what you just said, that it’s just conversation. And we’re probably better at it than LLMs at this point, but they’re getting really good, at least the language part of it. But hey, Tane, I can’t believe it. We’re coming up close to time already. Before we wrap up though, is there anything last that you want to revisit from the conversation or just want to make sure we get to before we close?
Tane:
I think just to go back and what we’ve been doing for maybe the past year is you’ve maybe seen some stuff in the press, but you’re doing some stuff with AI. And I think that is, like I say, the retailer side of things are kind of doing some experimentation around that. But as a company at the core of the brand and the business, I think what we’re doing is really taking a pragmatic approach here of what benefits can we as not a tech company, at the end I might have engineering leader in my title, but at the end of the day I sell furniture, that’s my job. That’s what the enablement is in what we do. So how do we as a business benefit from that without the negative consequences of using this stuff? We know it takes a lot of resources to run these things. We know that they hallucinate and drift.
Tane:
We know that there’s legislation that’s coming up that will hit us very hard if we do this wrong. So taking all of those things into account, I think inside the area, we are really thinking what’s the future of this that we can take forward and bring benefits for many people and not just for a few? And I think if we can help solve that problem, if we can be part of the solution, I’d be very happy to see that kind of work coming from us. I think you know I’m quite outspoken sometimes when it comes to some of what the other people in this area are doing, and especially with some more of that tech growth companies where they’re pushing out all this stuff without thinking about the consequences. And I think that’s the key part is if you’re in a big business, think about the consequences because the potential harm to come from these systems without the right understanding, without the right guardrails, it puts a lot of things at risk.
Larry:
Yeah, I love that. Closing on that pragmatic business focus level, because the tech guys, they’ll sell you anything you’ll buy, though. But, one very last thing, Tane, if folks want to follow you online or connect, what’s the best way to reach you?
Tane:
These days LinkedIn, you can find me there. Feel free to connect. Feel free to send me a message if you have any questions. I usually try to respond within a couple of days. But yeah, I’m happy to connect with anyone in the space who’s interested in finding out a bit more.
Larry:
Well, thanks so much Tane. Really enjoyed the conversation.
Tane:
Thank you.
